WPS8507


Policy Research Working Paper               8507




                 The Ecological Impact
            of Transportation Infrastructure
                              Sam Asher
                             Teevrat Garg
                             Paul Novosad




Development Economics
Development Research Group
July 2018
Policy Research Working Paper 8507


  Abstract
 There is a long-standing debate over whether new roads                             on local deforestation. In contrast, the highway upgrades
 unavoidably lead to environmental damage, especially forest                        caused substantial forest loss, which appears to be driven by
 loss, but causal identification has been elusive. Using mul-                       increased timber demand along the transportation corridors.
 tiple causal identification strategies, this paper studies the                     In terms of forests, last mile connectivity had a negligible
 construction of new rural roads to over 100,000 villages                           environmental cost, while expansion of major corridors had
 and the upgrading of 10,000 kilometers of national high-                           important environmental impacts.
 ways in India. The new rural roads had precise zero effects




 This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the
 World Bank to provide open access to its research and make a contribution to development policy discussions around the
 world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors
 may be contacted at sasher@worldbank.org.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
          The Ecological Impact of Transportation Infrastructure∗
                                              Sam Asher†

                                            Teevrat Garg‡

                                            Paul Novosad§

                                              June 2018




   ∗
     We are grateful for feedback on earlier versions of this paper from Prashant Bharadwaj, Jenn Burney,
Jonah Busch, Eric Edmonds, Paul Ferraro, Gordon Hanson, Maulik Jagnani, Erzo Luttmer, Erin Mansur,
Gordon McCord, and seminar participants at UC San Diego, George Washington University, and the 2017
CU Boulder Environmental and Energy Economics Workshop. Ryu Matsuura provided excellent research
assistance.
   †
     World Bank, sasher@worldbank.org
   ‡
     UC San Diego, teevrat@ucsd.edu
   §
     Dartmouth College, paul.novosad@dartmouth.edu
I   Introduction

Does human economic progress have an unavoidable environmental cost? This is a central
question for policymakers pursuing sustainable development and has been a long-standing
debate in both the conservation and the economics literature (Arrow et al., 1995; Grossman
and Krueger, 1995; Stern, Common and Barbier, 1996; Andreoni and Levinson, 2001; Foster
and Rosenzweig, 2003; Dasgupta, 2007; Alix-Garcia et al., 2013). A key pillar of economic
development is large-scale investment in transportation infrastructure that reduces the costs
of moving goods and people across space. Concern has been expressed about the potential
environmental cost of such investments, and of increased trade more generally (Copeland
and Taylor, 1994; Antweiler, Copeland and Taylor, 2001; Copeland and Taylor, 2004; Frankel
and Rose, 2005), but researchers have struggled to identify causal estimates of the impact of
transportation infrastructure on local environmental quality.
    The most omnipresent of transportation investments are roads. We focus on the impact
of road construction and expansion on forest loss as it is among the primary environmental
concerns associated with new road construction. Forest cover loss is globally and locally
important, generating global greenhouse emissions (IPCC, 2007; Jayachandran et al., 2017)
and local health externalities (Bauch et al., 2015; Garg, 2017). Emissions from global for-
est cover loss are comparable to those from the entire U.S. economy, or from the global
transportation sector (IPCC, 2007; U.S. Environmental Protection Agency, 2017).
    Because of the high cost and high expected return of roads, their placement typically de-
pends on various economic and political factors, making causal identification of their impacts
difficult. For example, new roads may be targeted to regions with expanding agricultural
land use; these roads may be a response to activities that are already damaging forest cover,
making it difficult to isolate the direct impact of the roads. While many earlier studies
have documented changes in forest cover following the construction of new roads, none have
addressed the endogeneity of road placement beyond the inclusion of control variables and
in a few cases, location fixed effects. Further, most of these studies have focused on large

                                              1
highways built into the Amazon rainforest; while these highways are important in terms of
potential deforestation, their impacts are of uncertain relevance for the set of potential rural
roads and highways that policy-makers in developing countries are considering today. The
majority of road projects in the decades ahead are likely to be last-mile roads to people not
currently connected to the road network and upgrades of existing transportation corridors
into modern highways.
   In this paper, we take advantage of a validated satellite-based measure of forest cover
(Vegetation Continuous Fields or VCF), which makes it possible to study the impacts of two
large-scale transportation projects in India. The first of these was an initiative to upgrade
two major transportation corridors: the 6000 km “Golden Quadrilateral” network (GQ)
connecting the country’s four largest cities, and the comparably-sized “North-South and
East-West” network (NS-EW) connecting the country’s four cardinal endpoints in a cross.
Both corridors were already used for cross-city transportation before 2000, but over the
following fifteen years they were upgraded into world class divided highways. The second
project was a rural road construction program, under which over 100,000 new paved rural
feeder roads were built, ten kilometers in length on average, providing new connections to
over 100 million rural residents. Each project has exceeded ten billion dollars in cost to date
and has caused a significant reallocation of local economic activity (Asher and Novosad,
2017; Ghani, Goswami and Kerr, 2016).
   Theoretically, the effect of road investments on local forest cover can be positive or
negative. New roads can increase forest cover loss by: (i) providing external markets for forest
resources, especially timber and firewood; (ii) providing external markets for agricultural
products, motivating extensification of agriculture into forested land; and (iii) increasing
the value of land for settlement and industry, resulting in forest clearing. On the other
hand, paved roads could also reduce forest cover loss by (i) improving local household and
industry access to substitutes for local forest resources, especially firewood; (ii) providing
access to external output and labor markets, lowering the relative returns to clearing forests


                                               2
for agricultural land as well as to harvesting other forest products such as firewood. Given
the substantially different nature of rural feeder roads and national highways, we can also
expect the importance of any of these channels to vary by the type of road.
   To evaluate the impact of rural roads, we first use a regression discontinuity approach, ex-
ploiting an implementation rule that discontinuously raised the probability of road construc-
tion in villages with population above an arbitrary threshold. Second, we use a difference-in-
differences specification that exploits the exact timing of road construction. Both approaches
show zero effects of new roads on forest cover. The estimates are precise; we can rule out
gains larger than 0.6% and losses greater than 0.2% in forest cover up to five years after
roads are completed. Further, we find zero effects for sample subgroups where we might
expect losses to be greater, such as villages with greater baseline forest cover or with very
poor or forest-dependent residents. We also find zero change in household firewood use in
treated villages. We do identify marginal (0.5%) reductions in forest cover during the road
construction period; these reductions are reversed soon after roads are completed, but there
is no evidence that forest cover continues to rise. We show that ignoring these construction
period effects could lead to biased impact estimates. These roads have no effect on forest
cover in spite of significantly altering economic opportunities for people in villages (Asher
and Novosad, 2017; Adukia, Asher and Novosad, 2017).
   Causal identification for impacts of highways is much more difficult than for rural roads,
because in almost all cases, new highways are small in number and are built along existing
transportation corridors. We take the approach of comparing changes in forest cover in areas
that are near and that are far from the new highways. While we do not have data covering
the period before the construction of the Golden Quadrilateral, the North-South/East-West
highway route provides a plausible counterfactual, in that it is a highway of comparable size
and importance that was announced simultaneously and on a similar construction schedule
to the GQ, but its construction was pushed back by approximately eight years due to bu-
reaucratic delays. Ghani, Goswami and Kerr (2016) take a similar approach in comparing


                                              3
these two networks to study the impacts of the GQ on manufacturing activity.1
    In sharp contrast to rural roads, we find that the highway upgrades have had substantial
negative effects on forest cover. Following construction of the GQ, we find a 20% decline in
forest cover in a 100 kilometer band around the highway, an effect that persists for at least
eight years. We find no change in forest cover along the NS-EW corridor until construction
accelerates in 2008, at which point we also observe local forest cover loss. The timing of
relative forest loss around the construction of each corridor supports a causal interpretation
of these estimates. Because forest cover in India is rising on average during the sample period,
these are net effects on forest cover, combining increases in deforestation and reductions in
afforestation.
    These highways appear to have depleted forest cover by increasing timber demand in
their vicinity, which has wide ranging effects into the hinterlands of the transport corridors.
Following the construction of the GQ, we find a substantial upward trend break in employ-
ment in proximate firms that use timber and wood as primary inputs, as well as employment
in logging firms. Additional tests reject the competing mechanisms; there are no increases in
agricultural land use or changes in local firewood consumption along the highway corridor.
    This paper makes two central contributions. First, we generate the first causal estimates
of the impact of large scale transportation infrastructure investments on natural resource
depletion.2 In so doing, we contribute to a long literature on the trade-offs and synergies
between economic development and environmental conservation.3
   1
      On the impacts of the Golden Quadrilateral on firms in India, see also Datta (2012) and Khanna (2016).
   2
      Many studies describe cross-sectional relationships between roads and forest cover or forest loss (Chomitz
and Gray, 1996; Angelsen and Kaimowitz, 1999; Pfaff, 1999; Cropper et al., 2001; Geist and Lambin, 2002;
Deng et al., 2011; Barber et al., 2014; Li et al., 2015; Dasgupta and Wheeler, 2016). A small number of studies
examine forest loss in areas with new roads but do not address the endogeneity of road placement (Pfaff
et al., 2007; Weinhold and Reis, 2008). The closest study to ours is ongoing work by Kaczan (2017), who
uses a difference-in-differences design similar to our first strategy (but does not look at highways), finding
that India’s new rural roads marginally increased forest cover. The differences may arise because Kaczan
(2017) does not distinguish between construction and post-construction periods, and includes villages that
never receive roads as part of the control group. We show in Section IV that both of these choices may lead
to biased treatment effects.
    3
      On the general relationship between economic development and the environment, see Den Butter and
Verbruggen (1994), Arrow et al. (1995), Grossman and Krueger (1995), Stern, Common and Barbier (1996),
Andreoni and Levinson (2001), Dasgupta et al. (2002), Foster and Rosenzweig (2003) and Stern (2004).


                                                       4
    Second, this is the first paper to show that the impact of roads on deforestation is a
function of which markets are being connected by those roads. Last-mile rural roads provide
connectivity to small local markets, facilitating exits from agriculture but without signifi-
cantly changing industry’s access to forest products (Asher and Novosad, 2017). In contrast,
highways dramatically change the geographic distribution of industry (Ghani, Goswami and
Kerr, 2016); in India at least, this appears to have important environmental consequences.
Our estimates are particularly relevant as the infrastructure agenda in Sub-Saharan Africa
and South and Southeast Asia is likely to prioritize exactly the kinds of infrastructure in-
vestments that we study here — new feeder roads and expansion of existing corridors — as
opposed to the large highways through virgin rainforest that have been the subject of much
of the earlier work on roads and deforestation.
    Finally, we make a methodological note in the literature on estimating impacts of infras-
tructure. Large-scale infrastructure often takes many years to build and involves significant
land clearing and economic activity during the construction process. In both our examina-
tion of highways and of rural roads, we find that forest loss begins during the construction
period; in either case, estimates based strictly on the timing of infrastructure completion
would underestimate the environmental impact of roads.
    The rest of the paper is organized as follows. The next section describes India’s rural road
and highway construction programs. In Section III, we describe the data on forest cover and
roads, as well as other secondary datasets used in our analysis. Section IV presents empirical
strategy and results describing the impact of rural roads on deforestation. Section V presents
the empirical strategy and impacts of highway expansions, and Section VI concludes.
On deforestation specifically, see Koop and Tole (1999), Burgess et al. (2012), Alix-Garcia et al. (2013),
and Jayachandran et al. (2017). Assun¸     ˜o et al. (2017) provide causal evidence that rural electrification
                                          ca
mitigated forest loss in Brazil. For an exhaustive review on drivers of deforestation, see Ferretti-Gallon and
Busch (2014). For a literature review on impacts of highways and rural roads on outcomes other than the
environment, see Asher and Novosad (2017).




                                                      5
II       Background: Road Construction Programs in India

In 1999 and 2000, the Government of India launched two major road construction programs
— one aimed at upgrading several national highway corridors and the other at connecting
the remainder of India’s population to the road network. Together, these programs marked
the largest expansion of road infrastructure in Indian history and came at a joint cost ex-
ceeding $50 billion. This section provides background information on both road construction
programs.

II.A       Rural Roads

In 2000, the Indian government launched the Pradhan Mantri Gram Sadak Yojana (PMGSY),
or the Prime Minister’s Village Roads Scheme. The primary objective of the program was to
provide new paved roads to previously unconnected villages, although in practice this also
involved upgrading low quality roads in already connected villages. By 2015, over 400,000
kilometers of new roads were built, providing access to the national road network to over
100 million rural people in over 100,000 villages.
     Rural road construction began toward the end of 2001 and was continuing steadily
through the end of the sample period in 2014 (See Appendix Figure A1). Villages were
selected for roads based on a set of guidelines issued by a national government body, the Na-
tional Rural Roads Development Authority. Notably, the program prioritized construction
of roads to larger villages; district-level implementation plans were to first target all villages
with populations greater than 1000, followed by villages with population greater than 500,
and finally those with population greater than 250.4
     The rules were applied on a state-by-state basis, allowing states to move from one thresh-
old to another on their own timelines. In practice, there were several other prioritization
     4
      Strictly speaking, the allocation was based on habitation population rather than village population. A
habitation is a smaller unit of aggregation than the village; there are between one and three habitations in
each village. In practice, habitation populations were pooled to the village level in many cases (see below).
We aggregate to the village level because neither additional data nor maps are available at the habitation
level.




                                                     6
guidelines and political patronage undoubtedly played a role, so that a village’s population
relative to the threshold significantly influenced its likelihood of receiving a road but was not
definitive. For instance, smaller villages could be connected if they were along the least-cost
path between larger prioritized villages, and proximate villages could combine their popu-
lations to attain the eligibility thresholds. For more details, see Asher and Novosad (2017)
and National Rural Roads Development Agency (2005).

II.B   National Highways

In 1999, the Indian government announced a plan to modernize its major highways, the
National Highways Development Project. The first component of the project was the up-
grading and widening of the Golden Quadrilateral highway corridor (henceforth, GQ), so
named because it connected the four major cities in India: New Delhi, Mumbai, Chennai
and Kolkata. The second component was a similar upgrading of the the North-South and
East-West corridor (NS-EW), which would connect the furthest corners of the country from
Srinagar in the north to Kanyakumari in the south, and from Porbandar in the west to
Silchar in the east.
   While the GQ and NS-EW projects were commissioned around the same time, the gov-
ernment prioritized the implementation of the GQ and construction of the NS-EW was
substantially delayed. Construction on the GQ began in 2001; 80% was completed by 2004
and 95% by 2006. In contrast, by 2006 only 10% of the NS-EW corridor was completed,
almost half of which was a set of highways which were shared with the GQ (Ghani, Goswami
and Kerr, 2016). By 2010, 72% of the NS-EW was completed, and 90% was completed by
2015. The delay in the construction of the NS-EW allows us to use the NS-EW corridor
as a counterfactual for changes in forest cover in the GQ corridor during and immediately
following substantial completion of the GQ.
   Before these highways were widened and upgraded, the GQ and NS-EW routes were
already significant transportation corridors, but their road quality and congestion were highly
variable. The upgrading of these networks dramatically improved their quality and reliability;

                                              7
these were the first major long-distance divided highway networks to be developed in India.
The construction of the GQ changed national supply networks and led to a substantial
reallocation of manufacturing firms into the GQ corridor (Datta, 2012; Khanna, 2016; Ghani,
Goswami and Kerr, 2016). The economic impact of the NS-EW corridor has so far been little
studied due to its later completion date.

III        Data

To estimate effects of new roads on forest cover, we combine five different national data
sources. We use a validated high resolution satellite-based measure of forest cover. Data
on rural roads come from the administrative implementation data generated by the rural
road construction program, and geographic data on new major highway networks come from
national highway maps. While these datasets form the basis of our core specifications, we
also use data from the 1991, 2001 and 2011 Population Censuses and 3rd through 6th rounds
of the Economic Census to control for location characteristics and explore mechanisms of
treatment effects. All of these are census datasets that describe the entire population of
India and are geocoded to the village, town and subdistrict levels. This section describes
the details of how we prepare and combine all of these datasets. Table 1 shows summary
statistics for all variables used.

III.A          Forest Cover

Detailed and reliable administrative records on forest cover and deforestation rarely exist,
especially in developing countries. Instead, we obtain high resolution time series estimates
of forest cover using a standardized publicly-available satellite-based dataset. Vegetation
Continuous Fields (VCF) is available at 250m resolution and provides annual tree cover
from 2000-2014 in the form of the percentage of each pixel under forest cover (Townshend
et al., 2011). For our primary specification, we define forest cover as the total log pixel value
plus one in a given geographic area.5 Results are robust to using the average percentage of
      5
          Results are robust to using the inverse hyperbolic sine transformation instead of log plus one.




                                                          8
forest cover in each village.
    The VCF measure is a prediction of the percentage of a pixel that is covered by forest,
generated from a machine learning model based on a combination of images from global
Landsat images and samples from higher resolution satellites. The measure employs not only
the visible bandwidth but also other bandwidths. For example, VCF uses thermal signatures
because forested areas tend to be cooler than non-forested plantation areas, allowing VCF
to (partially) distinguish between forest cover and plantations. To the extent that thermal
signatures and other correlates can distinguish forests from non-forest plantations, VCF
substantially improves upon the Normalized Differenced Vegetation Index (NDVI) that has
been widely used in understanding the causes of deforestation (for example, Foster and
Rosenzweig (2003)). For all analyses, we restrict the sample of villages to those that had
non-zero forest cover in 2000, a year predating the construction of all roads considered in
this research. This is also the earliest year that these forest cover datasets are available.6
    Some earlier studies have used the Global Forest Cover (GFC) dataset, which describes
baseline forest cover in the year 2000, and a binary indicator for the year of deforestation for
each pixel, at a 30m resolution, where a pixel is considered deforested if over 90% of 2000
forest was lost by a given year, or reforested if a pixel goes from zero forest in 2000 to positive
forest cover by 2012 (Hansen et al., 2013). While GFC and VCF are both based on the same
underlying Landsat images, GFC is less useful for the study of forest cover in India, because
forest change in India is not well summarized by a binary deforestation indicator. The VCF
measures suggest that forest cover rose 15% over the sample period, an estimate consistent
with official and international sources. Because most of these gains are in areas that had
some pre-existing forest, they are not recorded by GFC, nor are partial forest losses. We can
replicate GFC estimates using losses only in the VCF data, but they miss a significant share
of forest change in the sample period. Because 92% of villages are larger than the VCF cell
size, the resolution advantage of GFC would be minimal.
   6
     Fewer than 10% of villages have zero forest cover in 2000; 95% of these villages have less than 1% forest
cover in 2014; the mean of forest cover for pixels with non-zero forest is 12.76% in 2000.


                                                      9
    We matched forest cover data to village, town and subdistrict boundaries using geographic
boundary data purchased from ML InfoMap. In remote parts of India, we received only
settlement centroids rather than village boundaries. We generated Thiessen polygons for
these villages; all results are robust to excluding this set of villages.

III.B    Rural Roads

We scraped village-level administrative data describing the construction of rural roads from
the program’s online management portal.7 For each road, the data provide the names of
connected villages, the date when the contract for road construction was awarded, and the
date of road completion. While data were reported at the sub-village (habitation) level,
we aggregated the data to the village level to match our other data sources. We define
a village as treated if any habitation in the village was provided with a new road. The
data construction and scraping approach is described in detail in Asher and Novosad (2017).
The dataset describes over 100,000 new roads built between 2001 and 2014; we limit our
sample to areas with non-zero forest cover and no paved road in the baseline year, leaving
approximately 65,000 new roads in the analysis sample.8

III.C    Highways

Construction dates and geocoordinates for the Golden Quadrilateral and North-South and
East-West corridors were generously shared with us by Ghani, Goswami and Kerr (2016).
We linked these to the village, town and subdistrict polygons described above by calculating
straight line distances from polygon centroids to the nearest point on each highway.

III.D     Population and Economic Censuses

We matched all villages and towns from the 1991, 2001 and 2011 population censuses using a
combination of incomplete keys provided by the Registrar General and a set of fuzzy matching
   7
     The data is publicly available at http://omms.nic.in.
   8
     Results are robust to including upgrades and/or villages with no forest cover at baseline. These would
be expected to attenuate non-zero treatment effects, thus their exclusion if anything biases us against finding
zero effects.




                                                     10
algorithms based on village and town names. The population censuses describe village and
town public goods, amenities and household characteristics, including the primary source of
cooking fuel. Fuel use is reported as the share of households in a location using firewood
(68% of households at baseline), imported fuels (chiefly propane, 8%) or local nonwood fuels
(crop residue and dung, 22%) as a primary source of energy. Fuel use is reported at the
subdistrict level in 2001 and at the village level in 2011.
     The Economic Censuses are complete enumerations of all nonfarm establishments under-
taken in 1990, 1998, 2005 and 2013, including informal and non-manufacturing firms. We
matched these on village names to the three population censuses using a fuzzy matching
algorithm. The Economic Census reports total employment and industry for all firms. We
create variables describing total employment in (i) firms engaged in logging and (ii) firms
whose primary input is raw lumber, which include sawmilling and planing of wood, manu-
facture of wooden products such as furniture and wooden containers, manufacture of cork,
and manufacture of pulp and paper products. The industry categorization for the 2005
Economic Census places logging firms in the same industry category as firms engaged in the
conservation of forest plantations, management of forest tree nurseries and other afforestation
categories. We therefore exclude 2005 from analysis of employment in logging firms.

IV       Impacts of Rural Feeder Roads on Forest Cover

This section describes the impact of new feeder roads on local deforestation. The main
challenge to causal identification of the impacts of rural roads is endogeneity. Because roads
are costly to build, their placement is typically correlated with other factors that could
also be predictors of deforestation. For example, roads could be targeted to places that
are expected to grow or to places that are lagging economically. Road placement may also
depend on geographic (e.g. slope, terrain, soil quality) or political factors. Any of these
scenarios would bias OLS estimates of the effect of new roads on deforestation.9 Causal
     9
     Appendix Table A1 shows estimates from cross-sectional OLS regressions of village-level log forest cover
in 2001 on an indicator variable that takes the value one if a village has a paved road in 2001. While the
bivariate relationship is strongly negative and highly statistically significant, the estimate gets progressively


                                                      11
identification of the impact of new roads therefore relies on some kind of variation in road
placement or timing that is plausibly exogenous. To study the impact of rural roads, we
rely on (i) an implementation rule that led to a discontinuity in the probability of a village
getting a new road based on arbitrary population cutoffs; and (ii) variation in the specific
year that a targeted village was treated. We focus our analysis on forest cover in the vicinity
of connected villages; because newly connected rural villages are mostly small and isolated,
these roads are unlikely to have had important general equilibrium effects on more distant
areas.

IV.A       Rural Roads: Regression Discontinuity Specification

We begin by exploiting the eligibility rule that prioritized villages for new roads based on
arbitrary population thresholds. Given the imperfect compliance with these eligibility rules
(described in Section II), we employ a fuzzy regression discontinuity (RD) design. We limit
the RD analysis to states in which administrators adhered closely to population threshold
rules.10
    We use an optimal bandwidth local linear regression discontinuity specification (Imbens
and Lemieux, 2008; Imbens and Kalyanaraman, 2012; Gelman and Imbens, 2014) to identify
the change in forest cover caused by a new road at the treatment threshold. We use the
following two stage least squares specification:


T reatmentvds = γ0 + γ1 · 1(popvds ≥ Ts ) + γ2 (popvds − Ts ) + γ3 (popvds − Ts ) · 1(popvds ≥ Ts )

                 + νd +   vds                                                                             (1)

    F orestvds = β0 + β1 · T reatmentvds + β2 (popvds − Ts ) + β3 (popvds − Ts ) · 1(popvds ≥ Ts )

                 + µd + ηvds                                                                              (2)

closer to zero as we add village-level controls and fixed effects, implying substantial selection on observables
in the presence of roads. Selection on unobservables is plausibly also important, making the OLS estimates
unreliable for causal inference.
   10
      We identified these states with the help of officials at NRRDA. They include Chhattisgarh, Gujarat,
Madhya Pradesh, Maharashtra, Orissa and Rajasthan. The difference-in-differences analysis below uses all
states that built any roads in the sample period.


                                                     12
F orestvds is forest cover in village v , district d and state s, and T reatmentvds is an indicator
equal to one if a new road was built in village v . popvds is the population of village v and Ts is
the treatment threshold used in state s.11 µd and νd are district fixed effects; we find virtually
identical results with fixed effects at higher or lower geographic scales. We also add controls
for village characteristics in 2001, before any roads were built; like the fixed effects, these are
unnecessary for identification but improve precision. Controls include baseline forest cover,
indicators for village amenities (primary school, medical center and electrification), the log
of total agricultural land area, the share of agricultural land that is irrigated, distance in
kilometers from the nearest town, the illiteracy rate and the share of inhabitants that belong
to a scheduled caste. This is a cross-sectional regression where β1 identifies the effect of new
roads on forest cover in a given year. To further improve precision, we stack outcome data
from 2010 through 2013 and cluster standard errors at the village level.12
    Appendix Figure A2 shows regression discontinuity balance tests for a set of variables
measured in the baseline period; Appendix Table A2 presents the regression estimates on
these tests using Equation 1. None of the regression discontinuity estimates are significantly
different from zero at baseline. Appendix Figure A3 shows that the density of the running
variable is continuous around the treatment threshold (McCrary, 2008).
   11
       The treatment threshold varies with state because some states used a threshold of 500 and others
were using a threshold of 1000. States used the lower treatment threshold when they had few villages with
population over 1000 that did not already have roads. Officials at the National Rural Roads Development
Agency provided us with information on which states were using which cutoffs, which we then verified in the
data. Madhya Pradesh used both the 500 and 1000 treatment thresholds for roads built in the same period;
we include separate fixed effects for the set of villages in the neighborhood of each threshold. Because the
optimal regression discontinuity bandwidth is close to 100, there is no overlapping between these two groups.
Few villages around the lowest population threshold of 250 received roads so we do not use this threshold
for analysis.
    12
       Running the test in any of these years separately or for similar sets of years produces virtually identical
point estimates with marginally higher standard errors, and thus does not change any of our conclusions.
We exclude years before 2010 because the first stage for road construction is too weak; the majority of the
roads built before 2010 were built by states that did not follow the treatment threshold rules. The first stage
is comparable in size from 2010 through 2013 (see Panel B of Figure 1, which shows the first stage in each
year).




                                                       13
IV.B    Rural Roads: Regression Discontinuity Results

Figure 1 shows a graphical representation of the regression discontinuity estimates of the
impact of rural roads on forest cover. Panel A shows the first stage; the Y axis shows the
share of sample villages that received new roads by 2013 under PMGSY as a function of
their population relative to the treatment threshold. Villages above the threshold are about
16% more likely to receive new roads and the discontinuity is evident. Panel B shows the
first stage estimate separately for each outcome year; each point in the figure represents the
β1 coefficient from Equation 1, where the dependent variable takes the value one if a village
received a new road by the year indicated on the X axis. We can see that roads built before
2007 were not prioritized according to the population threshold rule; the first stage of the
RD becomes noticeable after 2008 and continues to rise until 2014.
   Panel C of Figure 1 plots village-level log forest cover in 2013 against the population
relative to the treatment threshold, in population bins. If roads significantly affected local
forest cover, we would expect to see a discontinuity at the treatment threshold analogous
to that in Panel A; no such treatment effect is evident. Panel D shows the reduced form
treatment effect of above-threshold population on forest cover in each year separately; as in
Panel B, each point is an estimate from a separate regression, where the dependent variable
is the log of forest cover for the year on the X axis. If the new rural roads significantly
affected forest cover, we would expect to see a change in the coefficient following 2008 when
administrators began to adhere to the population implementation rule. Instead, the effect is
very close to zero both before and after 2008, indicating that new rural roads had negligible
effects on forest cover.
   Table 2 shows analogous regression estimates where the dependent variable is average
forest cover from 2010 to 2013; standard errors are clustered at the village level. Column
1 shows the first stage estimate of a 16% increase in the probability of road treatment for
villages just above the eligibility threshold. Columns 2 and 3 confirm there is no reduced
form effect on either log or average forest cover. Columns 4 through 6 test for treatment

                                             14
effects in villages that might be expected to respond more to new roads. These are: villages
with above-median baseline forest cover (Column 4); villages with above-median population
shares of constitutionally described “backward” communities (Scheduled Tribes) who often
derive livelihoods from forests (Column 5); and villages with below median assets, who might
depend more on forests for fuelwood (Column 6). There is no evidence of impacts of roads
in any of these groups. Columns 7 and 8 show IV estimates on log and average forest cover.
The IV estimates respectively rule out a 0.14 gain and a 0.11 loss in log forest cover with 95%
confidence, or approximately a one percentage point change in average forest cover. Results
are robust to different controls or fixed effects and different bandwidth choices.13 Appendix
Table A4 uses the RD specification to show further that there are no changes in household
fuel use following completion of a new road.

IV.C       Rural Roads: Difference-in-Differences Specification

The regression discontinuity design estimates causal impacts of roads under minimal assump-
tions, but is limited to estimating a LATE in the neighborhood of the treatment threshold
in states that closely followed implementation rules on population thresholds. We can make
greater use of our data and obtain tighter treatment estimates using a difference-in-differences
specification that exploits the differential timing of road treatment in each village. For this
empirical test, we limit the sample of villages to those that received a road at some point
during the road construction program, and use outcomes in later-treated villages as a control
group for villages that were treated earlier. We specifically estimate the following equation:


            F orestvdt = β1 · Awardvdt + β2 · Completevdt + αv + γ dt + X v · ν t + ηvdt   (3)


F orestvdt is a measure of forest cover in village v and district d in year t. Awardvdt is
an indicator that takes the value one for the years where a contract has been awarded
for the construction of a road to village v but the road construction is not yet complete.
  13
       Results at many different bandwidths are shown in Appendix Table A3.



                                                   15
Completevdt is an indicator that takes the value one for all years following the completion
of a new road to village v . We separate these two periods because the road construction
process may have effects on forest cover (such as clearing of forested area to make room for
the physical placement of roads) that are theoretically distinct from the economic effects
of a village having a new road. Village fixed effects (αv ) control for all village-level time-
invariant unobservables, while district-year fixed effects (γ dt ) control for any pattern of
regional shocks.14 We also interact a vector of baseline village controls X v (baseline forest
cover, village population and distance from the village to the nearest towns) with year fixed
effects. These control for any differential time path of forest cover that is correlated with
baseline village characteristics. These controls are particularly important because larger
villages are more likely to be treated earlier due to program implementation rules. Standard
errors are clustered at the village level to account for serial correlation.
    We can interpret β1 and β2 as the effects of road construction activities and the effects of
new roads, respectively; both coefficients describe outcomes relative to the period before any
construction began. We restrict our sample from the universe of villages in India to those
that had no road in 2000 and had a road completed during the study period. We do this so
as not to compare villages that received new roads with those that did not; the endogeneity
problem in such a comparison is severe.15 Identification rests on the assumption that, among
the set of villages that received roads in the sample period, there are no other systematic
changes specific to villages in the years that roads were awarded and completed that are not
caused by the roads themselves.

IV.D     Rural Roads: Difference-in-Differences Results

The difference-in-difference estimates of the impact of rural roads on village-level forest
cover are summarized by Figure 2. These graphs show the residual of log forest cover —
  14
    Results are unchanged by replacing these with state-year or subdistrict-year fixed effects.
  15
    As we show above, a minority of roads were allocated strictly due to the village population thresholds.
There are enough of these to estimate a regression discontinuity test on local compliers, but not enough to
assume that all treated villages are selected as good as randomly.



                                                    16
after taking out fixed effects and controls described above — as a function of the number of
years elapsed since a road was completed in a given village. Panel A shows all previously-
unconnected villages that received new roads between 2001 and 2014. Panel B restricts the
set of villages to those with above median forest cover in 2000. We show only four years before
and after road construction because wider windows have more variable sample composition
across estimates; this occurs because we observe different length of pre- and post-periods for
different villages depending on their date of treatment.16 Two patterns are evident in the
figure. First, there is a statistically significant reduction in forest cover approximately two
years before road construction is complete. Second, forest cover marginally increases in the
four years after road completion recovering some or all of the pre-treatment drop.
   Given that these rural roads took one to two years to build, this pattern is consistent
with a small degree of forest loss (approximately 0.5%) during the road construction period,
with partial or complete recovery afterward. We test this directly in Table 3, which shows
estimates from Equation 3. Our main estimate in Column 1 shows that villages lose 0.5%
of their forest cover during the period between the awarding of a road construction contract
and the completion of a road. However, that forest loss is fully restored in the period after
the road has been completed; the estimate of 0.002 log points on the completion indicator
can be interpreted as the difference in forest cover between the post-road and the pre-award
periods. Relative to the pre-award period, we can rule out gains larger than 0.6% and
declines larger than 0.2% in forest cover. In Column 2, we show that failing to account
for the award period would lead to the estimation of a marginal forest cover gain of 0.5%
because it would incorrectly attribute the construction period loss to the pretrend. This
result highlights the importance of accounting for the construction period when studying
the environmental impacts of new infrastructure. Columns 3 and 4 present estimates where
forest cover is measured as the average share of each pixel that is covered by forest; results
are similar. These estimates are based on different lengths of post-construction periods in
  16
       Appendix Figure A4 shows a wider time window around treatment; the pattern is the same.



                                                   17
different villages, but on average they show effects for four years after treatment.17
     Table 4 shows these estimates along the same dimensions of heterogeneity described
above. Effects are broadly similar whether we cut the sample on baseline forest cover,
population share of Scheduled Tribes, or asset poverty. There is thus no evidence that our
zero results are hiding differential positive and negative effects in different places. The panel
estimates confirm the finding in the regression discontinuity analysis, using a different set of
villages with a different local average treatment effect; the evidence is clear that new rural
roads have had a negligible effect on local forest cover.

V        Impacts of Major Highways on Forest Cover

In this section, we aim to identify the causal impact of highways on local forest cover. The
identification challenge is that highways are typically built to connect cities with current or
anticipated economic growth; if economic growth is correlated with forest cover changes for
any reason other than the direct effect of highways, then we cannot interpret the correlation
between highways and forests as a causal effect.
     We therefore focus on a set of places that happen to be in between the targeted endpoints
of India’s new highways, as in Ghani, Goswami and Kerr (2016). Both the Golden Quadrilat-
eral (GQ) and the North-South and East-West (NS-EW) corridors were upgraded with the
objective of improving connections between India’s major cities and regions; the connection
of secondary cities and intermediate places on the route was a secondary priority. Because
these intermediate regions were targeted incidentally rather than directly, the placement of
the highways is less likely to be driven by existing or anticipated economic growth.
    17
      Appendix Table A5 shows that these estimates are robust to a range of specifications including the use
of village time trends, subdistrict-year fixed effects (instead of district-year fixed effects) and using a limited
sample of roads for which we have at least 4 (or 5) years of both pre-treatment and post-treatment data.
Appendix Table A6 shows additional specifications. Column 1 adds villages that did not receive roads in the
sample period, the specification used in Kaczan (2017). Like Kaczan (2017), we find a positive treatment
coefficient; however, Column 2 shows that this is not robust to the inclusion of village-specific time trends,
indicating that never-treated villages are on different forest cover trends from treated villages. Columns 3
and 4 show that our main estimate is robust to village-specific time trends. Column 5 and 6 define the
treated area as a circle around the village with a radius of 5km and 50km, respectively; as in the main
specification, we find no treatment effects at these radii.



                                                      18
   We can further generate a plausible counterfactual that describes how forest cover would
have changed in the absence of the highway upgrades. Like the GQ, the NS-EW route
was an important transportation corridor in 2000 and was to upgraded before 2005 as part
of NHDP, but the project did not begin in earnest until several years after the GQ was
completed. Our main estimates examine forest changes along the GQ corridor during and
after the construction years, as compared to regions further from the GQ. We then test
for effects along the NS-EW route using a similar specification, showing there are no effects
along the second corridor until after 2008, as would be expected given the construction delay.
   As a starting point, Figure 3 plots kernel-smoothed local regression estimates of mean
forest cover and forest cover change as a function of distance from each highway. Initial
forest cover (Panel A) is broadly similar across the two highways. Panel B shows forest
cover change from 2000-2008, also by distance to each highway. Relative to the NS-EW
(dashed line), forest cover within 100 km of the GQ (solid line) falls substantially between
2000 and 2008. At further distances the effects are similar across the two highways, though
there may be smaller relative gains for the GQ. We present this as suggestive evidence of
relative forest loss along the GQ corridor during and after its construction. The rest of this
section generates formal tests for change, controlling for fixed effects and other factors that
may have simultaneously influenced forest change.

V.A    Highways: Empirical Specification

The simplest form of the difference-in-differences specification is described by the following
equation:


            F orestist = β0 + β1 CLOSEis + β2 P OSTt + β3 CLOSEis ∗ P OSTt +       ist      (4)


In this specification, i indexes a subdistrict in state s and time t, CLOSEis is an indicator for
subdistricts close to the highway, and P OST indicates years following the completion of the
highway. F orestist is a measure of forest cover in subdistrict i and state s at time t, usually


                                              19
log total forest cover. β3 describes the differential change in forest between locations that are
near and far from the highway network after the highway is built, controlling for the same
geographic difference before the highway was built. If new highways cause deforestation, we
expect β3 to be less than zero. We conduct our analysis at the subdistrict level, because
subdistricts are contiguous regions that cover the whole of India for which we can calculate
a range of demographic and socioeconomic controls. We weight results by subdistrict area.18
There are approximately 4000 subdistricts in India.
    We extend this simple specification in three ways. First, because we do not have strong
priors on which distances are near and which are far, we use a flexible set of distance in-
dicators to nonparametrically identify highway effects at a range of distances. Estimates
can still be interpreted as the difference from a given band to the omitted (most remote)
distance band. This ensures that our result is not dependent upon a particular definition
of closeness. Second, because the construction of India’s national highways were multiyear
projects, we separate the P OSTt indicator into multiple periods to capture construction and
post-construction effects. Third, we add a wide set of fixed effects and controls to improve
precision and reduce bias from omitted variables. The most flexible estimating equation is:

                  D    2014
  F orestist =                 βd,t 1(DISTi ∈ (d− , d+ ), Y EAR = t) + γst + X i · ν t + ψd + ηist           (5)
                  d=1 t=2001


The distance to the highway is divided into D bands, the boundaries of which are indexed
by d. We include a distance band fixed effect ψd , state-year fixed effect γst and a vector of
subdistrict controls (X i ) interacted with year fixed effects (ν t ). The latter control for any
differential time path of forest cover that is correlated with baseline subdistrict characteris-
tics. Controls are the same as in Equation 3. We include locations up to a distance D + E
from the Golden Quadrilateral; the outer boundary (D, E ) is the omitted distance category
against which the other estimates can be compared. Unless otherwise specified, we define
   18
      Results from a town- and village-level analysis with subdistrict clusters deliver nearly identical results.
We could in principle conduct analysis at the grid cell level, but this would require imputation for control
variables not available at the grid cell level.


                                                       20
(D, E ) as the 200-300km distance band.19 βd,t identifies the change in forest cover from 2000
to year t, at distance range d from the highway, relative to the omitted distance range (D, E ).
The βd,t coefficients can thus be directly interpreted as the effect of highway construction
on forest cover after t years. If new highways cause proximate forest cover loss, we expect
βd,t to take on negative values for low values of d in the periods t after highway construction
has begun. For graphs, we include a set of indicator variables βd,2000 which describe baseline
forest cover as a function of distance from the highway.20 Standard errors are clustered at
the subdistrict level to account for serial correlation. Because the regression above may have
hundreds of coefficients, we pool years or distances in difference specifications to improve
interpretability.
    We exclude areas within 200km of the nodal towns on the highway routes, as we wish to
identify effects on intermediate regions rather than at the highway end points, as in Goswami
Ghani and Kerr (Ghani, Goswami and Kerr, 2016). Estimates of NS-EW treatment effects
omit areas that are within 200km of the GQ as they are plausibly being treated by the other
highway network. We do not omit NS-EW regions from the GQ regressions because NS-EW
construction has barely begun during the periods of interest for the GQ analysis; however,
regression results are not changed by omitting places within 200km of NS-EW.

V.B     Highways: Estimates on Forest Cover

Panel A of Figure 4 plots coefficient estimates from a single estimation of Equation 5, with
distances from the GQ highway divided into 10km bands, and years divided into a single
   19
      Alternate choices of the range of the omitted group, including using the remainder of the country does
not appreciably affect our estimates.
   20
      We do not include subdistrict fixed effects because we want to generate coefficients on the distance
band indicators for the omitted year 2000 — these coefficients describe the baseline differences in forest
cover between places that were near and far from the highway. However, inclusion of subdistrict fixed effects
does not meaningfully change the results. We use state-year fixed effects rather than district-year fixed effects
because we wish to test for meaningful effects of distance from highways that may extend beyond the radius
of districts. District-year fixed effects would absorb true effects of the GQ that span distances larger than
districts. The analysis of Goswami Ghani and Kerr (Ghani, Goswami and Kerr, 2016) is entirely district
level, giving us reason to expect meaningful cross-district effects. As expected, the inclusion of district-year
fixed effects attenuates our results slightly but does not change the direction of effects nor eliminate statistical
significance.



                                                       21
pre-construction year (2000), the construction period (2001-2004), and two post construction
periods (2005-2008 and 2009-2012). All estimates describe the difference between a given
10km distance band from the GQ and the omitted category of 250-300km.21 The solid black
line describes baseline forest cover as a function of distance from the GQ corridor. The
remaining lines show that forest cover within 100 km of the GQ declines rapidly during
the GQ construction period and then continues to fall in the years following construction.
Effects are slightly smaller in the 100-150km bandwidth, and statistically indistinguishable
from zero at a distances greater than 150km from the highway.
    To alleviate the concern that these forest losses are explained by existing trends in forest
loss along existing highway corridors, we run the same estimation for subdistricts along the
NS-EW corridor and show results in Panel B. As predicted, there are no differential changes
in forest cover close to the NS-EW route before 2008. Net forest loss along NS-EW begins in
the 2009-2012 period, and the distance effects then look similar to those of the GQ.22 This
distance pattern of forest cover loss is similar to that found in the Amazon (Pfaff et al., 2007).
Effects along the NS-EW corridor may be slightly smaller than along the GQ both because
construction took place slowly and was still incomplete at the end of the sample period
in 2014; the network structure of highways mean that the value of any particular segment
depends on the completion status of other segments. Concentration of industry along the
GQ corridor may have also reduced the importance of NS-EW as an intercity transportation
corridor by the time the NS-EW upgrades took place.
    Table 5 presents regression estimates from Equation 5, with distances in 50km bands for
legibility. Each estimate describes the difference in forest cover between a given distance
band and the omitted category of 200-300km. Columns (1) and (2) describe the impact
  21
      We include coefficients for the 200-250 km bands in order to plot treatment effects at these ranges.
Effects in closer bands are very similar if we restrict the distance indicators to 200km and use 200-300km as
the omitted group, because there are few differences across years in the 200-250km range.
   22
      Standard errors are omitted in the figure for visual clarity. For the GQ, differences between the 2000
estimates and the 2001-2004 estimates are statistically significant at the 1% level for all estimates up to
150km. For the NS-EW, differences between the 2000 estimates and the 2009-2012 treatment estimates are
statistically distinguishable at the 1% level until the 180km estimate.



                                                    22
of the GQ on forest cover. The top four rows of the table show estimates of construction
period impacts on forest cover at various distance bands. Places within 50km of the new
highway network lose 27 log points of forest cover (Column 1) or 1.3 percentage points of
forest cover (Column 2, on a base of 7.5%), and the effects shrink at greater distances. The
next four rows show similar effects (still relative to year 2000) in the post-construction period
of 2005-2008. The final four rows show estimates of baseline differences between the GQ and
the regions further away; the level differences are small relative to the treatment effects.
   Columns 3 and 4 of Table 5 show comparable estimates along the NS-EW corridor for
the same time periods. There are no detectable changes in forest cover close to the NS-
EW in the time period when significant forest loss took place near the GQ. Along with
Figure 4, this should alleviate any concern that the GQ treatment effects are driven by
generalized deforestation along existing highway corridors from 2001-2008. These results
are robust to instrumenting for highway location using straight line instruments connecting
the nodal cities of the highway network, as employed by Ghani, Goswami and Kerr (2016);
we present analogous reduced form estimates to the above in Appendix Table A7. These
estimates alleviate the concern that the particular routing of the GQ (but not the NS-EW)
was specifically targeted to places that may have already been losing forest cover.
   Our estimates describe the difference between forest change in the proximity of the up-
graded highways to forest changes far from the highways. One concern with these estimates
is that they could be describing displacement of forest loss from the hinterlands to the high-
way corridors, or even net afforestation in the hinterlands. This concern arises frequently
in studies of transportation projects with national scale, and is typically only resolved by
assumption through a structural modeling approach, which is beyond the scope of this pa-
per. This said, large displacement effects are made less plausible by the low quality of the
broader road network, and by the high transportation costs during the sample period, weak-
ening market connections with the hinterlands of these highways. While we cannot entirely
rule out that there may be some displacement effects, Panel B of Figure 3 suggests that


                                              23
effects are driven by the highway corridor regions rather than the hinterlands; forest cover
diverges between the places close to the GQ and NS-EW, but not in those far away.

V.B.1       Mechanisms for Highway Effects

We consider four possible mechanisms for the forest cover loss caused by India’s major
highway networks: (i) increased demand for timber products by firms due to local growth;
(ii) increased demand for firewood due to shifts in household fuel consumption; (iii) expansion
of agriculture into previously forested lands; and (iv) clearing of trees for settlements and
industry. This section presents suggestive evidence that the deforestation along India’s major
highways is predominantly caused by increased logging driven by local timber demand.
   To identify potential mechanisms, we use the regression specification used to identify
effects of highways on forest cover (Equation 5), with data from the economic and population
censuses which were undertaken in various periods between 1990 and 2013. Because the
results above suggest forest cover changes within 100km of the highway network, we use
100km distance bands and define 200-300km as the omitted distance band.23 The years in
the sample are determined by census availability.
   We first look at changes in employment in the list of industries that are directly down-
stream from timber harvesting (described in Section III). Panel A of Figure 5 shows results
from a regression of log employment in major wood-consuming sectors on the usual set
of year-distance-band fixed effects. We graph the point estimates on the 0-100km coeffi-
cient in each year that the Economic Census is available (i.e. the coefficients β0−100km,1990 ,
β0−100km,1998 , β0−100km,2005 , and β0−100km,2013 from Equation 5). These estimates can be in-
terpreted as the difference in residual log employment between the 0-100km distance band
and the 200-300km distance band, after controlling for state-year fixed effects and the con-
trols described above. In 1990 and 1998 (before the GQ was begun), there is no significant
difference between areas close to the highway corridor and areas that are far from it, nor is
there a significant trend. By 2005, we see a 5% increase in employment in wood-consuming
  23
       We find virtually identical results when we use 50km distance bins.


                                                     24
firms in the GQ corridor relative to the hinterland, which continues to rise through 2013.
Panel B shows similar results for employment in logging firms; we omit 2005 because logging
firms were not distinguished from firms engaged in afforestation in the 2005 Economic Cen-
sus. These two graphs suggest that demand for wood from downstream firms is a plausible
explanation for local deforestation after construction of the GQ.24 Note that logging firms
were more common along the transport corridor even before the GQ was built, but there is
no suggestion of a pretrend that could explain what we see after highway construction. Note
that employment in other sectors of the economy, which also consume wood, exhibit similar
treatment effects close to the new highways Ghani, Goswami and Kerr (2016).
   Panels C through E of Figure 5 show the effects of the GQ upgrades on household fuel
consumption, for which data are available in 2001 and 2011. We observe marginal increases in
firewood and imported wood use, and comparable reductions in use of local non-wood fuels.
These effects are not statistically significantly different from zero, nor are they large enough
to explain a 20% reduction in forest cover in the neighborhood of the GQ corridor. Panel F
shows that land use shifts slightly away from agricultural uses following the construction of
the GQ, breaking a previous upward trend, making agricultural extensification an unlikely
explanation for the treatment effects.
   It is difficult to directly test the last hypothesis that net forest loss has come from the
expansion of land dedicated to settlement and industry, because data on land dedicated to
settlement and industry only becomes available in the 2011 Population Census. However,
it is implausible that settlement and industrial expansion could explain a 20% reduction in
forest cover in a distance band as wide as 100km around a 6000km long highway corridor.
In 2011, only 6.7% of of rural land was used for settlement and industry.
   Note that the large relative forest losses along the GQ corridor are unlikely to be direct
effects of the road construction process. Although we found evidence of these effects in the
  24
    Appendix Tables A8 and A9 show complete regression results for all mechanism tests. As would be
expected, employment in logging firms is more geographically diffuse, as those firms reach further into the
GQ’s hinterland.



                                                  25
construction of rural roads in Section IV, these were very local, temporary, and an order of
magnitude lower in size. It is also implausible that direct construction effects would extend
more than a few kilometers from the roadway.
     In conclusion, we find evidence that expansion of industry demand for timber can ex-
plain forest loss in the GQ corridor, and we can rule out agricultural expansion, changes in
household fuel consumption and settlement expansion as mechanisms.

VI     Conclusion

The development, maintenance and expansion of transportation infrastructure is an impor-
tant driver and correlate of economic development around the world. In this paper, we
provide causal estimates of the ecological impact of two transportation investments with
global significance: India’s massive expansion of rural roads and its upgrading of national
highways. Using identification strategies established in the literature, we find that, (i) the
new rural roads had negligible effects on forest cover and; (ii) the highway expansions had a
large negative effect on forest cover, which may have been driven by the expansion of wood-
using industries. Methodologically, we demonstrate the critical importance of accounting
for endogeneity and separately estimating the effects of the construction period from the
post-completion period.
     Globally, road expansion is expected to dramatically increase through the course of the
21st century. Some additional 25 million kilometers (16 million miles) of road infrastructure
is projected to be built by 2050, a 60% increase over 2010 levels. Nine out of ten of these roads
will be built in developing countries (Laurance et al., 2014). At the same time, tropical forests
in developing countries are increasingly under threat. These forests not only provide global
carbon benefits but also provide important local ecosystems which support biodiversity as
well as the generally poor populations that rely on them (Barrett, Garg and McBride, 2016).
Against the background of this tension between economic development and environmental
conservation, understanding the relationship between roads and forests is fundamental to a
successful strategy for sustainable development.

                                               26
   Crucially, we show that the impact of road construction depends on what those roads
connect. The fiscal costs of the two large scale transportation investments that we study
were similar, but they had vastly different environmental consequences. Expansion of existing
highway corridors caused changes in the spatial distribution of industry, which had dramatic
effects on forest use in India. In contrast, building roads to connect smallholder farmers to
new markets had virtually no impact on local forests, even for those farmers most likely to
draw some part of their livelihoods from those forests.
   Our results highlight the importance of context in the study of infrastructure and envi-
ronment. Even within one country, we find substantially different impacts on forest cover of
different kinds of road investments. The focus of earlier studies on the Brazilian Amazon is
undoubtedly justified given the global significance of that ecosystem. However, for under-
standing the potential environmental impacts of transportation investments elsewhere in the
world, additional research is surely required.




                                                 27
References
Adukia, Anjali, Sam Asher, and Paul Novosad. 2017. “Educational Investment Re-
   sponses to Economic Opportunity: Evidence from Indian Road Construction.” Mimeo.

Alix-Garcia, Jennifer, Craig McIntosh, Katharine R.E. Sims, and Jarrod R.
    Welch. 2013. “The Ecological Footprint of Poverty Alleviation: Evidence from Mexico’s
    Oportunidades Program.” Review of Economics and Statistics, 95(2): 417–435.

Andreoni, James, and Arik Levinson. 2001. “The Simple Analytics of the Environmen-
   tal Kuznets Curve.” Journal of Public Economics, 80(2): 269–286.

Angelsen, Arild, and David Kaimowitz. 1999. “Rethinking the Causes of Deforestation:
   Lessons from Economic Models.” The World Bank Research Observer, 14(1): 73–98.

Antweiler, Werner, Brian R. Copeland, and Scott M. Taylor. 2001. “Is Free Trade
   Good for the Environment?” American Economic Review, 91(4).

Arrow, Kenneth, Bert Bolin, Robert Costanza, Partha Dasgupta, Carl Folke,
    Crawford S. Holling, Bengt-Owe Jansson, Simon Levin, Karl-G¨      oran M¨ aler,
    Charles Perrings, et al. 1995. “Economic Growth, Carrying Capacity, and the En-
    vironment.” Ecological Economics, 15(2): 91–95.

Asher, Sam, and Paul Novosad. 2017. “Rural Roads and Structural Transformation.”
   Mimeo.

Assun¸ ao, Juliano, Molly Lipscomb, Ahmed Mushfiq Mobarak, and Dimitri Sz-
      c˜
    erman. 2017. “Agricultural Productivity and Deforestation in Brazil.” Mimeo.

Barber, Christopher P., Mark A. Cochrane, Carlos M. Souza, and William F.
   Laurance. 2014. “Roads, Deforestation, and the Mitigating Effect of Protected Areas
   in the Amazon.” Biological Conservation, 177: 203–209.

Barrett, Christopher B, Teevrat Garg, and Linden McBride. 2016. “Well-being
    dynamics and poverty traps.” Annual Review of Resource Economics, 8: 303–327.

Bauch, Simone C., Anna M. Birkenbach, Subhrendu K. Pattanayak, and Erin O.
   Sills. 2015. “Public Health Impacts of Ecosystem Change in the Brazilian Amazon.”
   Proceedings of the National Academy of Sciences, 112(24): 7414–7419.

Burgess, Robin, Matthew Hansen, Benjamin A. Olken, Peter Potapov, and Ste-
   fanie Sieber. 2012. “The Political Economy of Deforestation in the Tropics.” The
   Quarterly Journal of Economics, 127(4): 1707–1754.

Chomitz, Kenneth M., and David A. Gray. 1996. “Roads, Land Use, and Deforestation:
   a Spatial Model Applied to Belize.” The World Bank Economic Review, 10(3): 487–512.

Copeland, Brian R., and M. Scott Taylor. 1994. “North-South Trade and the Envi-
   ronment.” The Quarterly Journal of Economics, 109(3).


                                           28
Copeland, Brian R., and Scott M. Taylor. 2004. “Trade, Growth, and the Environ-
   ment.” Journal of Economic Literature, 42(1).

Cropper, Maureen, Jyotsna Puri, Charles Griffiths, Edward B. Barbier, and
   Joanne C. Burgess. 2001. “Predicting the Location of Deforestation: The Role of
   Roads and Protected Areas in North Thailand.” Land Economics, 77(2): 172–186.

Dasgupta, Partha. 2007. “The Idea of Sustainable Development.” Sustainability Science,
   2(1): 5–11.

Dasgupta, Susmita, and David Wheeler. 2016. “Minimizing Ecological Damage from
   Road Improvement in Tropical Forests.” World Bank Policy Research Working Paper
   No. 7826.

Dasgupta, Susmita, Benoit Laplante, Hua Wang, and David Wheeler. 2002.
   “Confronting the Environmental Kuznets Curve.” Journal of Economic Perspectives,
   16(1): 147–168.

Datta, Saugato. 2012. “The Impact of Improved Highways on Indian Firms.” Journal of
    Development Economics, 99(1): 46–57.

Den Butter, F.A.G., and Harmen Verbruggen. 1994. “Measuring the Trade-Off Be-
   tween Economic Growth and a Clean Environment.” Environmental and Resource
   Economics, 4(2): 187–208.

Deng, Xiangzheng, Jikun Huang, Emi Uchida, Scott Rozelle, and John Gibson.
   2011. “Pressure Cookers or Pressure Valves: Do Roads Lead to Deforestation in China?”
   Journal of Environmental Economics and Management, 61(1): 79–94.

Ferretti-Gallon, Kalifi, and Jonah Busch. 2014. “What Drives Deforestation and What
    Stops it? A Meta-Analysis of Spatially Explicit Econometric Studies.” Mimeo.

Foster, Andrew D., and Mark R. Rosenzweig. 2003. “Economic Growth and the Rise
    of Forests.” The Quarterly Journal of Economics, 118(2): 601–637.

Frankel, Jeffrey A., and Andrew K. Rose. 2005. “Is Trade Good or Bad for the Envi-
    ronment? Sorting Out the Causality.” Review of Economics and Statistics, 87(1).

Garg, Teevrat. 2017. “Ecosystems and Human Health: The Local Benefits of Forest Cover
   in Indonesia.” Mimeo.

Geist, Helmut J., and Eric F. Lambin. 2002. “Proximate Causes and Underlying Driving
    Forces of Tropical Deforestation: Tropical Forests are Disappearing as the Result of
    Many Pressures, Both Local and Regional, Acting in Various Combinations in Different
    Geographical Locations.” BioScience, 52(2): 143–150.

Gelman, Andrew, and Guido Imbens. 2014. “Why High-Order Polynomials Should Not
   Be Used in Regression Discontinuity Designs.” NBER Working Paper No. 20405.


                                          29
Ghani, Ejaz, Arti Grover Goswami, and William R. Kerr. 2016. “Highway to Success:
   The Impact of the Golden Quadrilateral Project for the Location and Performance of
   Indian Manufacturing.” Economic Journal, 126(591): 317–357.

Grossman, Gene M., and Alan B. Krueger. 1995. “Economic Growth and the Envi-
   ronment.” The Quarterly Journal of Economics, 110(2): 353–377.

Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A.
   Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kom-
   mareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Townshend.
   2013. “High-Resolution Global Maps of 21st-Century Forest Cover Change.” Science,
   342(6160): 850–853.

Imbens, Guido, and Karthik Kalyanaraman. 2012. “Optimal Bandwidth Choice for
   the Regression Discontinuity Estimator.” Review of Economic Studies, 79(3): 933–959.

Imbens, Guido W., and Thomas Lemieux. 2008. “Regression Discontinuity Designs:
   A Guide to Practice.” Journal of Econometrics, 142(2): 615–635.

IPCC. 2007. The Fourth Assessment Report of the Intergovernmental Panel on Climate
   Change. Geneva, Switzerland.

Jayachandran, Seema, Joost de Laat, Eric F. Lambin, Charlotte Y. Stanton,
    Robin Audy, and Nancy E. Thomas. 2017. “Cash for Carbon: A Randomized Trial
    of Payments for Ecosystem Services to Reduce Deforestation.” Science, 357(6348): 267–
    273.

Kaczan, David. 2017. “Can roads contribute to forest transitions?” Mimeo.

Khanna, Gaurav. 2016. “Road Oft Taken: The Route to Spatial Development.” Mimeo.

Koop, Gary, and Lise Tole. 1999. “Is There an Environmental Kuznets Curve for Defor-
   estation?” Journal of Development Economics, 58(1): 231–244.

Laurance, William F., Gopalasamy Reuben Clements, Sean Sloan, Christine S.
   O’connell, Nathan D. Mueller, Miriam Goosem, Oscar Venter, David P.
   Edwards, Ben Phalan, Andrew Balmford, et al. 2014. “A Global Strategy for
   Road Building.” Nature, 513(7517): 229.

Li, Man, Alessandro De Pinto, John M. Ulimwengu, Liangzhi You, and
    Richard D. Robertson. 2015. “Impacts of Road Expansion on Deforestation and
    Biological Carbon Loss in the Democratic Republic of Congo.” Environmental and
    Resource Economics, 60(3): 433–469.

McCrary, Justin. 2008. “Manipulation of the running variable in the regression disconti-
   nuity design: A density test.” Journal of Econometrics, 142(2): 698–714.

National Rural Roads Development Agency. 2005. “Pradhan Mantri Gram Sadak
    Yojana - Operations Manual.” Ministry of Rural Development, Government of India.

                                           30
Pfaff, Alexander, Juan Robalino, Robert Walker, Steven Aldrich, Marcellus Cal-
    das, Eustaquio Reis, Stephen Perz, Claudio Bohrer, Eugenio Arima, William
    Laurance, et al. 2007. “Road Investments, Spatial Spillovers, and Deforestation in the
    Brazilian Amazon.” Journal of Regional Science, 47(1): 109–123.

Pfaff, Alexander S.P. 1999. “What Drives Deforestation in the Brazilian Amazon?: Ev-
    idence From Satellite and Socioeconomic Data.” Journal of Environmental Economics
    and Management, 37(1): 26–43.

Stern, David I. 2004. “The Rise and Fall of the Environmental Kuznets Curve.” World
    Development, 32(8): 1419–1439.

Stern, David I., Michael S. Common, and Edward B. Barbier. 1996. “Economic
    Growth and Environmental Degradation: the Environmental Kuznets Curve and Sus-
    tainable Development.” World Development, 24(7): 1151–1160.

Townshend, J., M. Hansen, M. Carroll, C. DiMiceli, R. Sohlberg, and C. Huang.
   2011. “User Guide for the MODIS Vegetation Continuous Fields product Collection 5
   version 1.”

U.S. Environmental Protection Agency. 2017. Inventory of U.S. Greenhouse Gas
    Emissions and Sinks: 1990-2015. Washington, D.C.

Weinhold, Diana, and Eustaquio Reis. 2008. “Transportation Costs and the Spatial
   Distribution of Land Use in the Brazilian Amazon.” Global Environmental Change,
   18(1): 54–68.




                                           31
                                                                            Figure 1
                                                         Regression Discontinuity Estimates of Impact of
                                                                  Rural Roads on Forest Cover
                                                 Panel A: First Stage (2013)                                                           Panel B: First Stage (by year)



                                     .5
                                                                                                                                  .2




                                     .4




                                                                                                     RD First Stage Coefficient
                             New Road


                                                                                                                                  .1
                                .3   .2




                                                                                                                                  0
                                     .1




                                                  −200     −100            0             100   200                                     2000         2005             2010    2015
                                                              Population Minus Threshold                                                                     Year
32




                                                Panel C: Reduced Form (2013)                                                  Panel D: Reduced Form (by year)
                                     .6




                                                                                                                                        .1




                                                                                                     RD Reduced Form Coefficient
                                                                                                      (y = Log Forest Cover 2013)
                                     .5




                                                                                                                                       .05
                             Log Forest Cover
                                    .4




                                                                                                                                         0
                                     .3




                                                                                                                                    −.05
                                     .2




                                                                                                                                       −.1
                                                  −200     −100            0             100   200                                           2000     2005            2010   2015
                                                              Population Minus Threshold                                                                      Year




     The figure shows regression discontinuity estimates of the impact of new rural roads on local deforestation. Panel A shows the first stage probability of a
     village receiving a new road before 2013 as a function of its population relative to the population threshold. Each point shows the mean of the Y variable
     in a given population bin. Panel B shows the first stage RD estimate of a village receiving a new road by the year indicated on the X axis. Each point
     is an estimate from an RD first stage regression. Panel C is analogous to Panel B; the dependent variable is the log of forest cover in 2013. The points
     show the mean of this variable in each population bin; population is shown relative to the population treatment threshold. Panel D shows reduced form
     RD estimates of the impact of being above the population threshold on forest cover in each year on the X axis. All estimates in Panels B and D use the
     same specification as Table 2, and include district-population threshold fixed effects and a control for baseline forest cover.
                             Figure 2
           Difference-in-Differences Estimates of Impact of
                   Rural Roads on Forest Cover
                                                          Panel A: Full Sample

                             .015      .01
               Residual Log Forest Cover
                 0         .005
                             −.005




                                             <= −4   −3    −2       −1       0      1       2   3   >= 4
                                                                Years after Road Completion




                                                     Panel B: High Baseline Forest
                             .02
               Residual Log Forest Cover
                   0         −.01 .01




                                             <= −4   −3    −2       −1       0      1       2   3   >= 4
                                                                Years after Road Completion




    The figure shows year-by-year estimates of log forest cover in villages that received
new roads between 2001 and 2013. Villages are grouped on the X axis according to the
year relative to road completion. Each point thus shows the average value of log forest
cover in villages in a given year relative to the treatment year, controlling for village fixed
effects, district*year fixed effects, baseline population * year and baseline log forest cover
* year interactions. Standard errors are clustered at the village level. The year before
road completion is omitted (t=-1); forest cover is thus shown relative to this period.



                                                                        33
                                                                Figure 3
                                                  Forest Cover and Forest Cover Change
                                                  Along Highway Corridors (2000-2008)
                                                      Panel A: Pre-Expansion Forest Cover (2000)


                                             11
            Log Total Forest (2000)




                                             10




                                              9
                                                  0            50              100             150   200
                                                                    Distance (km) to Highway

                                                                         Golden Quadrilateral
                                                                         North−South/East−West


                                                      Panel B: Forest Cover Change (2000-2008)
                                             1
            Change in Log Forest 2000−2008




                                             .5




                                             0


                                                  0           50               100             150   200
                                                                    Distance (km) to Highway

                                                                         Golden Quadrilateral
                                                                         North−South/East−West




     Panel A shows a kernel-smoothed regression of log subdistrict forest cover in 2000
on distance to the corridors where the Golden Quadrilateral and North-South/East-West
highways will be expanded. Panel B plots kernel-smoothed regression estimates of change
in log subdistrict forest cover from 2000 to 2008 against distance to each highway network.
By 2008, there was very little construction on the NS-EW corridor, so we treat it here
as a control group. The plots display means that are unadjusted for any fixed effects or
controls. 95% confidence intervals are displayed in the shaded areas.



                                                                           34
                             Figure 4
           Difference-in-Differences Estimates of Impact of
           Highways on Forest Cover, by Distance Bands
                                     Panel A: Golden Quadrilateral
                           .5


                          .25
            Log Forest




                           0


                         −.25


                          −.5
                                0    50         100          150       200   250
                                              Distance from GQ (km)

                                                    2000
                                                    2001−2004
                                                    2005−2008
                                                    2009−2012


                                    Panel B: North-South/East-West
                           .5


                          .25
            Log Forest




                           0


                         −.25


                          −.5
                                0    50          100         150       200   250
                                            Distance from NS−EW (km)

                                                    2000
                                                    2001−2004
                                                    2005−2008
                                                    2009−2012




    The figure shows point estimates from Equation 5, with distance from the Golden
Quadrilateral highway network (Panel A) and distance from the North-South/East-West
highway network (Panel B) divided into 10km bands. Each point on the graph shows, for
a given set of years (shown in the legend), the average value of log forest cover at a given
distance band from the given highway network, relative to the omitted distance band of
290 to 300 km from the highway. All estimates control for state*year fixed effects, baseline
population * year and baseline log forest cover * year interactions.



                                                  35
                                                               Figure 5
                                         Mechanism Tests for Impact of Highways on Deforestation
     Panel A: Employment in Wood-Using Firms                                                                       Panel B: Employment in Logging Firms




                                                                                                                                                                .07
                                                  .1
        Coefficient on Indicator 1(0−100km from GQ)




                                                                                                                    Coefficient on Indicator 1(0−100km from GQ)
                                                                                                                                                      .06
                                    .05




                                                                                                                                    .04       .05
                     0




                                                                                                                         .03
                                 −.05




                                                                                                                                             .02
                                                          1990   1995   2000          2005   2010   2015                                                               1990   1995   2000          2005   2010   2015
                                                                               Year                                                                                                         Year



       Panel C: Share of Energy from Firewood                                                                   Panel D: Share of Energy from Imported Fuels
                                                    .04




                                                                                                                                                                .02
        Coefficient on Indicator 1(0−100km from GQ)




                                                                                                                    Coefficient on Indicator 1(0−100km from GQ)
                                                                                                                                                     .015
                                         .02




                                                                                                                                          .01
                −.02           0




                                                                                                                            .005
                                 −.04




                                                                                                                                             0




                                                          1990   1995   2000          2005   2010   2015                                                               1990   1995   2000          2005   2010   2015
                                                                               Year                                                                                                         Year



Panel E: Share of Energy from Local Non-Wood Sources                                                             Panel F: Agricultural Share of Village Land
                                                                                                                                                                −.04
                                                    .04
        Coefficient on Indicator 1(0−100km from GQ)




                                                                                                                    Coefficient on Indicator 1(0−100km from GQ)
                                                                                                                                                     −.06
                                         .02




                                                                                                                                         −.08
                −.02           0




                                                                                                                             −.1
                                 −.04




                                                                                                                                             −.12




                                                          1990   1995   2000          2005   2010   2015                                                               1990   1995   2000          2005   2010   2015
                                                                               Year                                                                                                         Year




       The figure shows point estimates from Equation 5, with distances from the Golden Quadri-
       lateral highway network specified in 100km bands. Each figure shows the point estimate on
       the 0-100km distance indicator, interacted with the year shown in the X axis. The omitted
       category is the set of places that are 200-300 kilometers from the Golden Quadrilateral
       network. The dependent variables in Panels A and B are log employment in respectively
       wood-consuming firms and in logging firms. Logging was not specified in the 2005 Eco-
       nomic Census so this point is omitted. In Panels C through E the dependent variable is
       the share of households’ cooking fuel that takes the form of (C) firewood; (D) imported
       fuels, primarily propane; and (E) crop residue and animal waste. In Panel F, the depen-
       dent variable is the share of village land dedicated to agriculture. All estimates are from
       regressions with state-year fixed effects and standard errors are clustered at the subdis-
       trict level. Appendix Table A9 shows the full set of estimates from the regressions that
       produced these graphs.


                                                                                                           36
                                                                                                              37




                                            Table 1
                                        Summary Statistics

Village-level Statistics
                                                        Mean      Standard Deviation          Observations
New road before 2011                                    0.18             0.38                   257256
Road completion year                                    2007               2                     45459
Population share with no assets (2002)                  0.69             0.31                   169387
Population share Scheduled Tribes (2001)                0.22             0.39                   257256
Agricultural share of village land (2001)               0.64             0.28                   372246
Share energy from firewood (2001)                        0.67             0.26                   409298
Share energy from imports (2001)                        0.07             0.09                   409298
Share energy from local nonwood (2001)                  0.26             0.26                   409298


Subdistrict-level Statistics
                                                       Mean       Standard Deviation          Observations
Average forest cover (2000)                             12.76            14.66                   4019
Average forest cover (2014)                             14.69            14.49                   4019
Distance to Golden Quadrilateral                       218.50           212.33                   4019
Distance to North-South East-West                      191.48           155.68                   4019
Employment in wood-using firms                          141.41           299.77                   4019
Employment in logging firms                              9.78             92.66                   4019
The table shows summary statistics for the samples used for village- and subdistrict-level analyses. Road
completion year is shown only for villages that received new roads between 2001 and 2011. The sample
for the first four village-level variables consists of the set of villages that did not have a road at baseline.
The sample for agricultural land and energy shares consists of all villages with non-zero forest cover at
baseline.
                                                       Table 2
                                    Regression Discontinuity Estimates of Impact of
                                             Rural Roads on Forest Cover

                                First Stage                              Reduced Form                                              IV
                                Any Road       Log Forest   Avg Forest    High Baseline     High ST    Low Assets     Log Forest     Avg Forest
Above Population Threshold       0.163***        0.003        0.042           -0.006         0.007        0.017
                                  (0.010)       (0.011)      (0.060)         (0.013)        (0.014)      (0.017)
New Road                                                                                                                 0.016         -0.062
                                                                                                                        (0.065)       (0.539)
N                                   89476          89476        89476          44880         44388         35520         89476         89476
r2                                   0.25           0.80         0.56           0.69          0.83          0.78          0.80          0.36
∗
  p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01
The table shows regression discontinuity treatment estimates of the effect of new village roads on local forest cover, estimated with Equation 1.
In Column 1, the dependent variable is an indicator that takes the value one if a village received a new road in the sample period. Above
Population Threshold is an indicator for a village population being above the treatment threshold. Columns 2 through 6 show reduced form
estimates of the effect of being above the treatment population threshold. The dependent variables in Columns 2 and 3 respectively are log
village forest cover and average covered share of each village pixel; the data source is Vegetation Continuous Fields. Columns 4 through 6
run the log forest cover specification on subgroups defined respectively by (i) above-median forest cover villages; (ii) above median share of
Scheduled Tribes in a village; and (iii) below median baseline village assets. Columns 7 and 8 show IV estimates of the treatment effects
of new roads, using respectively log and average forest cover as dependent variables. The sample for Columns 2 through 8 includes forest
cover estimates for years 2011 through 2013 for increased precision; robust standard errors are clustered at the village level to account for
serial correlation. All estimates include district-population threshold fixed effects and a control for baseline forest cover.




                                                                                                                                                   38
                                                                             39




                      Table 3
   Difference-in-Differences Estimates of Impact of
           Rural Roads on Forest Cover

                            Log Forest               Average Forest
                          (1)        (2)               (3)      (4)
Award Period         -0.005***                     -0.035***
                       (0.002)                      (0.013)
Completion Period       0.002     0.005***           0.009     0.014
                       (0.002)     (0.002)          (0.015)   (0.012)
District-Year F.E.       Yes         Yes              Yes       Yes
Village F.E.             Yes         Yes              Yes       Yes
N                      689745      689745           689745    689745
r2                       0.94        0.94             0.92      0.92
∗          ∗∗        ∗∗∗
  p < 0.10, p < 0.05, p < 0.01
The table shows difference-in-differences estimates of the impact of new
village roads on local forest cover. We define forest cover as log village
forest cover (Columns 1 and 2) and average covered share of each village
pixel (Columns 3 and 4); the data source is Vegetation Continuous Fields.
The sample consists strictly of villages that received new roads between
2001 and 2013, and were not accessible by paved road in 2001. Award
Period is an indicator variable that takes the value one for years after a
road contract was awarded and before the road was completed. Completion
period is an indicator variable that marks the years after a village’s new
road was built. All regressions include district*year fixed effects, village
fixed effects, baseline population * year fixed effects, and baseline forest
* year fixed effects. Standard errors are clustered at the village level to
correct for serial correlation.
                                                                                                        40




                                 Table 4
                     Rural Roads and Deforestation:
            Heterogeneity of Difference-in-Differences Estimates

                        Baseline Forest                 ST Share                 Asset Poverty
                       High       Low                High     Low              Poor     Not Poor
Award Period         -0.005* -0.005***              -0.003 -0.006***          -0.004 -0.006***
                     (0.003)     (0.002)           (0.003)   (0.002)         (0.003)     (0.002)
Completion Period -0.002          0.002             0.000     0.003            0.001      0.001
                     (0.004)     (0.002)           (0.003)   (0.003)         (0.004)     (0.003)
N                    342345      346860            344760    344580          264810      424545
r2                     0.86        0.92              0.93     0.95             0.94        0.95
∗          ∗∗        ∗∗∗
  p < 0.10, p < 0.05, p < 0.01
The table shows difference-in-differences estimates of the impact of new village roads on local forest
cover, along three dimensions of heterogeneity. Forest cover is defined as log village forest cover;
the data source is Vegetation Continuous Fields. Columns 1 and 2 respectively show estimates for
villages with above and below median baseline forest cover. Columns 3 and 4 respectively show
estimates for villages and above and below median population share of members of Scheduled
Tribes. Columns 5 and 6 respectively show estimates for below- and above-median shares of
households who report no assets in the 2002 Below Poverty Line survey. The sample consists
strictly of villages that received new roads between 2001 and 2013, and were not accessible by
paved road in 2001. Award Period is an indicator variable that takes the value one for years
after a road contract was awarded and before the road was completed. Completion period is
an indicator variable that marks the years after a village’s new road was built. All regressions
include district*year fixed effects, village fixed effects, baseline population * year fixed effects, and
baseline forest * year fixed effects. Standard errors are clustered at the village level to correct for
serial correlation.
                                      Table 5                                                            41
                   Difference-in-Differences Estimates of Impact of
                             Highways on Forest Cover

                                          GQ (Treatment)                                NSEW (Placebo)
                                     Log Forest Avg Forest                        Log Forest Average Forest
GQ Construction Period * (0-50km)    -0.265***    -1.306***                         -0.038         0.183
                                       (0.055)      (0.248)                        (0.072)       (0.305)
GQ Construction Period * (50-100km)  -0.278***    -1.215***                         -0.001        0.077
                                       (0.056)      (0.247)                        (0.065)       (0.286)
GQ Construction Period * (100-150km) -0.221***    -1.085***                         -0.001        -0.176
                                       (0.051)      (0.235)                        (0.060)       (0.267)
GQ Construction Period * (150-200km)  -0.102**     -0.444**                          0.023        0.004
                                       (0.044)      (0.198)                        (0.057)       (0.264)
GQ Post Period * (0-50km)            -0.210***    -1.161***                         -0.013         0.413
                                       (0.061)      (0.253)                        (0.062)       (0.319)
GQ Post Period * (50-100km)          -0.185***    -1.023***                          0.022         0.106
                                       (0.060)      (0.249)                        (0.061)       (0.308)
GQ Post Period * (100-150km)          -0.131**    -0.855***                          0.019        -0.200
                                       (0.058)      (0.226)                        (0.060)       (0.294)
GQ Post Period * (150-200km)            -0.008       -0.198                          0.028        -0.012
                                       (0.051)      (0.197)                        (0.060)       (0.301)
Distance 0-50km                         0.022        0.145                          -0.014         0.265
                                       (0.021)      (0.100)                        (0.020)       (0.722)
Distance 50-100km                       0.021        0.151                          -0.021         0.394
                                       (0.020)      (0.099)                        (0.018)       (0.717)
Distance 100-150km                      0.026        0.133                          -0.006        -0.582
                                       (0.019)      (0.083)                        (0.021)       (0.784)
Distance 150-200km                      0.015        0.075                          -0.000        -0.560
                                       (0.014)      (0.060)                        (0.016)       (0.697)
N                                       26766        26766                          19062         19062
r2                                       0.89         0.91                            0.92          0.85
∗          ∗∗        ∗∗∗
  p < 0.10, p < 0.05, p < 0.01
The table shows treatment estimates for the impact of the construction of the GQ highway network on forest
cover in the proximity of the highway, according to Equation 5. We define forest cover as log village forest cover
(Columns 1 and 3) and average covered share of each village pixel (Columns 2 and 4); the data source is Vegetation
Continuous Fields. The distance variables are indicators that identify places within a given distance band from the
GQ (Columns 1 and 2) or the NS-EW highway network (Columns 3 and 4). The omitted category is the band of
places at a distance of 200-300km from the highway network. These distance band indicators are then interacted
with time period indicators. The construction period (rows 1 through 4) is 2001 to 2004. The post period (rows
5 through 8) is 2005 to 2008. Columns 3 and 4 estimate a placebo specification with distances to the NS-EW
highway network, where construction had barely begun by 2008. The sample includes data from 2000 to 2008;
2000 is the omitted period. We omit years after 2008 as the placebo group is treated in those years. In Columns 3
and 4, we exclude places within 150km of the GQ network to prevent sample contamination. All estimates include
state-year fixed effects and standard errors are clustered at the subdistrict level to account for serial correlation.
A   Appendices: Additional Tables and Figures




                                    42
                                                                                                                                     43




                                                                 Figure A1
                                                        Road Program Completion Dates


                                               15,000
                   Number of Roads Completed




                                               10,000




                                                5,000




                                                   0
                                                        2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014




The figure shows the number of roads completed under the PMGSY road construction program, by year.
                                                                                                                                          44




                                             Figure A2
                                Regression Discontinuity Balance Tests


                                  Log Forest Cover (2000)                                     Average Forest Cover (2000)

             .1                                                         .2


             0                                                          0

            −.1                                                        −.2
                  −100       −50          0           50         100         −100         −50          0           50               100
                              Population Minus Threshold                                   Population Minus Threshold


                              Change in Log Forest (2000−2005)                          Share of Households Using Firewood (2001)
                                                                        .01
             .02
                  0
                                                                             0
            −.02
            −.04                                                       −.01
                      −100   −50         0          50           100             −100      −50         0          50                100
                             Population Minus Threshold                                    Population Minus Threshold


                                  Log Night Lights (2000)                                      Average Night Light (2000)

             .1                                                         .2

             0                                                          0

            −.1
                                                                       −.2
                  −100       −50          0           50         100         −100         −50          0           50               100
                              Population Minus Threshold                                   Population Minus Threshold




The figure displays a graphical form of the regression discontinuity balance test. Each graph shows the
means of a variable measured at baseline in bins defined by population relative to the rural road program
treatment threshold. The linear fits and standard errors are estimated from Equation 1. The vertical line
shows the treatment threshold; the jump in the fit at this line is the regression discontinuity treatment
estimate. The dependent variable in each panel (left-to-right, top-to-bottom) is (A) log forest cover in 2000;
(B) average forest cover in 2000; (C) change in log forest cover from 2000 to 2005; (D) the share of households
whose primary cooking fuel is firewood; (E) the log of night light luminosity in 2000; and (F) average night
light luminosity in 2000. In Panel C, we omit villages with roads built before 2006, to ensure that balance
estimates are not contaminated by the small number of treated villages before this date. All estimates
include district-population threshold fixed effects.
                                                                                                        45




                                              Figure A3
                                  Regression Discontinuity Density Test


                     .01
          Density




                    .005




                      0
                           −100          −50            0                  50               100
                                               Normalized Population


The figure displays a graph from a regression discontinuity density test (McCrary, 2008). The X axis shows
the population relative to the road program treatment eligibility threshold. The Y axis shows a kernel
estimate of the density of villages in a given normalized population band. The lines display non-parametric
fits to the density function along with 95% confidence intervals.
                                      Figure A4                                                                      46
                     Difference-in-Differences Estimates of Impact of
                       Rural Roads on Forest Cover (Long Panel)

                                                                          Full Sample
                                    .03      .02
                      Residual Log Forest Cover
                      0          .01−.01




                                                   <= −5   −4   −3     −2    −1      0     1      2   3   4   >= 5
                                                                        Years after Road Completion




                                                                     High Baseline Forest
                                    .03
                       Residual Log Forest Cover
                      0          .01−.01    .02




                                                   <= −5   −4   −3     −2    −1      0     1      2   3   4   >= 5
                                                                        Years after Road Completion




The figure shows year-by-year estimates of log forest cover in villages that received new roads between 2001
and 2013. The figure is identical to Figure 2, but with an additional estimate for the 5th year before and
after treatment. Villages are grouped on the X axis according to the year relative to road completion. Each
point thus shows the average value of log forest cover in villages in a given year relative to the treatment year,
controlling for village fixed effects, district*year fixed effects and baseline population * year and baseline log
forest cover * year interactions. Standard errors are clustered at the village level.
                                                                                                          47




                                 Table A1
          OLS Regressions of Forest Cover on Rural Road Indicators

                                                (1)           (2)       (3)       (4)                     (5)
Paved Road in 2001                          -0.162***     -0.270*** -0.041*** -0.025***               -0.018**
                                             (0.010)       (0.010)   (0.009)   (0.009)                 (0.009)
Population                                                0.476*** 0.862*** 0.896***                  0.910***
                                                           (0.013)   (0.012)   (0.012)                 (0.012)
Population2                                               -0.037*** -0.085*** -0.092***              -0.095***
                                                           (0.004)   (0.003)   (0.003)                 (0.003)
Distance in km to town of 10,000                                                                      0.007***
                                                                                                       (0.000)
Distance in km to town of 100,000                                                                       -0.000
                                                                                                       (0.000)
Constant                                     3.405***      3.083***      2.786***      2.764***       2.533***
                                              (0.004)       (0.008)       (0.007)       (0.007)        (0.013)
Fixed Effects                                   None          None          State         State        District
N                                             270871        270871        270871        270871         270871
r2                                             0.00           0.02          0.19         0.28            0.28
∗
  p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01
The table shows estimates from OLS regressions of village-level log forest cover in 2001 on an indicator variable
that takes the value one if a village has a paved road in 2001. Column 1 presents the bivariate estimates, and
Columns 2 through 5 present estimates with progressively greater numbers of controls and fixed effects. Forest
cover is calculated from Vegetation Continuous Fields. Population is measured in millions of people.
                                                                                                        48




                                       Table A2
                         Regression Discontinuity Balance Tests
                     Variable                                         RD Estimate
                     Log Forest (2000)                                      -0.012
                                                                          (0.027)
                     Average Forest (2000)                                  -0.092
                                                                          (0.101)
                     Share Cooking with Firewood                            -0.001
                                                                          (0.003)
                     Log Forest Change (2000-2005)                           0.012
                                                                          (0.012)
                     Mean Night Light (2000)                                -0.020
                                                                          (0.060)
                     Log Night Light (2000)                                 -0.033
                                                                          (0.029)
                     Number of Observations                                 55222
                     ∗
                       p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01

The table shows estimates from a regression discontinuity balance test. We run the regression discontinuity
specification defined by Equation 1 on variables measured before any rural road construction took place,
and report the reduced form treatment estimates. Row 4 (Log Forest Change 2000-2005) tests for pretrends
in forest cover; we exclude villages with roads built before 2006 for this sample. All estimates include
district-population threshold fixed effects.
                                           Table A3
                         Regression Discontinuity Estimates of Impact of
                       Rural Roads on Forest Cover (Alternate Bandwidths)

                                     Log Forest (2013)                                    Average Forest (2013)
                               (1)     (2)       (3)     (4)                          (5)     (6)      (7)       (8)
Above Population Threshold 0.015      0.003    0.002   -0.002                       0.124    0.042    0.035     0.017
                             (0.015) (0.011) (0.009) (0.008)                       (0.085) (0.060) (0.050) (0.043)
Bandwidth                       50     100      150     200                           50      100      150       200
N                             45112   89476 133908 178292                           45112    89476 133908 178292
r2                             0.80    0.80     0.80    0.80                         0.56     0.56     0.57      0.57
∗          ∗∗        ∗∗∗
  p < 0.10, p < 0.05, p < 0.01
The table shows reduced form regression discontinuity estimates of the impact of rural roads on forest cover. The specifi-
cations are comparable to those in Columns 2 and 3 of Table 2 but with alternate bandwidths defined in the bandwidth
row of the table. The sample includes forest cover estimates for years 2011 through 2013 for increased precision; robust
standard errors are clustered at the village level to account for serial correlation. All estimates include district-population
threshold fixed effects and a control for baseline forest cover.




                                                                                                                                  49
                                                                                        50




                         Table A4
       Regression Discontinuity Estimates of Impact of
            Rural Roads on Household Fuel Use

                             Imports              Local Non-Wood         Firewood
Above Population Threshold     -0.002                   0.002              0.000
                              (0.002)                 (0.006)             (0.006)
N                              22318                   22318               22318
r2                              0.28                    0.42                0.42
∗          ∗∗        ∗∗∗
  p < 0.10, p < 0.05, p < 0.01
The table shows reduced form regression discontinuity treatment estimates of the
effect of new village roads on village-level household fuel use, estimated with Equa-
tion 1. The dependent variable is the share of households in a village that use im-
ported fuel sources (primarily propane, Column 1); dung and crop residue (Column
2); and firewood (Column 3) as primary fuel sources for cooking. The dependent
variables are measured in 2011. In addition to district-population threshold fixed
effects, controls include the baseline fuel share reported in 2001 (at the subdistrict
level) and forest cover in 2000.
                                                                                          51




                          Table A5
        Difference-in-Differences Estimates of Impact of
        Rural Roads on Forest Cover (Robustness Tests)

                              (1)       (2)        (3)                       (4)
Award Period               -0.004** -0.006***  -0.010***                 -0.007***
                            (0.002)  (0.002)    (0.003)                   (0.002)
Completion Period            0.001    -0.000     -0.002                    -0.002
                            (0.002)  (0.002)    (0.004)                   (0.003)
District-Year F.E.            Yes       No        Yes                       Yes
Subdistrict-Year F.E.          No      Yes         No                        No
Village F.E.                  Yes      Yes        Yes                       Yes
Village Time Trends           Yes       No         No                        No
Panel Sample                  Full     Full   +/- 5 Years               +/- 4 Years
N                           689745   683025     374385                    482130
r2                            0.95     0.96       0.94                      0.94
∗
  p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01
The table shows difference-in-differences estimates of the impact of new village roads
on local forest cover, under alternate sample definitions. We define forest cover as log
village forest cover; the data source is Vegetation Continuous Fields. Specifications
are identical to those in Table 3 with the following changes. Column 1 includes
village-specific time trends. Column 2 uses subdistrict-year fixed effects instead of
district-year fixed effects. Column 3 restricts the sample to villages with roads for
which we can observe at least 5 years of data before road completion and 5 years
after. Column 4 does the same, with 4 years. The sample consists strictly of villages
that received new roads between 2001 and 2013, and were not accessible by paved
road in 2001. Award Period is an indicator variable that takes the value one for years
after a road contract was awarded and before the road was completed. Completion
period is an indicator variable that marks the years after a village’s new road was
built. All regressions include district*year fixed effects, village fixed effects, baseline
population * year fixed effects, and baseline forest * year fixed effects. Standard errors
are clustered at the village level to correct for serial correlation.
                                                                                                          52




                                  Table A6
                Difference-in-Differences Estimates of Impact of
             Rural Roads on Forest Cover (Alternate Specifications)

                             (1)              (2)            (3)           (4)             (5)                (6)
Award Period             0.004***           -0.002       -0.005***      -0.004**        -0.005**            -0.001
                          (0.001)          (0.001)        (0.002)        (0.002)         (0.002)           (0.001)
Completion Period        0.017***            0.000         0.002          0.001           0.003             -0.000
                          (0.001)          (0.002)        (0.002)        (0.002)         (0.003)           (0.002)
District-Year F.E.          Yes               Yes           Yes            Yes             Yes               Yes
Village F.E.                Yes               Yes           Yes            Yes             Yes               Yes
Village Time Trends.         No               Yes            No            Yes              No                No
Village Definition        Boundary         Boundary       Boundary       Boundary      5 km radius       50 km radius
N                         3359370          3359370        689745         689745          688275            688275
r2                          0.95             0.96           0.94           0.95            0.94              0.97
∗          ∗∗        ∗∗∗
  p < 0.10, p < 0.05, p < 0.01
The table shows difference-in-differences estimates of the impact of new village roads on local forest cover, under
alternate sample definitions. We define forest cover as log village forest cover; the data source is Vegetation Continuous
Fields. Specifications are identical to those in Table 3 with the following changes. Column 1 includes villages that did
not receive PMGSY roads at any time as part of the control group. Column 2 adds village-specific time trends to show
that the positive treatment estimate in Column 1 is driven by differential trends in never-treated villages. Columns
3 and 4 repeat these two specifications with the standard set of villages to show that village-specific time trends do
not affect our main estimates. Column 5 estimates the standard specification from Table 3 with treated villages only,
but the dependent variable includes forest cover in a 5km radius from the village centroid. Column 6 uses a 50km
centroid. Award Period is an indicator variable that takes the value one for years after a road contract was awarded
and before the road was completed. Completion period is an indicator variable that marks the years after a village’s
new road was built. All regressions include village fixed effects, baseline population * year fixed effects, and baseline
forest * year fixed effects. Standard errors are clustered at the village level to correct for serial correlation.
                                    Table A7                                                          53
                  Difference-in-Differences Estimates of Impact of
                            Highways on Forest Cover:
                       Straight Line Instrumental Variables

                                        GQ (Straight Line)                        NSEW (Straight Line)
                                     Log Forest Avg Forest                     Log Forest Average Forest
GQ Construction Period * (0-50km)    -0.144***    -0.442***                      -0.013         0.379
                                       (0.026)      (0.142)                     (0.055)       (0.314)
GQ Construction Period * (50-100km)  -0.181***    -0.749***                      -0.010         0.199
                                       (0.027)      (0.142)                     (0.052)       (0.297)
GQ Construction Period * (100-150km) -0.129***    -0.495***                       0.014        -0.336
                                       (0.029)      (0.164)                     (0.048)       (0.272)
GQ Construction Period * (150-200km) -0.086***     -0.362**                      -0.030        -0.157
                                       (0.031)      (0.178)                     (0.044)       (0.268)
GQ Post Period * (0-50km)            -0.113***    -0.422***                      -0.015         0.528
                                       (0.026)      (0.133)                     (0.053)       (0.332)
GQ Post Period * (50-100km)          -0.110***    -0.518***                      -0.061        -0.051
                                       (0.027)      (0.137)                     (0.054)       (0.329)
GQ Post Period * (100-150km)          -0.057**       -0.138                      -0.027        -0.374
                                       (0.028)      (0.154)                     (0.051)       (0.316)
GQ Post Period * (150-200km)           -0.050*       -0.186                      -0.051        -0.361
                                       (0.029)      (0.162)                     (0.042)       (0.320)
Distance 0-50km                       0.019***     0.179***                    -0.056***       -0.166
                                       (0.007)      (0.051)                     (0.017)       (0.801)
Distance 50-100km                     0.033***     0.282***                    -0.057***      1.903**
                                       (0.008)      (0.055)                     (0.016)       (0.886)
Distance 100-150km                    0.033***     0.273***                    -0.045***        0.480
                                       (0.007)      (0.051)                     (0.016)       (0.819)
Distance 150-200km                    0.024***     0.195***                     -0.019*        -0.342
                                       (0.005)      (0.039)                     (0.011)       (0.866)
N                                       26397        26397                       14958         14958
r2                                       0.90         0.90                         0.94          0.86
∗          ∗∗        ∗∗∗
  p < 0.10, p < 0.05, p < 0.01
The table shows reduced form estimates from regressions of distance band * time period interactions on forest
cover. Distance bands are calculated to straight line approximations of the Golden Quadrilateral (Columns 1 and
2) and North-South/East-West (Columns 3 and 4) highway corridors. The estimating equation is Equation 5. We
define forest cover as log village forest cover (Columns 1 and 3) and average covered share of each village pixel
(Columns 2 and 4); the data source is Vegetation Continuous Fields. The omitted distance category is the set of
subdistricts at a distance of 200-300km from each set of straight line approximations. The construction period
(rows 1 through 4) is 2001 to 2004. The post period (rows 5 through 8) is 2005 to 2008. Columns 3 and 4 estimate
a placebo specification with distances to the NS-EW highway network, where construction had barely begun by
2008. The sample includes data from 2000 to 2008; 2000 is the omitted period. We omit years after 2008 as the
placebo group is treated in those years. In Columns 3 and 4, we exclude places within 150km of the GQ network
to prevent sample contamination. All estimates include state-year fixed effects and standard errors are clustered
at the subdistrict level to account for serial correlation.
                                                                                    54



                      Table A8
Mechanism Tests for Impact of Highways on Deforestation:
          Employment in Wood-Using Firms

                                                       Wood Use       Logging
  (0-100km from GQ) * 1(Year == 1990)                     0.003       0.033***
                                                         (0.016)       (0.005)
  (100-200km from GQ) * 1(Year == 1990)                  -0.028*      0.018***
                                                         (0.014)       (0.003)
  (0-100km from GQ) * 1(Year == 1998)                     0.007       0.032***
                                                         (0.018)       (0.006)
  (100-200km from GQ) * 1(Year == 1998)                -0.044***      0.023***
                                                         (0.015)       (0.005)
  (0-100km from GQ) * 1(Year == 2005)                   0.052***
                                                         (0.017)
  (100-200km from GQ) * 1(Year == 2005)                   -0.011
                                                         (0.014)
  (0-100km from GQ) * 1(Year == 2013)                   0.081***      0.055***
                                                         (0.016)       (0.008)
  (100-200km from GQ) * 1(Year == 2013)                   -0.009      0.041***
                                                         (0.012)       (0.008)
  N                                                     1037954        766458
  r2                                                       0.17         0.06
  ∗
    p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01
  The table shows estimates of the impact of the Golden Quadrilateral highway
  on log employment in timber-related firms. The table shows the full specifi-
  cations used to generate Figure 5, panels A and B. The dependent variable is
  log employment in firms for which timber is the primary input (sawmilling,
  pulp and paper, manufacture of wooden containers, wooden furniture and cork
  boards, Column 1) and log employment in logging firms (Column 2). Each
  row shows the interaction of an indicator for a given distance band from the
  Golden Quadrilateral, interacted with an indicator for a given Economic Cen-
  sus year. The omitted distance category is 200-300km. The estimates thus
  show the difference between log employment in each sector/year/distance band
  with log employment in the same sector/year at distsance 200-300km from the
  highway. Logging is not specifically identified in the 2005 Economic Census,
  so this estimate is omitted. All regressions include state-year fixed effects and
  cluster standard errors at the subdistrict level.
                                                  Table A9
                            Mechanism Tests for Impact of Highways on Deforestation:
                                              Land and Fuel Use

                                                   Ag Land Share        Fuel: Firewood      Fuel: Imported      Fuel: Local Non-wood
(0-100km from GQ) * 1(Year == 1991)                  -0.085***
                                                      (0.016)
(100-200km from GQ) * 1(Year == 1991)                -0.080***
                                                      (0.014)
(0-100km from GQ) * 1(Year == 2001)                  -0.069***               -0.008             0.010***                  -0.001
                                                      (0.011)               (0.015)              (0.003)                 (0.015)
(100-200km from GQ) * 1(Year == 2001)                -0.066***             0.051***               -0.003                -0.049***
                                                      (0.009)               (0.012)              (0.003)                 (0.012)
(0-100km from GQ) * 1(Year == 2011)                  -0.097***               0.001              0.013***                  -0.014
                                                      (0.010)               (0.014)              (0.004)                 (0.014)
(100-200km from GQ) * 1(Year == 2011)                -0.054***             0.060***                0.004                -0.064***
                                                      (0.008)               (0.012)              (0.003)                 (0.012)
N                                                     813691                638458               638458                  638458
r2                                                      0.30                  0.36                 0.22                    0.40
∗
  p < 0.10,∗∗ p < 0.05,∗∗∗ p < 0.01
The table shows estimates of the impact of the Golden Quadrilateral highway on land and fuel use. The table shows the full specifications
used to generate Figure 5, panels C through F. The dependent variable is the share of village land dedicated to agriculture (Column 1); the
share of households in a village that use firewood (Column 2); imported fuel sources (primarily propane, Column 3); and dung and crop
residue (Column 4) as primary fuel sources for cooking. Each row shows the interaction of an indicator for a given distance band from the
Golden Quadrilateral, interacted with an indicator for a given Population Census year. The omitted distance category is 200-300km. The
estimates thus show the difference between the outcome variable in each sector/year/distance band with the outcome variable in the same
sector/year at distsance 200-300km from the highway. All regressions includes state-year fixed effects and cluster standard errors at the
subdistrict level.




                                                                                                                                              55