American Economic Association

Educational Investment Responses to Economic Opportunity

Author(s): Anjali Adukia, Sam Asher and Paul Novosad
Source: American Economic Journal: Applied Economics , January 2020, Vol. 12, No. 1
(January 2020), pp. 348-376
Published by: American Economic Association

Stable URL: https://www.jstor.org/stable/10.2307/26866470

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/10.2307/26866470?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Economic Association is collaborating with JSTOR to digitize, preserve and extend access
to American Economic Journal: Applied Economics

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
American Economic Journal: Applied Economics 2020, 12(1): 348–376
https://doi.org/10.1257/app.20180036

Educational Investment Responses to Economic

Opportunity: Evidence from Indian Road Construction†
By Anjali Adukia, Sam Asher, and Paul Novosad*

The rural poor in developing countries, once economically ­isolated,

are increasingly being connected to outside markets. Whether these
new connections crowd out or encourage educational investment is
a central question. We examine the effects on educational choices of
115,000 new roads built under India’s flagship road construction pro-
gram. We find that children stay in school longer and perform better
on standardized exams. Heterogeneity in treatment effects supports
a standard human capital investment model: enrollment increases
most when nearby labor markets offer high returns to e­ ducation and
least when they imply high opportunity costs of schooling. (JEL I25,
I26, J24, O15, O18, R42)

I ncreased access to international markets has important influences on schooling

decisions, which are central to supporting l­ong-run economic growth.1 A large
share of the world’s rural poor are not w ­ ell connected to international ­markets,
­however, and depend instead on domestic linkages to nearby towns and cities.2
The impacts of domestic market integration are less studied than the impacts of
­connections to international markets. A key ­trade-off for individuals is between
­long-run investment in human capital and immediate economic opportunities that
might discourage increased schooling. Connections to new markets should encour-
age educational attainment if they increase returns to education, or otherwise raise
household income or liquidity. However, immediate earnings opportunities for the

* Adukia: University of Chicago, 1307 East 60th Street, Chicago, IL 60637, and NBER (email: adukia@
uchicago.edu); Asher: Johns Hopkins University SAIS, 1717 Massachusetts Ave NW, Washington, DC 20036
(email: [email protected]); Novosad: Economics Department, Dartmouth College, 6106 Rockefeller Center, Room
301, Hanover, NH 03755 (email: [email protected]). Dave Donaldson was coeditor for this article.
We thank Martha Bailey, Chris Blattman, Liz Cascio, Esther Duflo, Eric Edmonds, Rick Hornbeck, Ruixue Jia,
Ofer Malamud, Mushfiq Mobarak, Doug Staiger, Bryce Steinberg, and participants of seminars at AEFP, APPAM,
Boston University, CIES, Columbia University, CSWEP, Dartmouth, DePaul, EBRD, EEA, the Federal Reserve
Banks of Chicago (CHERP) and New York, Georgetown University, Michigan State University, NEUDC, the
NBER Education Program Meetings, Northwestern University School of Law, the OECD, PAA, PacDev, Stanford,
University of California-Berkeley, University of Chicago, University of Connecticut, University of Illinois at
Chicago, University of Michigan, University of Missouri, the Urban Institute, and Yale, and the anonymous referees
for their ­helpful comments and suggestions. We thank Srinivas Balasubramanian, Jack Landry, Anwita Mahajan,
Olga Namen, and Taewan Roh for excellent research assistance. We thank Arun Mehta and Aparna Mookerjee for
help in data acquisition.
†
Go to https://doi.org/10.1257/app.20180036 to visit the article page for additional materials and author
disclosure statement(s) or to comment in the online discussion forum.
1
See, for example, Edmonds and Pavcnik (2006); Edmonds, Pavcnik, and Topalova (2010); and Shastry (2012).
2
See, for example, Atkin et al. (2015), who shows that domestic trade costs in developing countries can be
­considerably higher than international trade costs.

348

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 349

young could motivate an earlier exit from schooling. As educational investment

responds to market integration, it shapes the ­long-run economic impacts of policies
that are increasingly integrating markets in developing countries.
We examine the human capital investment response when a paved road is built
to a previously unconnected village, effectively connecting it to outside markets.
India’s national rural road construction program (PMGSY) built high quality roads
to 115,000 villages across the country between 2001 and 2015, connecting over
30 million rural households to nearby towns. We focus on new rural feeder roads,
which provide terminal connections between the broader transportation network
and previously unconnected villages. The impacts of new road connections on
­schooling are theoretically ambiguous: they may raise the returns to education, raise
the ­opportunity cost of schooling, and/or have important income or liquidity effects.
A major challenge in estimating causal effects of new roads is the endogeneity
of road placement. If roads are targeted to wealthier or poorer regions, for ­example,
then comparisons of villages with and without roads will be biased. To overcome
this bias, we exploit the timing of road completion in each village, estimating a
panel regression with village and ­state-time fixed effects. Village fixed effects
­control for unobserved ­village-specific factors that may have influenced the timing
of road ­construction. ­State-time fixed effects control for ­time-variant ­state-specific
shocks and policies. We thereby compare educational outcomes in villages before
and after a road is built, flexibly controlling for ­time-variant regional shocks and
static differences between villages that receive roads in different years.
We use ­village-level school enrollment data from India’s national annual ­census
of primary and middle schools (District Information System for Education (DISE)
2002–2015). Through a combination of human and machine fuzzy matching, we
linked DISE data to administrative data from the national rural road construction
program. The result is a panel of 300,000 villages across India. The use of ­census
data is essential to our analysis, since variation in the road program is at the ­village
level. It also gives us power to precisely estimate heterogeneous impacts in subsa-
mples of the data. Our sample spans a broad range of economic conditions in India
today, similar to the variation across many places worldwide that remain unreached
by paved roads.
We find that road construction significantly increases middle school ­enrollment.
We estimate that connecting a village with a new paved road causes a 7 percent
increase in middle school enrollment over the following three years. The estimates
are precise and statistically significant. We also estimate increases in the number of
students taking and scoring highly on middle school completion exams, ­indicating
that educational performance is also improving.3 The results are robust to a range
of specifications and sample definitions, as well as a regression discontinuity
­specification that exploits a program rule that caused villages above specific village
population thresholds to be targeted for road construction.

3
In many cases, interventions that improve attendance and enrollment do not improve student test scores
(e.g., Miguel and Kremer 2004; Behrman, Parker, and Todd 2008; and Adukia 2017), perhaps due to congestion.
Congestion effects in our study may be counterbalanced by ­already enrolled children working harder.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
350 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

We do not find enrollment effects for primary school children, for whom there
are fewer labor market opportunities. We do find small increases in primary school
performance, however, suggesting that students may be increasing school effort on
the intensive margin.
We then explore heterogeneity in the treatment effects on middle school ­children,
guided by predictions from a standard model of human capital investment. The
model predicts four primary mechanisms through which educational investment in
rural areas may be affected by road connections to nearby labor markets. We model
roads as leading to factor price equalization across areas, which is then predicted
to: raise the ­low-skill wage and thereby increase the opportunity cost of schooling;
raise the skill premium and thus increase the returns to education; increase lifetime
household earnings (an income effect); and ease a liquidity constraint.4 The model
suggests that the importance of each of these effects will be different across regions,
depending on local market characteristics. Newly connected villages will experi-
ence larger opportunity cost effects when the u­ rban-rural low skill wage gap is large.
Returns to education effects will be larger when the ­urban-rural gap in Mincerian
returns to education is larger. To predict the expected importance of income and
liquidity effects, we use a measure of asset poverty.5
The estimated variation in treatment effects across these three measures supports
the predictions of the model. Partitioning our data according to these measures, we
estimate substantial treatment effect heterogeneity across villages, with effects that
are positive and statistically significant in 39 percent of villages and positive but
insignificant in 52 percent of villages. Market integration has (small and statistically
insignificant) negative effects only in the 9 percent of villages where we expect
opportunity cost effects to be high, and returns to education and income/liquidity
effects to be low—exactly where the theory predicts treatment effects would be
most negative.
We explore several other treatment mechanisms, for which we do not find support
in the data: ­supply-side improvements in school infrastructure; migration; displace-
ment to/from nearby villages; and improved access for children on the outskirts of
villages.
Our findings suggest that integrating the rural poor with regional markets has the
potential to drive further ­long-run growth through increased educational ­attainment.
Enrollment and exam performance respond positively to increased economic
­opportunities. Our results also provide a causal mechanism that underlies the strong
correlation around the world between education, growth, and trade.
This study is related to the literature on the impact of labor demand shocks on
schooling decisions, which finds both positive and negative schooling impacts from
new economic opportunities.6 The estimated heterogeneity in treatment effects

4
Because roads may change factor prices in many markets, many other effects are also possible. We focus on
effects predicted from the literature on road construction.
5
Income and liquidity effects are theoretically distinct but difficult to disentangle without detailed
­household-level data (Edmonds 2006), so we consider them together.
6
The opening of new outsourcing facilities in India and garment factories in Bangladesh has driven increases
in schooling (Jensen 2012, Oster and Steinberg 2013, Heath and Mobarak 2015). Positive agricultural demand
shocks in India, expansion of natural gas fracking in the United States, and expanded export manufacturing
in Mexico have increased dropout rates, especially for middle school children and older children (Cascio and

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 351

in our study is consistent with the heterogeneity found in the literature, and well
explained by the standard human capital model: new labor market opportunities
affect the opportunity costs of schooling, but also affect the ­long-run benefits of
schooling and demand for schooling through income and liquidity effects.
Our paper also contributes to the literature on the development impacts of
­transport infrastructure.7 Relative to earlier work on roads and schooling, our
large ­village-level sample and research design allow a more precise estimation of
the causal effects of road construction. Our results suggest that road construction
and domestic market integration may have greater l­ong-run impacts on economic
­development by increasing educational investment. Finally, we contribute to a wide
body of research on improving educational attainment in developing countries (see
Evans and Popova 2016 and Glewwe and Muralidharan 2016 for reviews of this
literature). Our results highlight that investments outside the education sector can
have important effects on schooling decisions.
This paper is organized as follows. Section I presents a conceptual framework
describing human capital investment decisions and the role of market integration.
Section II provides background on road construction and education in India. We
describe the data in Section III and the empirical strategy in Section IV. Section V
presents results, Section VI explores the mechanisms suggested by the model of
human capital investment, and Section VII concludes.

I. Conceptual Framework: Schooling Decisions and Economic Opportunity

We outline a standard conceptual framework to help explain how human capi-

tal investment decisions respond to changes in labor market opportunities (Becker
1964). This framework helps to explain why the impacts of labor demand shocks
on schooling vary across the empirical literature, and motivates our later analysis
of how roads’ impacts on rural schooling decisions are affected both by character-
istics of villages and by characteristics of local labor market conditions outside the
village.
In this framework, the key decision point is an individual’s t­rade-off between the
­long-run benefits of schooling and the ­short-run return to labor. A ­two-period model
is sufficient to highlight the essential comparative statics. In the first period, a p­ erson

Narayan 2015, Atkin 2016, Shah and Steinberg 2017). Our estimates are also related to impacts on human cap-
ital ­accumulation from India’s national public works program Mahatma Gandhi National Rural Employment
Guarantee Act (MGNREGA), which has found small decreases in enrollment for middle school students across
India (Das and Singh 2013, Islam and Sivasankaran 2015, Shah and Steinberg 2015, Adukia 2018, Li and Sekhri
forthcoming).
7
Some examples include Jacoby (2000); Jacoby and Minten (2009); Gibson and Olivia (2010); Mu and van de
Walle (2011); Casaburi, Glennerster, and Suri (2013); Donaldson and Hornbeck (2016); and Donaldson (2018).
For a detailed review, including studies on the impacts of highways and regional roads, see Hine et al. (2016). Asher
and Novosad (forthcoming) finds that Pradhan Mantri Gram Sadak Yojana (PMGSY) road construction leads to
changes in occupations but has little effect on village assets, incomes, or consumption using regression discontinu-
ity exploiting village population thresholds. Mukherjee (2012) uses a ­similar approach to find that PMGSY road
construction in India increases school enrollment. We present ­comparable regression discontinuity estimates, but
we focus on panel estimates that have much greater statistical precision and allow for analysis of treatment hetero-
geneity. Using ­district-level data from India, Aggarwal (2018) finds an ­association between road construction and
school enrollment. Khandker, Bakht, and Koolwal (2009) and Khandker and Koolwal (2011) show that s­ mall-scale
road construction in Bangladesh is associated with increased school enrollment.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
352 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

chooses between working for a ­low-skill wage and obtaining schooling. In the sec-
ond period, the person works and receives either a high or a low wage, depending
upon the schooling choice in the first period. The person consumes in both periods,
drawing from an initial endowment and wages earned in each period that the person
works. The person can save at some interest rate, but may be restricted in borrowing.
The person’s initial endowment can reflect household wealth or wages of household
adults who have completed their schooling. Education may also be a normal good,
which households value independently of its impact on future wages.8
When a village becomes connected to an external market via a new road, there
is a change in the parameters underlying this ­trade-off between education and early
participation in the labor market. Reduced transportation costs affect worker wages
in both periods by inducing factor price equalization across areas.9 In equilibrium,
urban areas have higher wages than rural areas for both low- and ­high-skilled work-
ers, and higher Mincerian returns to education (see online Appendix Table A1).10
Connecting a village to its external market is therefore likely to increase the ­low-skill
wage and increase the returns to education.11
An increase in the ­low-skill wage raises the opportunity cost of schooling and
­discourages human capital investment, which we call the opportunity cost effect.
A relative increase in the h­igh-skill wage raises the returns to education and
­encourages human capital investment, which we call the returns to education effect.
Changes in wages could also affect human capital investment through income
effects or liquidity effects. As wages rise, income effects will increase human ­capital
investment if schooling is a normal good. Increases in household liquidity may also
affect human capital investment if credit constrained households cannot afford to
pay school fees or require children to work. In principle, these effects could go
in either direction, but based on u­ rban-rural wage gaps and skill gaps in India, we
expect the opportunity cost effect to reduce schooling, and the returns to education,
income, and liquidity effects to increase schooling.
Predictions for how human capital investment is affected by factor price
­equalization in goods and capital markets are less clear, as many prices can change
simultaneously.12 Rural road construction could also affect schooling decisions
through many other channels, such as information flows, marriage markets, and
healthcare access, but we focus on impacts through labor markets and wages.

8
This framework underlies much of the theoretical literature on child labor and human capital invest-
ment ­decisions. See, for example, Ranjan (1999) or Baland and Robinson (2000). We abstract away from
­intra-household bargaining.
9
Wage convergence could come from permanent migration, temporary migration (e.g., daily commuting to
larger markets along new roads), or changes in factor prices due to goods market integration. Asher and Novosad
(forthcoming) shows that new PMGSY roads increase the number of people working for wages outside of villages.
10
It is possible that these static price differentials reflect unobserved differences in skills of workers in different
locations, even controlling for education. For example, the quality of education in rural areas is probably lower than
in urban areas. However, unobserved education quality differences are unlikely to drive the entire differential, given
the presence of higher skilled jobs in cities and towns, and the high returns to r­ ural-to-urban migration documented
in other studies, e.g., Bryan, Chowdury, and Mobarak (2014).
11
We can think of these effects as changes in real wages, such that changes in local goods prices due to new
roads are subsumed in the above effects.
12
For example, in general equilibrium, there may be various changes in the prices of different intermediate
goods and final goods. Capital market integration could also affect interest rates, changing the impact of liquidity
constraints and changing the return on savings.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 353

To explore which of these mechanisms are more important, we identify places

where these mechanisms are likely to generate heterogeneous impacts on schooling
decisions. We generate measures of regional labor market conditions, which should
influence the magnitude of changes in l­ow-skill wages and the returns to education
when a village becomes more integrated with nearby labor markets. The opportunity
cost effect should be particularly large when the l­ow-skill regional wage is much
larger than the ­low-skill wage in the village becoming connected. The effect on
returns to education should be larger when regional returns to education are greater
than returns to education in the village becoming connected. We expect that income
and liquidity effects on schooling would be greater in villages that are liquidity
­constrained or have low incomes, though the economic opportunities created by new
roads may also differ in these villages. In the absence of shocks that separately affect
liquidity and income, these liquidity and income effects are difficult to disentangle
(Edmonds 2006), and so we consider them together.

II. Background and Details of the Road Construction Program

School enrollment increased substantially in India over our study period, from
2002 to 2015, paralleling a global increase in educational attainment. Increasing
educational attainment has been a national priority in India, with several national
initiatives aimed toward achieving universal primary education. Educational attain-
ment and rates of economic development vary substantially across India. Indian
­policymakers in the past have allocated public goods with an aim to mitigate spatial
inequality, but large disparities remain and are at the center of public debate in India
(Banerjee and Somanathan 2007, Drèze and Sen 2013).
Many rural villages have limited connections to regional markets even while
major cities in India have become increasingly connected to world markets. High
costs of road construction and rapid degradation have historically constrained the
ability of the Indian government to connect every village. In 2001, 49 percent of
Indian villages remained inaccessible by ­all-season roads. These villages were
­characterized by greater poverty and lower educational attainment.
In 2000, the Government of India launched the Pradhan Mantri Gram Sadak
Yojana (Prime Minister’s Village Road Program, or PMGSY), a national program
that aimed to eventually build a paved road to every village in India. While the fed-
eral government issued implementation guidelines, decisions on v­ illage-level allo-
cations of roads were ultimately made at the district level. The unit of targeting for
road construction was the habitation, which is the smallest rural administrative unit
in India. A village is typically comprised of between one and three habitations; there
are approximately 600,000 villages in India and 1.5 million habitations. We focus
on villages as the unit of analysis because: many villages have only one habitation;
many habitations were pooled to the village level for the purposes of the program;
and little economic data is available at the habitation level.
Road construction was targeted initially toward villages with larger popula-
tions. In some states, this took the form of a strict population threshold for road
construction eligibility, while other criteria were used in other states. Given the
­program rules, ­early-treated villages tended to have larger populations, but were not

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
354 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

s­ ubstantially different from ­late-treated villages in other characteristics.13 There

were initially 80,000 villages eligible for the road construction program, a number
that has grown as guidelines have been expanded to include smaller villages.
By 2015, over 115,000 villages had paved roads built or upgraded under the
PMGSY program. These construction projects were most often managed through
subcontracts with larger firms, and were built with ­capital-intensive methods and
external labor; the building of the road itself was therefore not a major local labor
demand shock. These PMGSY roads are distinct from new roads being built under
the Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA),
which are lower quality roads built with ­labor-intensive methods.14
Figure 1 shows the distribution of road construction by state and over time. The
median road length was 4.4 kilometers. Given the difficult terrain around many of
these villages, a new paved road represents multiple hours saved on a round trip to
or from the village.

III. Data

We constructed a ­village-level panel dataset, combining data on road construction

with education outcomes and other village characteristics. We matched an annual
census of Indian schools, the District Information System for Education (DISE,
­2002–2015), to administrative data from the implementation of the road program
(­2001–2015) and three successive Indian Population Censuses (1991, 2001, 2011).
We matched locations based on village, block, and district names using a set of
fuzzy matching algorithms.15
DISE is an annual census of primary and middle schools in India. It includes
data on student enrollment, exam completion, and school infrastructure. This dataset
was created by the Ministry of Human Resource Development of the Government
of India and is administered by the National University of Educational Planning
and Administration. DISE data cover every registered Indian government primary
and middle school beginning in 2005.16 We also have DISE data for a smaller
­sample of schools from ­2002–2004, a period when the ­data-collection system was
still being rolled out on a d­ istrict-by-district basis. We are able to replicate national
­survey-based statistics on enrollment, suggesting that the DISE data are reliable.17

13
District fixed effects explain 30 percent of the variation in year of treatment among treated villages.
A ­population quartic explains another 9 percent of the variation, after which inclusion of additional control
­variables has virtually no additional predictive power.
14
Major highway projects during this period, such as the Golden Quadrilateral, were planned and executed
independently of PMGSY. There is not evidence of coordination of PMGSY road construction with the construction
of the Golden Quadrilateral or other district road improvement projects.
15
For fuzzy matching, we used a combination of the reclink program in Stata and a custom fuzzy ­matching
script based on the Levenshtein algorithm but modified for the languages used in India. The fuzzy matching
­algorithm is posted at github.com/paulnov/­masala-merge. We were able to match 83 percent of villages in the road
administrative data to the population censuses and 65 percent of villages in DISE. We matched 80 percent of census
blocks; within census blocks, we matched 81 percent of villages.
16
We refer to academic years (which begin in June or July) according to the beginning of the school year (e.g.,
we refer to academic year ­2007–2008 as 2007).
17
Our preferred sample drops villages that reported total enrollment (first through eighth grades) greater
than 60 percent of total population (the ninety-ninth percentile of this statistic), which appear to be measured
with error. By comparison, in 2001 only 22.4 percent of the population was of primary school age or middle

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 355

Panel A

15,000
Number of new roads

10,000

5,000

0
01

02

03

04

05

06

07

08

09

10

11

12

13

14

15
20
20

20

20

20

20
20

20

20

20

20

20

20

20

20

Panel B
Andhra Pradesh
Arunachal Pradesh
Assam
Bihar
Chhattisgarh
Gujarat
Haryana
Himachal Pradesh
Jammu Kashmir
Jharkhand
Karnataka
Madhya Pradesh
Maharashtra
Manipur
Meghalaya
Orissa
Punjab
Rajasthan
Tamil Nadu
Uttar Pradesh
Uttarakhand
West Bengal

0 5,000 10,000 15,000 20,000

Number of new roads

Figure 1. Summary of Road Construction under PMGSY

Notes: The panels in this figure describe the distribution of new roads built under PMGSY between 2001 and 2015
across years and states. Graphs show new roads according to their registered completion dates.
Source: PMGSY Online Monitoring and Management System

school age (ages 6–15). Demographic data from the Below Poverty Line Census (2002) suggests that fewer than
40 ­percent of village residents are between 6 and 15 years of age in 99 percent of villages. Our results are not
­materially changed by these decisions.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
356 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

DISE data are based on interviews with school headmasters, and there are potential
concerns of misreporting and inflated enrollment. Because new roads lower the cost
of monitoring enrollment numbers, we expect that changes in misreporting would
bias downward the estimated impacts of roads. It is also unlikely to affect only
­middle school students.
Our main outcome variable is log middle school enrollment, which we define
as the natural logarithm of one plus the total number of middle school children
enrolled in all schools in a village. Our main focus is on outcomes for middle school
­children (grades ­6–8) because the transition to middle school is a natural breakpoint
in a child’s schooling at which educational milestones are often measured, but we
report some outcomes for primary school children as well. Younger children also
have fewer labor market opportunities. DISE did not report enrollment informa-
tion for high school over our sample period. DISE does not report the total number
of ­school-age children in a village, so we are unable to calculate enrollment rates
directly. However, we can track total village population at 1­ 0-year intervals using
the Population Census, allowing us to make indirect inferences about enrollment
rates.
DISE collects information on examination outcomes in the set of states with
­end-of-school examinations for primary schools and middle schools. These exams
are used for promotion decisions and completion verification. The information
­collected includes the number of students that appeared for the exam, that passed
the exam, and that scored high marks. We also use DISE data on school infrastruc-
ture, which describe the ­school-level presence of piped water, sanitation facilities,
electricity, a library, a computer, a boundary wall, and a playground.
For data on road construction, we use administrative records of the PMGSY
­program that are used to track and implement the program.18 Road data are reported
at the village level, or at the smaller habitation level that we aggregate to the village
level. We define a village as having a paved road at baseline if any habitation in that
village had a paved road. We define a village as receiving a new road by a given
year if any habitation in the village received a new road before September 30 of the
school year, which is the date on which DISE records enrollment numbers.
Online Appendix Figure A1 shows how we define our main sample of villages.
We restrict our sample to villages that did not have a paved road in 2001. We further
limit the sample to villages that received new PMGSY roads between September
2003 and September 2015, and we exclude villages where roads were categorized
as upgrades rather than as new roads. Our main sample includes 10,014 villages that
received roads between 2003 and 2015, for which we have enrollment data for at
least one ­pretreatment year and one ­posttreatment year. We find similar results when
we extend our sample to an unbalanced sample (n = 19,152) or include v­ illages
that never received PMGSY roads (n = 112,475).
For data on ­district-level rural wages and urban wages, we use data from all
individuals reporting wages from the fifty-fifth round of the NSS Employment and
Unemployment Survey (­1999–2000). For data on village population and other

18
We obtained these data from the government’s public reporting portal for PMGSY, hosted at http://omms.
nic.in.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 357

Table 1—Summary Statistics at Baseline

Mean (SD)
Population (2001 census) 1,291.4
(998.3)
Nonfarm employment (1998 economic census) 60.1
(173.8)
Number of primary and middle schools 1.7
(2.0)
Total enrollment (grades 1–8) 217.1
(389.0)
Total primary school enrollment (grades 1–5) 178.0
(286.8)
Total middle school enrollment (grades 6–8) 39.1
(125.6)
Middle school exam passers (2005) 7.3
(15.4)
Exam passers with distinction (2005) 1.5
(5.4)

Note: The table shows means and standard deviations (in parentheses) of village-level ­variables
at baseline in the sample of villages that were matched across all analysis datasets.
Source: Unless otherwise indicated, the data source is the District Information System for
Education (DISE), 2002.

c­ haracteristics, we use data from the Population Censuses of India in 1991, 2001,
and 2011, and the 1998 Economic Census.
Table 1 shows summary statistics of villages at baseline. The enrollment d­ rop-off
at middle school is substantial: the average primary school cohort has 36 children
per year, while the average middle school cohort has only 13 children.

IV. Empirical Strategy

Our goal is to estimate the causal impact of roads on educational choices.

­Cross-sectional estimates of the relationship between village roads and school-
ing decisions are likely to be biased estimates of the impact of roads on schooling
because villages that do not have access to paved roads are different from more
accessible villages along many dimensions. For example, villages without paved
roads are likely to be smaller, have more difficult terrain, and be more politically
marginalized. Our main empirical specification is a panel fixed effects regression
that exploits the timing of road construction, within the set of all villages that
received new roads by 2015 under the PMGSY program.
The panel estimation exploits variation in the year that a village was connected to
the road network. The panel estimator is defined by the following equation:

(1) ​​Y​i, s, t​​  = β ⋅ ​ROAD​i, s, t​​  + ​γs,​ t​​  + ​η​i​​  + ​ϵ​i, s, t​​  .​   

Here, ​​Yi​, s, t​​​is the outcome variable (such as school enrollment), measured in village i​ ​
and state ​s​in year ​t​; ​​ROAD​i, s, t​​​is an indicator variable for whether the village has

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
358 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

been connected by a paved road by year ​t​; ​​γ​s, t​​​is a ­state-year fixed effect, and ​​η​i​​​ is a
village fixed effect. The error term, ​​ϵ​i, s, t​​​, is clustered at the village level to account
for serial correlation in the dependent variable.
The coefficient of interest, ​β​, measures the impact of a new road on v­ illage-level
outcomes (such as log school enrollment). All villages have ​​ ROAD​i, 2002​​  = 0​
and ​​ROAD​i, 2015​​  = 1​, i.e., all sample villages received a road at some point under
the program between 2003 and 2015. We thus avoid making a potentially biased
comparison between villages that were and were not eligible for new roads.
The identification assumption is that in the absence of the PMGSY, ­village-level
outcomes would have followed the same path over time in villages that receive
a paved road in different years, after partialling out the location and time fixed
effects. The ­state-year fixed effects control flexibly for differential enrollment
growth across states. This alleviates concern that states with more effective govern-
ments ­simultaneously built roads and also provided other government services; it
also ­controls for any broader regional trends in enrollment that might be correlated
with road construction. The village fixed effects control for systematic differ-
ences between early- and l­ate-treated villages. No additional controls are included
because the village fixed effects account for all static village characteristics, and
we do not have annual data on any time varying characteristics of villages other
than school enrollment. We also present specifications that control for village time
trends and for baseline village characteristics interacted with year fixed effects.

V. Results

A. Average Impacts on School Enrollment

Table 2 shows estimates of the effect of road construction on village school

­enrollment, using equation (1). Column 1 reports that a new paved road leads to a
7 percent increase in middle school enrollment in a village (95 percent ­confidence
interval: 4.1–9.9 percent). This effect corresponds to approximately three additional
students in middle school, given the sample mean of 39 students enrolled in mid-
dle school.19 In columns 2 and 3 of Table 2, we split the analysis by gender, and
find similar effects on the enrollment of girls and boys. Columns 4 through 6 show
comparable estimates using the level of middle school enrollment as the dependent
variable, rather than log enrollment.
To explore further the changes in school enrollment before and after a new road
is built, we regress log middle school enrollment on a set of relative time dummies
that indicate the number of years before or after road construction in the village. The
estimating equation is

(2) ​​Y​i, s, t​​  = ​  ∑ ​​​​ζτ​ ​​​(1​(t = ​t​  treatment

i, s​  ​  + τ)​)​  + ​γs,​ t​​  + ​η​i​​  + ​ϵ​i, s, t​​  ,​
τ∈(−5, + 5),τ ≠−1

19
This effect reflects a treatment period of 3.7 years after a road is built, on average. The estimate is a weighted
difference between enrollment in all treated years and enrollment in untreated years. Estimating a weighted linear
combination of relative treatment time dummies according to Borusyak and Jaravel (2017) delivers a very similar
treatment estimate of 0.06.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 359

Table 2—Impact of New Roads on Middle School Enrollment

Dependent variable All, log Girls, log Boys, log All, levels Girls, levels Boys, levels
(1) (2) (3) (4) (5) (6)
New road 0.070 0.060 0.056 2.558 1.331 1.227
(0.015) (0.012) (0.013) (0.537) (0.287) (0.284)
Observations 146,678 146,678 146,678 146,678 146,678 146,678
R2 0.80 0.81 0.80 0.79 0.77 0.78

Notes: The table reports panel estimates of the effect of new road construction on village-level log middle school
enrollment, estimated with equation (1). Column 1 presents the primary balanced panel specification. The depen-
dent variable in columns 2 and 3 is log middle school enrollment for girls and boys, respectively. Columns 4–6
repeat these three specifications using the level of middle school enrollment as the dependent variable. All spec-
ifications have state-year fixed effects and village fixed effects. Standard errors are clustered at the village level.

0.2
log middle school enrollment

0.1

−5 −4 −3 −2 −1 0 1 2 3 4 5

Time since new road

Figure 2. Impact of Roads on Middle School Enrollment: Treatment Effect Time Series

Notes: The figure shows coefficient estimates from a panel regression of log middle school enrollment on a set
of indicator variables indicating the number of years before or since a road was constructed, along with a set of
­state-by-year fixed effects and village fixed effects. The estimating equation is equation (2). Year 0 is the first year
in which a road was present when enrollment data were collected on September 30. Years t = −1 and t = −5
are omitted. Ninety-five percent confidence intervals are displayed around each point estimate. Standard errors are
clustered at the village level.

where ​τ​ indicates the year relative to when a new road was built (i.e., ​τ = − 1​is the
year before road construction). As in equation (1), we include ­state-by-year fixed
effects and village fixed effects.
In Figure 2, we plot the identified ​τ​ coefficients. We omit the relative time coeffi-
cients from the year before treatment and the first year available, following the sug-
gestion of Borusyak and Jaravel (2017). The regression above can only be estimated
with two relative time coefficients omitted because all villages in our sample are

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
360 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

eventually treated.20 We can therefore identify trend breaks, but cannot test either
average trends or ­pre-trends. The ­F-test of the p­ retreatment coefficients in Figure 2,
which tests for n­ onlinear ­pre-trends, is insignificant ( p = 0.94).
Figure 2 shows that increases in school enrollment correspond to the timing of
new road construction, and these effects appear to be persistent. The timing and
persistence of this change in enrollment suggests that the treatment effects are not
driven by labor demand shocks from road construction itself, which would occur as
the road is being built and disappear thereafter.

B. Robustness: Specifications and Sample Definitions

Table 3 shows that the estimated average effect on middle school enrollment is
robust to a range of empirical specifications and sample definitions. In column 1,
we allow for v­ illage-specific linear time trends to control for potential differential
trends across villages that receive a paved road in different years.21 In column 2, we
control for interactions between year fixed effects and baseline village characteris-
tics: population, share of irrigated land, number of schools, log middle school and
primary school enrollment, literacy rate, population share of Scheduled Castes, and
distance to nearest town. In column 3, we expand the sample to an unbalanced panel
by including villages with missing data in one or more years, and column 4 shows
the unbalanced panel estimates with ­village-specific time trends. In ­column 5, we
restrict the data to years after 2004, when the DISE data have the highest c­ overage
of villages and schools. Column 6 restricts the sample to a set of villages for which
we have four observations before and four observations after the completion of
road construction; the sample is limited to those observations, thus providing nine
observations per village. The estimates are similar in magnitude and statistical
significance. The stability of the treatment effect suggests that these estimates are
not driven by different types of villages being treated at different times. Online
Appendix Table A2 reports specifications from Table 2, with ­district-by-year fixed
effects, and shows similar estimates.
To verify that ­p-values are estimated correctly, we run a randomization test. In
the spirit of the Fisher Randomization Test, we randomly generate a placebo year
of road completion for each village, and then estimate equation (1) as if the placebo
year were the treatment year. We run this estimation 1,000 times. Online Appendix
Figure A2 shows the distribution of ​β​, the placebo impacts of a new road on log
middle school enrollment growth, which gives us a ­nonparametric distribution of
test statistics under the sharp null hypothesis. The placebo estimates are centered
around zero and, consistent with Table 2, none of the thousand estimates attains

20
Borusyak and Jaravel (2017) shows that event study designs where all groups are eventually treated can be
identified only up to a linear trend in relative time. For instance, an upward linear trend in enrollment could either
be described by equation (2) with linearly increasing time fixed effects, or with linearly increasing relative time
effects. McKenzie (2006) makes a similar point by arguing that without normalization, only second differences in
relative time effects can be identified. The standard d­ ifference-in-differences specification (equation (1)) has an
implicit normalization with zero ­pre-trend.
21
We use village time trends as a robustness check, rather than in the main specification because of the
­possibility that the time trends in part pick up the effects of the new road over time (Wolfers 2006). This said, all
results presented below are similarly robust to the inclusion of village time trends.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 361

Table 3—Impact of New Roads on Middle School Enrollment: Robustness

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

New road 0.058 0.058 0.086 0.078 0.053 0.041
(0.012) (0.014) (0.013) (0.013) (0.013) (0.009)

State-year fixed effects Yes Yes Yes Yes Yes Yes

Village fixed effects Yes Yes Yes Yes Yes Yes
Village time trends Yes No No Yes No No
Baseline variables No Yes No No No No
× year dummies
Panel sample Balanced Balanced Unbalanced Unbalanced Balanced 4 years
post-2004 pre/post
Observations 146,678 142,748 237,281 237,281 115,247 148,910
R2 0.91 0.83 0.76 0.88 0.87 0.84

Notes: The table reports panel estimates of the effect of new road construction on village log middle school
­enrollment, estimated with equation (1). Estimates are analogous to those in Table 2, with the following modifica-
tions. Column 1 adds a separate linear time trend for each village. Column 2 adds interactions between year fixed
effects and each of the following continuous village-level variables measured at baseline: population, number of
schools, log middle and primary school enrollment, literacy rate, population share of scheduled castes, irrigated
land share, and distance to nearest town. Column 3 uses an unbalanced panel, adding additional villages that do not
have data in all years. Column 4 adds a village time trend to the unbalanced panel specification. Column 5 restricts
the sample to years 2005 or later. Column 6 includes data only for four years before each road is built and four
years after. Different years are thus included for different villages, but each village has nine observations. Due to
data availability, the sample in column 6 only includes villages with roads built between 2006 and 2012. All spec-
ifications have state-year fixed effects and village fixed effects. Standard errors are clustered at the village level.

our main estimate of the effect of a new road on log enrollment (0.07 increase in
log enrollment).

C. Robustness: Regression Discontinuity

In this section, we present regression discontinuity estimates of the impact of

new roads on schooling. Under PMGSY road construction guidelines, states were
instructed to first target villages with populations greater than 1,000 in the popula-
tion census, and then villages with population greater than 500. Only some states
followed these guidelines, however, and even then, states followed the guidelines
to different degrees because there were often several conflicting guidelines.22 In
states where there were few unconnected villages with populations over 1,000, they
tended to use the ­500-person threshold immediately. In most states, construction
proceeded in villages both above and below the population threshold simultane-
ously, but there were more villages treated above the threshold, and these were
treated sooner. Population above a guideline threshold is therefore an imperfect pre-
dictor of treatment status.
Figure 3 shows the relationship between the probability of receiving a new road
by 2011 and the population relative to the treatment threshold. There is a clear

22
For example, under certain circumstances, proximate habitations could pool their populations to exceed this
cutoff; we do not observe where this took place. We met several times with the National Rural Roads Development
Agency, the national coordinating body for the program, to identify the set of states that adhered to program
guidelines and which eligibility thresholds were used. The states in the sample are Chhattisgarh, Gujarat, Madhya
Pradesh, Maharashtra, Odisha, and Rajasthan.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
362 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

0.2

New village road

−0.2

−200 0 200

Village population minus treatment threshold

Figure 3. Regression Discontinuity First Stage: Share of Villages Treated by Population

Notes: The graph plots the conditional expectation function of an indicator variable indicating that a village has
received a road before 2011, conditioning on the village population as reported in the 2001 Population Census. Each
point represents the mean of all villages in the given population bin (328 villages per bin). Population has been
centered around the state-specific threshold used for road eligibility, which is either 500 or 1,000 depending on the
state. Points to the right of the center line represent villages with a higher prioritization under PMGSY, according
to program rules.

d­ iscontinuity in treatment status at the population threshold. By contrast, there is no

discontinuous change in the density of villages on either side of the cutoff, nor in
characteristics of villages prior to road construction.23
We estimate the impacts of road construction using the following implementation
of a local linear estimator:

(3) ​ln​(​Y​i, s, t​​)​  = ​γ​1​​  1{

​ ​pop​i, s, 2001​​  − P ≥ 0}​  + ​γ​2​​​(​pop​i, s, 2001​​  − P)​

 + ​γ3​ ​​​(​pop​i,s, 2001​​  − P)​  × 1​{​pop​i, s, 2001​​  − P ≥ 0}​

  

 + ​γ4​ ​​  ln​(​Y​i, s, 2002​​)​  + λ ​Xi,​ s, 2001​​  + ​ηs​​​  + ​υi,​ s​​  .​

Here, ​​Yi​, s, t​​​is log enrollment in village i​​, region s​​, and time t​​; ​P​is the popula-
tion ­threshold; ​​pop​i, s, 2001​​​is baseline village population (i.e., the running variable);
​​X​i, s, 2001​​​is a vector of village controls measured at baseline; and η​ ​​ s​​​is a region fixed

23
To test this formally, we fit a ­nonparametric function to the village population distribution, with allowance
for a discontinuity at the treatment threshold (McCrary 2008); the ­p-value testing the null of no discontinuity is
0.31. Online Appendix Figure A3 presents the population histogram and the graphical rendering of the McCrary
Test. Online Appendix Table A3 and Figure A4 present regression discontinuity estimates and graphs showing that
baseline village covariates do not vary systematically at the treatment threshold.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 363

Panel A. Reduced-form and first-stage Panel B. log middle school enrollment

estimates by year by population (2011–2015)
1

log middle enrollment,

0.2

mean 2011–2015
Point estimate

0.5

0
0

First stage

−0.2
Reduced form
−0.5
2005 2010 2015 −200 0 200
Village population minus treatment threshold

Figure 4. Regression Discontinuity Impacts of New Roads on log Middle School Enrollment Growth

Notes: Panel A shows reduced-form and first-stage estimates from equation (3), estimated on each sample year
from 2003 to 2015. Each square and solid error bar describes a single estimate from equation (3), where the depen-
dent variable is an indicator taking the value one if a village received a new road before the year on the x-axis. The
diamonds and dashed error bars describe the reduced-form regression discontinuity estimate of the effect of being
above the population threshold on village log middle school enrollment. Error bars show 95 percent confidence
intervals. Panel B plots the conditional expectation function of average log middle school enrollment between
2011–2015. Population is centered around the state-specific threshold used for program eligibility, which is either
500 or 1,000. Each point represents the mean of approximately 328 villages in the given population bin. Estimates
in both panels control for baseline log middle school enrollment, literacy rate, number of primary and middle
schools, the log number of nonfarm jobs in the village, and district fixed effects.

effect.24 The change in the outcome variable across the population threshold P ​ ​ is
captured by ​​γ​1​​​. The population controls allow for different slopes on either side of
the treatment threshold. We limit the sample to populations close to the treatment
threshold, using an optimal bandwidth calculation (Imbens and Kalyanaraman
2012).
Panel A of Figure 4 shows ­first-stage and r­educed-form regression discon-
tinuity estimates for all sample years. The ­first-stage estimates show that the
population threshold rule begins to be applied around 2007 and stabilizes in
importance from 2011 to 2015, during which years villages just above the thresh-
old are ­20–25 ­percentage points more likely to have received new roads. The
­reduced-form estimates on log middle school enrollment follow a similar pattern,
ramping up in 2007 and s­ tabilizing in 2011. To maximize power, we estimate the
regression discontinuity on the pooled set of enrollment estimates from 2011 to
2015, ­clustering equation (3) at the village level to account for serial correlation.
Panel B of Figure 4 plots log middle school enrollment as a function of population
relative to the treatment threshold, which shows an increase in enrollment above
the treatment threshold.

24
For control variables, we include baseline log enrollment, the literacy rate, number of primary schools,
n­ umber of middle schools (all from the 2001 Population Census), and the log number of n­ onfarm jobs in the village
(from the 1998 Economic Census).

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
364 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

Table 4—Impact of New Roads on Middle School Enrollment Growth: Regression

Discontinuity Estimates

First stage Reduced form IV

(1) (2) (3)
Panel A. RD estimates
Above population threshold 0.239 0.108
(0.015) (0.066)
New road by 2011 0.450
(0.276)

Observations 55,271 55,271 55,271

R2 0.26 0.28 0.28

Panel B. Placebo RD estimates

Above population threshold 0.014 0.009
(0.011) (0.059)

Observations 56,219 56,219

R2 0.27 0.25

Notes: Panel A shows regression discontinuity estimates of the impact of new road construc-
tion on log village middle school enrollment, estimated with equation (3). The sample includes
all villages and enrollment data from 2011 to 2015. Standard errors are clustered at the village
level to account for serial correlation. Column 1 reports first-stage estimates of the effect of
being above the state-specific population threshold (that defines road program eligibility) on
the probability of receiving a new road before 2011. Column 2 shows a reduced-form regres-
sion discontinuity estimate of the impact of being above the population eligibility threshold
on log middle school enrollment. Column 3 shows the instrumental variable estimate of the
impact of a new road on village log middle school enrollment. Panel B shows a placebo test
consisting of the same specification in columns 1 and 2 of panel A, but in the set of states
that did not adhere to PMGSY rules regarding the population eligibility threshold. All speci-
fications control for baseline log middle school enrollment, literacy rate, number of primary
schools, number of middle schools, and the log number of nonfarm jobs in the village.

In Table 4, panel A presents regression discontinuity estimates for the pooled

­2011–2015 sample.25 Column 1 reports the fi ­rst-stage estimate, where the
­dependent variable is a ­village-level indicator equal to one if a village received a
road. Thirty-three percent of villages in the sample received new roads; a village
just above the population treatment threshold is 24 percentage points more likely to
receive a new road. Column 2 reports the ­reduced-form impact on log middle school
enrollment from crossing the population threshold. Column 3 presents the IV esti-
mate, which yields a large but imprecisely estimated impact of road construction on
middle school enrollment ( p = 0.103).
The regression discontinuity estimate is considerably larger than the panel
­estimate, but it is also substantially less precise. While it is possible that the local
average ­treatment effect of roads on enrollment for regression discontinuity c­ omplier
­villages is substantially higher than for villages in the ­difference-in-difference
sample, we note that the 95 percent confidence interval of the regression
­

25
Online Appendix Figure A5 shows regression discontinuity estimates for each year from 2010 to 2015, as
well as the pooled ­2011–2015 estimate, under the optimal bandwidth and alternate bandwidths that are 25 percent
higher and lower.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 365

discontinuity estimate includes the panel estimate, and we are hesitant to put a large
weight on the specific point estimate.26
Villages may otherwise differ across the population thresholds, and so as
a ­placebo exercise we estimate equation (3) for states that did not follow the
PMGSY population threshold guidelines.27 Panel B of Table 4 shows that there is
no substantive first stage in these states (column 1) and a r­educed-form treatment
effect close to zero (column 2). This provides reassurance that villages above the
population threshold would not have otherwise experienced differential changes in
school enrollment.
The regression discontinuity estimates corroborate the results from the main
panel specification, indicating higher middle school enrollment following road
­construction. The strength of the regression discontinuity approach is its reliance
on few assumptions for causal inference, but the power of the test is limited by
­imperfect compliance, as well as the restriction of the sample to villages close to
threshold populations in states that followed the allocation rules. These factors also
make the regression discontinuity estimates less representative of impacts across
India. We therefore focus on the panel specifications in the section on treatment
heterogeneity below.

D. Average Impacts on School Achievement

Increasing middle school enrollment may not directly translate into greater
l­ earning, especially if school quality is low or if there is increased school crowding.
To measure student learning, we estimate impacts on student examination outcomes.
Table 5 presents panel estimates of the impact of new roads on a set of dependent
­variables describing students’ e­ xam-taking decisions and exam performance. We
focus on middle school ­end-of-year exams, which were required to certify com-
pletion of middle school. Column 1 shows the estimated effect of roads on the log
number of students who appear for the exam. Column 2 shows the effect on the
log number of students who pass the exam, and column 3 shows the effect on the
log number who score high marks.28 For exam appearance and exam passing, we
find similar effects to the enrollment effects: 6 percent more students take and pass
exams in villages after the construction of a new paved road.29 We find a positive but
smaller 3 percent increase in the number of students scoring high marks.

26
Indeed, in Section VI we find subgroup estimates approaching this level in villages where we expect treat-
ment effects to be particularly large. In online Appendix Table A4, we estimate panel specifications on samples of
villages that are similar to the regression discontinuity sample. While some of these have larger point estimates,
note that it is impossible to set the sample to the regression discontinuity complier villages, as we do not know
which villages above the population threshold would be untreated at lower populations, and which villages below
the threshold would be treated at higher populations. The regression discontinuity sample also includes villages that
never received roads, whereas our main panel estimates use only villages that received roads at some point.
27
Major states that built roads under PMGSY but did not follow program guidelines include Andhra Pradesh,
Assam, Bihar, Uttar Pradesh, and Uttarakhand.
28
Sample size is smaller for the exam estimates than for enrollment estimates because we were only able to
obtain examination results for all states in our sample for the years ­2004–2009. In each case, we report effects
on the log number of students plus one. The estimates are similar for an unbalanced panel.
29
The number of students achieving these exam outcomes is smaller than the enrollment effects because for
every ten students enrolled in the eighth grade, only six appear for the exam, five pass the exam, and two pass the
exam with distinction.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
366 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

Table 5—Impact of New Roads on Middle School Completion Examinations

Exam taken Exam passed High exam score

(1) (2) (3)
New road 0.060 0.058 0.035
(0.019) (0.019) (0.014)

State-year fixed effects Yes Yes Yes

Village fixed effects Yes Yes Yes
Panel sample Balanced Balanced Balanced

Observations 32,239 32,239 32,239

R2 0.73 0.72 0.61

Notes: The table reports panel estimates of the effect of new road construction on village-level
school examination performance, estimated with equation (1). All columns use a balanced
panel specification, analogous to column 1 in Table 2. The dependent variable in columns 1
through 3 is, respectively: (1) the log number of students sitting for the middle school comple-
tion examination; (2) the log number of students who pass this exam; (3) the log number of
students who pass this exam with high marks. All specifications have state-year fixed effects
and village fixed effects. Standard errors are clustered at the village level.

The estimated impacts on exam outcomes reflect the net impact on student
achievement and can be interpreted in two ways. The first possibility is that the stu-
dents induced to stay in school take and pass exams at the same rate as ­non-marginal
students (but receive fewer high marks), and there are no effects on the exam per-
formance of n­ on-marginal students. Alternately, the marginal students who were
induced to stay in middle school could do worse on exams (perhaps because they are
of lower ability than students not on the margin of dropping out), but students who
would have stayed in school regardless of road construction are now performing bet-
ter on exams. N
­ on-marginal students could perform better, for example, if they begin
to perceive higher returns to human capital accumulation due to increased access to
labor markets outside the village. It is difficult to disentangle these two scenarios.
Under both interpretations, we can reject the possibility that school enrollment is
increasing without corresponding increases in academic achievement.

E. Impacts on Primary School Outcomes

In this section, we explore impacts on primary school enrollment and exam scores.
In Table 6A, columns 1 through 3 show the estimated d­ ifference-in-differences
impact of road construction on log primary school enrollment. We do not find
effects on primary school enrollment, consistent with our expectation that children
under the age of 12 have few labor market opportunities. Columns 4 and 5 show
­reduced-form and IV results from the regression discontinuity estimation, which
also do not indicate an impact on primary school enrollment. The precision of
the regression discontinuity estimate is much lower, however, so we cannot reject
meaningful regression discontinuity impacts in either direction. The corresponding
regression discontinuity figure is in panel A of online Appendix Figure A6.
Table 6B shows estimated impacts on primary school completion exam outcomes.
We find weakly positive impacts on student exam outcomes with point estimates

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 367

Table 6A—Impact of New Roads on Primary School Outcomes: Primary School Enrollment

Panel Reduced form IV

(1) (2) (3) (4) (5)
New road −0.005 −0.004 −0.005 0.033
(0.004) (0.003) (0.005) (0.086)
Above population threshold 0.008
(0.020)

Observations 146,678 146,678 237,281 66,663 66,663

R2 0.87 0.92 0.88 0.30 0.30

Notes: The table reports estimates of the effect of new road construction on village log primary school enrollment.
Columns 1 through 3 present panel estimates, and columns 4 and 5 present regression discontinuity estimates.
Column 1 presents the main balanced panel specification. Column 2 adds village-specific time trends, and column 3
repeats the main specification in the unbalanced panel. Column 4 shows the reduced-form estimate of the effect
on log primary school enrollment growth of being just above the eligibility threshold, and column 5 presents the
regression discontinuity IV estimates of the impact of the new road. Standard errors are clustered at the village level.

Table 6B—Impact of New Roads on Primary School Outcomes: Primary School

Completion Examinations

Exam taken Exam passed High exam score

(1) (2) (3)
New road 0.028 0.021 0.024
(0.016) (0.016) (0.017)
State-year fixed effects Yes Yes Yes
Village fixed effects Yes Yes Yes
Panel sample Balanced Balanced Balanced

Observations 31,671 31,671 31,671

R2 0.73 0.71 0.61

Notes: The table reports estimates of the effect of new road construction on village log primary
school enrollment. Columns 1 through 3 present panel estimates, and columns 4 and 5 present
regression discontinuity estimates. Column 1 presents the main balanced panel specification.
Column 2 adds village-specific time trends, and column 3 repeats the main specification in the
unbalanced panel. Column 4 shows the reduced-form estimate of the effect on log primary
school enrollment growth of being just above the eligibility threshold, and column 5 presents
the regression discontinuity IV estimates of the impact of the new road. Standard errors are
clustered at the village level.

between 2 and 3 percent, but they are of marginal ­statistical ­significance. The results
are suggestive of increased effort among enrolled c­ hildren, which could be due to
anticipated increases in attending middle school or a­ nticipated increases in returns
to education in the labor market.

VI. Mechanisms

A. Human Capital Investment Incentives

In this section, we examine the mechanisms underlying the estimated average

impact of new rural roads on human capital accumulation. Our analysis is guided
by the conceptual framework outlined in Section I. We begin by focusing on three
primary channels and sources of heterogeneity: a negative impact on human capital

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
368 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

investment through increased opportunity costs of schooling, a positive impact on

human capital investment through increased returns to education, and an impact
on human capital investment through income or liquidity effects. We analyze the
combined impact of income and liquidity effects because, given the available data,
it is difficult to distinguish between the effects of higher lifetime income and greater
­cash-in-hand. Our estimation focuses on identifying subsets of villages where each
mechanism is likely to be especially prominent.
To examine these mechanisms, we begin by assuming that reductions in trans-
portation costs will lead to factor price equalization: when a rural village receives
a new road, its wages and returns to education will adjust toward the wages and
returns in the broader geographic area. If the ­low-skill wage gap between the village
and surrounding market is high, the village ­low-skill wage will rise more than if the
­low-skill wage gap is small. We therefore expect the largest increases in the oppor-
tunity cost of schooling to occur in places with the largest gaps in l­ow-skill wages
between the village and its surrounding labor market. We proxy for the expected size
of the opportunity cost effect with the ­district-level ­urban-rural wage gap, the most
granular level at which wages can be calculated. We use data on urban and rural
wages from the fifty-fifth round of the National Sample Survey (NSS), undertaken
in ­1999–2000, the last NSS round before any PMGSY roads were built.
To proxy for the expected size of the returns to education effect, we again aim
to identify the difference in returns to education between each village and its
regional labor market. The underlying assumption is that a new rural road will shift
the returns to education in a village toward the returns to education in the broader
regional labor market. We calculate d­ istrict-level returns to education by running
Mincerian regressions at the district level, separately for individuals in rural and
urban areas, using data from the fifty-fifth round of the NSS. We call this difference
the ­urban-rural returns gap, or the skill premium gap.30 We assume the returns to
education effect is stronger when this skill premium gap is higher.
Finally, to proxy for the importance of income and liquidity effects, we assume
that households with few assets are more likely to be liquidity constrained, and
that a given change in wages for these households has a larger income effect. We
measure average baseline assets at the village level using data from the 2002 Below
Poverty Line Census. We define a village as having low assets (and thus high
­potential income and liquidity effects) if the share of households reporting zero
durable assets is above the sample median.31 Similarly, we define binary indicator
measures for the opportunity cost proxy and the returns to education proxy, based
on whether each proxy is above the sample median.
We then estimate our previous panel regression, including additional ­interaction
terms between the treatment indicator for road construction and the i­ndicator
variable for each of the three mechanisms. If the estimated impact of road
­
­construction ­varies with our proxy measure, and the interaction term is important in

30
​​ 2​​  , and
Specifically, in each district we regress log wage for working individuals on years of education, age, age​​ 
the log of household land owned, separately for urban and rural locations. Mincerian returns are minimally affected
by alterations to this specification, such as excluding land owned or including state fixed effects. We exclude
­districts with no urban data.
31
The surveyed assets are a radio, a television, a telephone, and a motorcycle.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 369

Table 7—Impact of New Roads on Middle School Enrollment: Treatment Heterogeneity

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

New road 0.074 0.115 0.049 0.061 0.073
(0.017) (0.024) (0.023) (0.024) (0.035)
New road × high opportunity cost effect −0.085 −0.088
(0.034) (0.034)
New road × high returns to education effect 0.053 0.061
(0.034) (0.034)
New road × high income/liquidity effect 0.026 0.031
(0.034) (0.034)

State-year fixed effects Yes Yes Yes Yes Yes

Village fixed effects Yes Yes Yes Yes Yes
Panel sample Balanced Balanced Balanced Balanced Balanced

Observations 111,580 111,580 111,580 111,580 111,580

R2 0.81 0.81 0.81 0.81 0.81

Notes: The table reports panel estimates of the effect of new road construction on village log middle school enroll-
ment, interacted with binary district-level measures of different potential mechanisms. The size of the opportunity
cost effect is proxied by the district-level mean low-skill urban wage minus the mean low-skill rural wage. The size
of the returns to education effect is proxied by the difference between the urban and rural Mincerian returns to one
additional year of education. The size of income and liquidity effects are proxied by the share of households in a
village reporting zero assets in 2002. These interactions take the value of one if the underlying variable is above
the value of the median village. The specifications use equation (1). All columns use a balanced panel specifica-
tion, analogous to column 1 in Table 2. Column 1 repeats the main specification without interactions in the sample
with nonmissing interaction variables. Columns 2 through 4 show the effects of the individual interaction terms,
while column 5 jointly estimates all interaction terms. Wage and education data come from the fifty-fifth round
of the NSS Employment and Unemployment Survey (1999–2000), and asset data are from the Below Poverty
Line Census (2002). All specifications have state-year fixed effects and village fixed effects. Standard errors are
clustered at the village level.

magnitude, it p­ rovides suggestive evidence of that mechanism being an important

channel through which new roads affect schooling decisions.
Table 7 shows the results from estimating these interactions. Column 1 repeats
the main specification, without interaction terms, in the sample for which each
­interaction term is measured.32 Columns 2 through 4 include each interaction term
separately, and column 5 includes the three interaction terms together.
The estimated interaction effects are consistent with the predictions from a
­standard model of human capital investment. Road construction has the smallest
effects on middle school enrollment in districts where these roads are expected to
most raise the opportunity cost of schooling. The largest effects of road construction
on middle school enrollment are in districts where road connections are expected
to raise the skill premium the most and to have the largest income and liquidity
effects. The opportunity cost effect interaction is strongly statistically significant
( ​p < 0.01​), the returns to education effect interaction is marginally statistically
­significant ( ​p = 0.08​), and the income/liquidity effect interaction is in the expected
direction but statistically insignificant ( ​p = 0.37​ ).33 The greater magnitude of the
opportunity cost effect may be in part because the u­ rban-rural wage gap is much
larger than the ­urban-rural skill premium gap (see online Appendix Table A1).

32
This analysis excludes districts without NSS data for both urban and rural areas in ­1999–2000.
33
For completeness, online Appendix Table A5 shows results by quartile of each mechanism proxy.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
370 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

Table 8—Treatment Heterogeneity in Estimated Road Impacts: Subgroup Estimates

Returns to Income/liquidity
Opportunity cost effect education effect effects Treatment estimate Number of villages
Low Low Low 0.032 2,527
(0.050)
Low Low High 0.138 1,029
(0.043)
Low High Low 0.189 987
(0.050)
Low High High 0.094 523
(0.051)
High Low Low −0.018 751
(0.049)
High Low High 0.014 844
(0.045)
High High Low 0.035 751
(0.045)
High High High 0.093 558
(0.054)

Notes: The table reports panel estimates of the effect of new road construction on village log middle school enroll-
ment, fully interacted with binary predictors of the size of the opportunity cost effect, the returns to education effect,
and the income/liquidity effect (as described in Table 7). The table shows linear combinations of interaction terms
that describe the treatment effect in each of the eight partitions of the data according to the binary mechanism indi-
cators. The specification is based on equation (1), with added treatment interactions. All specifications have state-
year fixed effects and village fixed effects. Standard errors are clustered at the village level.

While the estimated interaction effects are consistent with a standard model of
human capital investment, note that there could be other ­district-level characteristics
that influence the size of treatment effects and are correlated with the proxies we use.
For example, high r­ural-urban wage gaps are correlated with greater ­remoteness,
worse infrastructure, lower returns to education, and tend to be in the North. The
estimated interaction effects are robust to the inclusion of interactions with these
other variables, but there are myriad other unobserved d­ istrict-level characteristics.
Therefore, we see these estimates not as definitive but rather as suggestive indica-
tions of the mechanisms underlying the main estimates.
By fully interacting the three binary mechanism variables, we can obtain treat-
ment effects in eight partitions of the sample based on the model’s predictions.
Table 8 shows the treatment effect in each subgroup from the fully interacted
­regression. The point estimate is negative (but small and statistically insignificant)
only in the partition with a high opportunity cost effect, low returns to education
effect, and low income/liquidity effect, which represents 9 percent of all villages.
This is precisely the group where a standard model predicts that roads would have
the most adverse effects on education. Treatment effects are positive and significant
only in the 39 percent of villages where at least two of the mechanisms are favorable.
This treatment heterogeneity is consistent with the heterogeneity in results from
earlier research on impacts of labor demand shocks on school enrollment. Jensen
(2012) and Oster and Steinberg (2013) find that increasing availability of call center
jobs lead to increased schooling. The dominant mechanism in these studies is likely
an increase in the return to education, since spoken English is a requirement for these

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 371

call center jobs. Conversely, Shah and Steinberg (2017) finds that children are more
likely to attend school in drought years, when there are fewer a­ gricultural jobs avail-
able. The more important mechanism in that setting is likely to be an ­opportunity
cost effect, as the ­low-skill wage is declining when there are fewer agricultural jobs
available (or less need for children to substitute into home production while parents
work agricultural jobs). The small negative effects on schooling from India’s work-
fare MGNREGA program (Das and Singh 2013, Islam and Sivasankaran 2015, Shah
and Steinberg 2015, Adukia 2018, Li and Sekhri forthcoming) are likely driven by
similar mechanisms, as MGNREGA hires workers for ­labor-intensive construction
and increases labor demand for l­ ower skill workers. The same model helps to explain
the variation in estimated impacts on schooling of labor demand shocks outside of
India (e.g., fracking jobs in the United States (Cascio and Narayan 2015), export
manufacturing jobs in Mexico (Atkin 2016), and garment manufacturing jobs in
Bangladesh (Heath and Mobarak 2015)). In each case, individual schooling choices
appear to respond to the skill requirements of the labor market opportunities.
The heterogeneity of economic opportunities across India allow us to identify
both large positive effects in the places where the relative return to h­ igh-skill work
goes up the most, and neutral to weakly negative effects on schooling in places
where the relative return to ­low-skill work rises the most. But our finding that
­treatment effects are negative (and small) in only a small share of villages is a
striking result given the number of recent studies finding adverse impacts of new
labor market opportunities.

B. Other Potential Mechanisms

In this section, we explore several other mechanisms through which road con-
struction might impact schooling outcomes.

School Quality.—We have focused on how road construction affects the incen-
tives for human capital investment (i.e., changes in the demand for schooling),
though road construction could also affect school quality or the number of schools
available (i.e., changes in the supply of schooling). We use v­ illage-level DISE data
to examine impacts of road construction on the number of schools and on measures
of school quality, as proxied by physical characteristics of a school.
Online Appendix Table A6 reports no impact of road construction on the num-
ber of schools, and no systematic impact on school infrastructure characteristics.34
While a minority of specifications show statistically significant effects on school
infrastructure, none approach the size of the enrollment effects presented above.
The standard errors in column 1 rule out a 2 percentage point change in the presence
of any of these kinds of infrastructure. Overall, we do not find systematic evidence
that road construction substantially affects the number of schools and their physical
characteristics.

34
We find similar estimates if we weight the school infrastructure variables by the number of students attending
the school to reflect the share of children in a village who have access to a particular kind of infrastructure.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
372 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

Other Government Programs.—To alleviate concern that other government pro-

grams could have been using the same eligibility criteria as the road program, or
simultaneously implemented other programs along with roads, we use the regres-
sion discontinuity approach to test for appearance of other public goods in treated
villages. We observe village public goods only in the decennial population cen-
sus, and so we are unable to estimate panel regressions. Online Appendix Table A7
shows no discontinuity in the presence of schools, as above using DISE data, and
also shows no discontinuity in village access to electric power, a primary health
center, or a commercial bank. We can rule out a 1 percentage point increase in the
existence of primary or secondary schools, health centers and banks, and a 4 per-
centage point increase in middle schools and electrification status.

Migration.—We next explore whether the estimated increases in middle school

enrollment could be driven by increased migration into villages that receive roads,
or reduced outmigration from those villages. Note that we did not find impacts of
road construction on primary school enrollment (Table 6A), which would presum-
ably also be affected if the increase in middle school enrollment was driven by
migration responses to road construction.
To test for migration effects, we use the regression discontinuity specification to
examine impacts of roads on village population in 2011.35 The result is shown in
panel A of online Appendix Figure A6; there is no effect of the treatment threshold
on population. The point estimate is close to zero, and the 95 percent confidence
­interval rejects the net entry or exit of more than four people from a treated village.
Changing migration patterns are thus unlikely to explain the effects of roads on
school enrollment.

­Cross-Village Displacement.—Relatedly, we explore whether our estimated

impacts on school enrollment could be driven by displacement effects, in which
increased enrollment in treated villages is counterbalanced by decreased enroll-
ment in nearby villages. We calculate total middle school enrollment for all other
villages within a 3 km or 5 km radius of each village that received a new road.36
Using the panel specification, columns 1 and 2 of online Appendix Table A8 report
the estimated impact of roads on log middle school enrollment in these surround-
ing villages. We do not find impacts on school enrollment in these surrounding
villages; the 95 percent ­confidence interval rules out a 4 percent decrease in enroll-
ment in a 3 km radius of the village, and a 2 percent decrease in a 5 km radius.

School Accessibility.—Finally, we examine the possibility that road construc-

tion increases school enrollment by decreasing the students’ costs of traveling to
school.37 Children generally walk to village schools, though paved roads could

35
As with public goods, village population is measured only in the decennial censuses, so we cannot use the
panel approach here.
36
The average road built through the PMGSY program had a length of 4.4 km, and the average Indian village
has a diameter of 2.1 km.
37
For example, Muralidharan and Prakash (2017) finds that the provision of bicycles made girls more likely to
attend middle school and high school.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 373

make schools more accessible, especially during the rainy season. We explore this
possibility by estimating whether the impact of roads varies across villages that are
more or less dispersed geographically. We expect that impacts through increased
school accessibility would be more pronounced for villages in which children have
further to walk to school. We measure village dispersion using village surface area,
and divide the sample into villages with ­above-median and ­below-median surface
area per capita.38 Columns 3 and 4 of online Appendix Table A8 show that the esti-
mated impact of roads is similar across more dispersed and more dense villages. The
point estimates are similar and we cannot reject equality between them. This sug-
gests that a decreased cost of reaching school is not a main channel through which
road construction is impacting school enrollment.
Similarly, road construction may increase middle school enrollment in a village
by increasing its accessibility to children from nearby villages that do not have
a middle school. This is related to the potential for c­ross-village displacement,
discussed above, but this effect would not decrease middle school enrollment in
those nearby villages. To explore this mechanism, we calculate the total number
of ­school-age children within a 5 km radius of sample villages who were living in
­villages without middle schools.39 Columns 5 and 6 of online Appendix Table A8
show that estimated impacts of roads on middle school enrollment are similar across
villages with more or fewer ­underserved children in nearby villages; the point esti-
mate is slightly higher in villages that do not have many underserved children living
nearby. This provides suggestive evidence that the schooling increases we observe
are not originating from nearby villages without middle schools.

VII. Conclusion

High local transportation costs are a central feature of the lives of the very poor
around the world, leaving them isolated from external markets. Connecting remote
villages to h­ igh-quality transportation networks is a major goal of d­ eveloping c­ ountry
governments and international development agencies. These roads can bring access to
new economic opportunities; however, a concern is that increased access to ­low-skill
labor market opportunities could decrease investment in the human ­capital that is
central to ­long-run increases in living standards and broader economic growth.
We examine this ­ trade-off in the context of India’s flagship rural road
­construction program, which built local paved roads to 115,000 villages in India
between 2001 and 2015. These roads connected villages with nearby labor
­markets, potentially changing the incentives for investment in human capital. We
find that rural road construction increased adolescent schooling outcomes. Further,
we find that a standard model of human capital investment has important predic-
tive power for how schooling decisions respond differently across villages to the
increases in economic opportunity from rural road construction. We highlight the

38
We find similar results using village surface area.
39
We proxy for the number of middle ­school-aged children using the number of children aged ­0 – 6 in 2001,
as reported in the Population Census. We find similar results if we use total village population in villages without
middle schools.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
374 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

competing influences of opportunity cost effects, returns to education effects, and

income/liquidity effects.
Our analysis draws on the substantial heterogeneity in economic opportunities
across India, allowing us to identify large positive effects on schooling in places
where the relative return to h­ igh-skill work increases the most, as well as neutral
or weakly negative effects on schooling in places where the return to l­ow-skill
work increases the most. Notably, treatment effects are negative (and small) in
only a small subset of villages. Across most of rural India, local market integration
­substantially promoted increased investment in human capital.
Our paper also highlights an important but understudied impact of rural
­infrastructure investment. Investments in road improvements are usually premised
on their potential to bring economic growth to rural areas, with a focus on contem-
poraneous economic gains in those areas. If road construction leads to increased
investment in human capital in rural areas, then the ­long-run economic impacts will
be greater than s­ hort-run estimates suggest and will reflect human capital dividends
over subsequent generations.

REFERENCES

Adukia, Anjali. 2017. “Sanitation and Education.” American Economic Journal: Applied Economics
9 (2): 23–59.
Adukia, Anjali. 2018. “Spillover Impacts on Education from Employment Guarantees.” Becker Fried-
man Institute Working Paper 2018-33.
Aggarwal, Shilpa. 2018. “Do Rural Roads Create Pathways Out of Poverty? Evidence from India.”
Journal of Development Economics 133: 375–95.
Asher, Samuel E., and Paul M. Novosad. F
 orthcoming. “Rural Roads and Local Economic Develop-
ment.” American Economic Review.
Atkin, David. 2016. “Endogenous Skill Acquisition and Export Manufacturing in Mexico.” American
Economic Review 106 (8): 2046–85.
Atkin, David, Azam Chaudhry, Shamyla Chaudry, Amit K. Khandelwal, and Eric Verhoogen. 2015.
“Markup and Cost Dispersion across Firms: Direct Evidence from Producer Surveys in Pakistan.”
American Economic Review 105 (5): 537–44.
Baland, Jean-Marie, and James A. Robinson. 2000. “Is Child Labor Inefficient?” Journal of Political
Economy 108 (4): 663–79.
Banerjee, Abhijit, and Rohini Somanathan. 2007. “The Political Economy of Public Goods: Some
Evidence from India.” Journal of Development Economics 82 (2): 287–314.
Becker, Gary. 1964. Human Capital: A Theoretical and Empirical Analysis with Special Reference to
Education. New York: Columbia University Press.
Behrman, Jere R., Susan W. Parker, and Petra E. Todd. 2008. “Medium-Term Impacts of the Opor-
tunidades Conditional Cash Transfer Program on Rural Youth in Mexico.” In Poverty, Inequality,
and Policy in Latin America, edited by Stephan Klasen and Felicitas Nowak-Lehmann, 219–70.
Cambridge: MIT Press.
Borusyak, Kirill, and Xavier Jaravel. 2017. “Revisiting Event Study Designs.” https://scholar.harvard.
edu/files/borusyak/files/event_studies_may8_website.pdf.
Bryan, Gharad, Shyamal Chowdury, and Ahmed Mushfiq Mobarak. 2014. “Underinvestment in a
Profitable Technology: The Case of Seasonal Migration in Bangladesh.” Econometrica 82 (5):
1671–1748.
Casaburi, Lorenzo, Rachel Glennerster, and Tavneet Suri. 2013. “Rural Roads and Intermediated
Trade: Regression Discontinuity Evidence from Sierra Leone.” https://scholar.harvard.edu/files/
lorenzocasaburi/files/casaburi_glennester_suri_roads.pdf.
Cascio, Elizabeth U., and Ayushi Narayan. 2015. “Who Needs a Fracking Education? The Educational
Response to Low-Skill Biased Technological Change.” NBER Working Paper 21359.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
VOL. 12 NO. 1 ADUKIA ET AL.: EDUCATIONAL RESPONSES TO ECONOMIC OPPORTUNITY 375

Das, Shreyasee, and Abhilasha Singh. 2013. “The Impact of Temporary Work Guarantee Programs on
Children’s Education: Evidence from the Mahatma Gandhi National Rural Employment Guarantee
Act from India.” https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2368011.
Donaldson, Dave. 2018. “Railroads of the Raj: Estimating the Impact of Transportation Infrastructure.”
American Economic Review 108 (4–5): 899–934.
Donaldson, Dave, and Richard Hornbeck. 2016. “Railroads and American Economic Growth: A ‘Mar-
ket Access’ Approach.” Quarterly Journal of Economics 131 (2): 799–858.
Drèze, Jean, and Amartya Sen. 2013. An Uncertain Glory: India and Its Contradictions. Oxford:
Oxford University Press.
Edmonds, Eric V. 2006. “Child Labor and Schooling Responses to Anticipated Income in South
Africa.” Journal of Development Economics 81 (2): 386–414.
Edmonds, Eric V., and Nina Pavcnik. 2006. “International Trade and Child Labor: Cross-Country Evi-
dence.” Journal of International Economics 68 (1): 115–40.
Edmonds, Eric V., Nina Pavcnik, and Petia Topalova. 2010. “Trade Adjustment and Human Capital
Investments: Evidence from Indian Tariff Reform.” American Economic Journal: Applied Econom-
ics 2 (4): 42–75.
Evans, David K., and Anna Popova. 2016. “What Really Works to Improve Learning in Develop-
ing Countries? An Analysis of Divergent Findings in Systematic Reviews.” World Bank Research
Observer 31 (2): 242–70.
Gibson, John, and Susan Olivia. 2010. “The Effect of Infrastructure Access and Quality on Non-farm
Enterprises in Rural Indonesia.” World Development 38 (5): 717–26.
Glewwe, Paul, and Karthik Muralidharan. 2016. “Chapter 10—Improving School Education Out-
comes in Developing Countries: Evidence, Knowledge Gaps, and Policy Implications.” In Hand-
book of the Economics of Education, Vol. 5, edited by Eric A. Hanushek, Stephen Machin, and
Ludger Woessmann, 653–743. Amsterdam: North Holland.
Heath, Rachel, and A. Mushfiq Mobarak. 2015. “Manufacturing Growth and the Lives of Bangladeshi
Women.” Journal of Development Economics 115: 1–15.
Hine, John, Masam Abedin, Richard Stevens, Tony Airey, and Tamala Anderson. 2016. Does the Exten-
sion of the Rural Road Network Have a Positive Impact on Poverty Reduction and Resilience for the
Rural Areas Served? If So How, and If Not Why Not? A Systematic Review. London: EPPI-Centre,
Social Science Research Unit, UCL Institute of Education, University College London.
Imbens, Guido, and Karthik Kalyanaraman. 2012. “Optimal Bandwidth Choice for the Regression
Discontinuity Estimator.” Review of Economic Studies 79 (3): 933–59.
Islam, Mahnaz, and Anitha Sivasankaran. 2015. “How Does Child Labor Respond to Changes in
Adult Work Opportunities? Evidence from NREGA.” https://scholar.harvard.edu/files/mahnazis-
lam/files/nrega_mahnaz_anitha_1.pdf.
Jacoby, Hanan G. 2000. “Access to Markets and the Benefits of Rural Roads.” Economic Journal 110
(465): 713–37.
Jacoby, Hanan G., and Bart Minten. 2009. “On Measuring the Benefits of Lower Transport Costs.”
Journal of Development Economics 89 (1): 28–38.
Jensen, Robert. 2012. “Do Labor Market Opportunities Affect Young Women’s Work and Family
Decisions? Experimental Evidence from India.” Quarterly Journal of Economics 127 (2): 753–92.
Khandker, Shahidur R., Zaid Bakht, and Gayatri B. Koolwal. 2009. “The Poverty Impact of Rural
Roads: Evidence from Bangladesh.” Economic Development and Cultural Change 57 (4): 685–
722.
Khandker, Shahidur R., and Gayatri B. Koolwal. 2011. “Estimating the Long-Term Impacts of Rural
Roads: A Dynamic Panel Approach.” World Bank Policy Research Working Paper 5867.
Li, Tianshi, and Sheetal Sekhri. Forthcoming. “The Unintended Consequences of Employment-Based
Safety Net Programs.” World Bank Economic Review.
McCrary, Justin. 2008. “Manipulation of the Running Variable in the Regression Discontinuity
Design: A Density Test.” Journal of Econometrics 142 (2): 698–714.
McKenzie, David J. 2006. “Disentangling Age, Cohort, and Time Effects in the Additive Model.”
Oxford Bulletin of Economics and Statistics 68 (4): 473–95.
Miguel, Edward, and Michael Kremer. 2004. “Worms: Identifying Impacts on Education and Health
in the Presence of Treatment Externalities.” Econometrica 72 (1): 159–217.
Mu, Ren, and Dominique van de Walle. 2011. “Rural Roads and Local Market Development in Viet-
nam.” Journal of Development Studies 47 (5): 709–34.
Mukherjee, Mukta. 2012. “Do Better Roads Increase School Enrollment? Evidence from a Unique
Road Policy in India.” https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2207761.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms
376 AMERICAN ECONOMIC JOURNAL: APPLIED ECONOMICS JANUARY 2020

Muralidharan, Karthik, and Nishith Prakash. 2017. “Cycling to School: Increasing Secondary School
Enrollment for Girls in India.” American Economic Journal: Applied Economics 9 (3): 321–50.
Oster, Emily, and Bryce Millet Steinberg. 2013. “Do IT Service Centers Promote Enrollment? Evi-
dence from India.” Journal of Development Economics 104: 123–35.
Ranjan, Priya. 1999. “An Economic Analysis of Child Labor.” Economics Letters 64 (1): 99–105.
Shah, Manisha, and Bryce Millett Steinberg. 2015. “Workfare and Human Capital Investment: Evi-
dence from India.” NBER Working Paper 21543.
Shah, Manisha, and Bryce Millett Steinberg. 2017. “Drought of Opportunities: Contemporaneous and
Long Term Impacts of Rainfall Shocks on Human Capital.” Journal of Political Economy 125 (2):
527–61.
Shastry, Gauri Kartini. 2012. “Human Capital Response to Globalization: Education and Information
Technology in India.” Journal of Human Resources 47 (2): 287–330.
Wolfers, Justin. 2006. “Did Unilateral Divorce Laws Raise Divorce Rates? A Reconciliation and New
Results.” American Economic Review 96 (5): 1802–20.

This content downloaded from

35.154.106.89 on Fri, 06 Sep 2024 06:11:45 UTC
All use subject to https://about.jstor.org/terms