Highway to Success: The Impact of the Golden Quadrilateral Project for the Location and Performance of Indian Manufacturing

We investigate the impact of transport infrastructure on the organisation and efficiency of manufacturing activity. The Golden Quadrilateral (GQ) project upgraded a central highway network in India. Manufacturing activity grew disproportionately along the network. These findings hold in straight�?line IV frameworks and are not present on a second highway that was planned to be upgraded at the same time as GQ but subsequently delayed. Both entrants and incumbents facilitated the output growth, with scaling among entrants being important. The upgrades facilitated better industrial sorting along the network and improved the allocative efficiency of industries initially positioned on GQ.


Introduction
Adequate transportation infrastructure is an essential ingredient for economic development and growth. Rapidly expanding countries like India and China face severe constraints on their transportation infrastructure. Business leaders, policy makers, and academics describe infrastructure as a critical hurdle for sustained growth that must be met with public funding, but to date there is a limited understanding of the economic impact of those projects.
We study how proximity to a major new road network a¤ects the organization of manufacturing activity, especially the location of new plants, through industry-level sorting and the e¢ ciency of resource allocation.
We exploit a large-scale highway construction and improvement project in India, the Golden Quadrilateral (GQ) project. The analysis compares districts located 0-10 km from the GQ network to districts 10-50 km away, and we utilize time series variation in the sequence in which districts were upgraded and di¤erences in the characteristics of industries and regions that were a¤ected. Our study employs establishment-level data that provide new insights into the sources of growth and their e¢ ciency improvements.
The GQ upgrades stimulated signi…cant growth in organized manufacturing (formal sector) in the districts along the highway network, even after excluding the four major cities that form the nodal points of the quadrangle.
Long-di¤erenced estimations suggest output levels in these districts grew by 49% over the decade after the construction began. This growth is not present in districts 10-50 km from the GQ network nor in districts adjacent to another major Indian highway system that was scheduled for a contemporaneous upgrade but subsequently delayed. We further con…rm this growth e¤ect in a variety of robustness checks, including dynamic analyses and straight-line instrumental variables (IV) based upon minimal distances between nodal cities. As the 0-10 km districts contained a third of India's initial manufacturing base, this output growth represented a substantial increase in activity that would have easily covered the costs of the upgrades.
Decomposing these aggregate e¤ects, districts along the highway system experienced a signi…cant boost in the rate of new output formation by young …rms, roughly doubling pre-period levels. These entrants were drawn from industries intensive in land and buildings, suggesting the GQ upgrades facilitated sharper industrial sorting between the major nodal cities and the districts along the highway. Despite a substantial increase in entrant counts, the induced entrants maintained comparable size and productivity to control groups. The young cohorts, moreover, demonstrated a post-entry scaling in size that is rare for India and accounted for an important part of the output growth. We also observe heightened output levels from incumbent …rms that existed in these 0-10 km districts before the reforms commenced. This growth combines slightly higher survival rates with increases in plant size. Despite this aid to incumbent growth, the incumbent share of local activity declines due to the stronger entry e¤ects.
Looking at industries as a whole, the GQ upgrades improved the allocative e¢ ciency (e.g., Hsieh and Klenow 2009) for industries that were initially positioned along the GQ network. Similar improvements were not present in earlier periods nor for industries that were mostly aligned on the placebo highway system. These results suggest that the GQ upgrades shifted activity towards more productive plants in the most a¤ected industries. Among district traits, the GQ upgrades helped activate intermediate cities of medium population density, where some observers believe India's development has underperformed compared to China. We also …nd that local education levels were important for explaining the strength of the changes, but that various other potential adjustment 1 costs (e.g., labor regulations) were not.
Our project contributes to the literature on the economic impacts of transportation networks in developing economies, which is unfortunately quite small relative to its policy importance. Two studies consider India and the GQ upgrades speci…cally. Datta (2011) …nds evidence of improved inventory e¢ ciency and input sourcing for manufacturing establishments located on the GQ network almost immediately after the upgrades commenced.
These results connect to our emphasis on the GQ upgrade's impact for the organization of formal manufacturing activity. Khanna (2014) examines changes in night-time luminosity around the GQ upgrades, …nding evidence for a spreading-out of economic development. Both studies are further discussed below. In related work, Ghani et al. (2012) identify how within-district infrastructure and road quality aid the allocative e¢ ciency of manufacturing activity in local areas between rural and urban sites.
Beyond India, several recent studies …nd mixed evidence regarding economic e¤ects for non-targeted locations due to transportation infrastructure in China or other developing economies. 1 These studies complement the larger literature on the United States and those undertaken in historical settings. 2 This study is the …rst to bring plant-level data to the analysis of these highway projects. This granularity is not feasible in the most-studied case of the United States as the major highway projects mostly pre-date the United States'detailed Census data. As a consequence, state-of-the-art work like Chandra and Thompson (2000) and Michaels (2008) utilize aggregate data and broad sectors. The later timing of the Indian reforms a¤ords data that can shed light on many margins like entry behavior, misallocation, and distributions of activity. Moreover, prior work mostly identi…es how the existence of transportation networks impacts activity, but we can quantify the impact from investments into improving road networks compared to placebo networks that are not enhanced. This provides powerful empirical identi…cation, and the comparisons are informative for the economic impact of road upgrade investments, which are very large and growing. 3 The remainder of this paper is as follows: Section 2 gives a synopsis of highways in India and the GQ project.
Related literatures consider non-transportation infrastructure investments in developing economies (e.g., Du ‡o and Pande 2007, Dinkelman 2011) and the returns to public capital investment (e.g., Aschauer 1989, Munell 1990, Otto and Voss 1994. Several studies evaluate the performance of Indian manufacturing, especially after the liberalization reforms (e.g. Kochhar et al. 2006, Ahluwalia 2000, Besley and Burgess 2004. Some authors argue that Indian manufacturing has been constrained by inadequate infrastructure and that industries that are dependent upon infrastructure have not been able to reap the maximum bene…ts of the liberalization's reforms (e.g. Gupta et al. 2008, Gupta and Kumar 2010, Mitra et al. 1998. 3 Through 2006 and including the GQ upgrades, India invested USD 71 billion for the National Highways Development Program to upgrade, rehabilitate, and widen India's major highways to international standards. A recent Committee on Estimates report for the Ministry of Roads, Transport and Highways suggests an ongoing investment need for Indian highways of about USD 15 billion annually for the next 15 to 20 years (The Economic Times, April 29, 2012).

Data Preparation
We employ repeated cross-sectional surveys of manufacturing establishments carried out by the government of India. Our work studies the organized sector surveys that were conducted in 1994-95 and in the 11 years stretching from 1999-00 to 2009-10. In all cases, the survey was undertaken over two …scal years (e.g., the 1994 survey was conducted during 1994-1995), but we will only refer to the initial year for simplicity. This time span allows us three surveys before the GQ upgrades began in 2001, annual observations for …ve years during which the highway upgrades were being implemented, and annual data from this point until 2009. Estimations typically use 1994 or 2000 as a reference point to measure the impact of GQ upgrades. This section describes some key features of these data.
The organized manufacturing sector of India is composed of establishments with more than ten workers if the establishment uses electricity. If the establishment does not use electricity, the threshold is 20 workers or more. These establishments are required to register under the India Factories Act of 1948. The unorganized manufacturing sector is, by default, comprised of establishments which fall outside the scope of the Factories Act. The organized sector accounts for over 80% of India's manufacturing output, while the unorganized sector accounts for a high share of plants and employment . The results reported in this paper focus on the organized sector. 7 The organized manufacturing sector is surveyed by the Central Statistical Organization through the Annual Survey of Industries (ASI). Establishments are surveyed with state and four-digit National Industry Classi…cation (NIC) strati…cation. For most of our analysis, we use the provided sample weights to construct population-level estimates of organized manufacturing activity at the district level. Districts are administrative subdivisions of Indian states or union territories that provide more-granular distances from the various highway networks. We also construct population-level estimates of three-digit NIC industries for estimations of allocative e¢ ciency. 8 ASI surveys record economic characteristics of plants like employment, output, capital, raw materials, and land and building value. For measures of total manufacturing activity in locations, we aggregate the activity of plants up to the district level. We also develop measures of labor productivity and total factor productivity (TFP). Weighted labor productivity is simply the total output divided by the total employment of a district.
Unweighted labor productivity is calculated through averages across plants and is used in robustness checks.
TFP is calculated primarily through the approach of Sivadasan (2009), who modi…es the Olley and Pakes (1996) and Levinsohn and Petrin (2003) methodologies for repeated cross-section data. 9 7 In a companion piece, Ghani et al. (2013) also consider the unorganized sector and …nd a very limited response to the GQ upgrades. There are traces of evidence of the organized sector …ndings repeating themselves in the unorganized sector (e.g., heightened entry rates, forms of industry sorting discussed below), but the results are substantially diminished in economic magnitudes. These null patterns also hold true regardless of the gender of the business owner in the unorganized sector. This di¤erential is reasonable given the greater optimization in location choice that larger plants conduct and the ability of these plants to trade inputs and outputs at a distance. 8 For additional detail on the manufacturing survey data, see Nataraj (2011), Kathuria et al. (2010), Fernandes and Pakes (2008), Hasan and Jandoc (2010), and Ghani et al. (2014). 9 As the Indian data lack plant identi…ers, we cannot implement the Olley and Pakes (1996) and Levinsohn and Petrin (2003) methodologies directly since we do not have measures of past plant performance. The key insight from Sivadasan (2009) is that one can restore features of these methodologies by instead using the average productivity in the previous period for a closely matched industry-location-size cell as the predictor for …rm productivity in the current period. Once the labor and capital coe¢ cients are recovered using the Sivadasan correction, TFP is estimated as the di¤erence between the actual and the predicted output. This correction removes the simultaneity bias of input choices and unobserved …rm-speci…c productivity shocks. We also consider a residual regression approach as an alternative. For every two-digit NIC industry and year, we regress log value-added (output minus Repeated cross-sectional data do not allow panel analyses of …rms or accurate measures of exiting plants. The data do, however, allow us to measure and study entrants. Plants are distinguished by whether or not they are less than four years old. We will use the term "young" plant to describe the activity of these recent entrants.
Estimations also consider incumbent establishments operating in districts from 2000 or earlier.
The sample for long-di¤erenced estimations contains 311 districts. This sample is about half of the total number of districts in India of 630, but it accounts for over 90% of plants, employment, and output in the organized manufacturing sector throughout the period of study. The reductions from the 630 baseline occur due to the following reasons. First, the ASI surveys only record data for about 400 districts due to the lack of organized manufacturing (or its extremely limited presence) in many districts. Second, we drop states that have a small share of organized manufacturing. 10 Finally, we require manufacturing activity be observed in the district in 2000 and 2007/9 to facilitate the long-di¤erenced estimations over a consistent sample. 11 We measure the distance of districts to various highway networks using o¢ cial highway maps and ArcMap GIS software. Reported results use the shortest straight-line distance of a district to a given highway network, measured from the district's edge. We …nd very similar results when using the distance to a given highway network measured from the district centroid. The Empirical Appendix provides additional details on data sources and preparation, with the most attention given to how we map GQ traits that we ascertain at the project level to district-level conditions for pairing with ASI data. 12 Empirical speci…cations use a non-parametric approach with respect to distance to estimate treatment e¤ects.
We de…ne indicator variables for the shortest distance of a district to the indicated highway network (GQ, NS-EW) being within a speci…ed range. Most speci…cations use four distance bands: nodal districts, districts located 0-10 km from a highway, districts located 10-50 km from a highway, and districts over 50 km from a highway. In an alternative setup, the last distance band is further broken into 50-125 km, 125-200 km, and over 200 km.
Our focus is on the non-nodal districts of a highway. We measure e¤ects for nodal districts, but the interpretation of these results is di¢ cult as the highway projects are intended to improve the connectivity of the nodal districts. For the GQ network, we follow Datta (2011) in de…ning the nodal districts as Delhi, Mumbai, Chennai, and Kolkata. In addition, Datta (2011) describes several contiguous suburbs (Gurgaon, Faridabad, Ghaziabad, and NOIDA for Delhi; Thane for Mumbai) as being on the GQ network as "a matter of design rather than fortuitousness."We include these suburbs in the nodal districts. As discussed later when constructing our instrument variables, there is ambiguity evident in Figure 1 about whether Bangalore should also be considered a nodal city. The base analysis follows Datta (2011) and does not include Bangalore, but we return to this question. For the NS-EW network, we de…ne Delhi, Chandigarh, NOIDA, Gurgaon, Faridabad, Ghaziabad, Hyderabad, and Bangalore to be the nodal districts using similar criteria to those applied to the GQ network. raw materials) of plants on their log employment and log capital, weighting plants by their survey multiplier. The residual from this regression for each plant is taken as its TFP. We then take the average of these residuals across plants for a district.
1 0 These excluded states are Andaman and Nicobar Islands, Dadra and Nagar Haveli, Daman and Diu, Jammu and Kashmir, Tripura, Manipur, Meghalaya, Nagaland and Assam. The average share of organized manufacturing from these states varies from 0.2% to 0.5% in terms of establishment counts, employment or output levels. We exclude this group to ensure reasonably well measured plant traits, especially with respect to labor productivity and plant TFP. With respect to the latter, we also exclude plants that have negative value added.
1 2 Appendix materials and tables identi…ed in this paper are available online at http://www.people.hbs.edu/wkerr/. Looking at di¤erences in growth patterns by distance from the GQ network, 0-10 km districts exceed 10-50 km districts in every column but total employment growth. Moreover, in most cases, the growth in these very proximate districts also exceeds that in districts over 50 km away. The associated share changes in Panel D tend to be quite strong considering the big increases in the nodal cities that are factored into these share changes. 13

Empirical Analysis of Highways'Impact on Economic Activity
We …rst consider long-di¤erenced estimations that compare district manufacturing activity before and after the GQ upgrades. We use this approach as well for our placebo analyses and IV estimations. We then turn to dynamic estimations that consider annual data throughout the 1994-2009 period, followed by the industry-level sorting analyses and examinations of allocative e¢ ciency.

Long-Di¤erenced Estimations
Long-di¤erenced estimations compare district activity in 2000, the year prior to the start of the GQ upgrades, with district activity in 2007 and 2009 (average across the years). About 95% of the GQ upgrades were completed by the end of 2006. We utilize two surveys after the conclusion of most of the GQ upgrades, rather than just our …nal data point of 2009, to be conservative. Dynamic estimations below …nd that the 2009 results for many economic outcomes are the largest in districts nearby the GQ network. An average across 2007 and 2009 is a more conservative approach under these conditions. These estimations will also show that benchmarking 1994 or 1999 as the reference period would deliver very similar results given the lack of pre-trends surrounding the GQ upgrades.
Indexing districts with i, the speci…cation takes the form: The set D contains three distance bands with respect to the GQ network: a nodal district, 0-10 km from the GQ network, and 10-50 km from the GQ network. The excluded category includes districts more than 50 km from the GQ network. pro…le, female-male sex ratio, population share in urban areas, population share in scheduled castes or tribes, literacy rates, and an index of within-district infrastructure. The variables regarding access to national and state highways and railroads are measured at the end of the period and thus include some e¤ects of the GQ upgrades.
The inclusion of these controls in the long-di¤erenced estimation is akin to including time trends interacted with these initial covariates in a standard panel regression analysis.
The column headers of Table 2 list dependent variables. Columns 1-3 present measures of total activity, Columns 4-6 consider new entrants, Columns 7 and 8 document productivity outcomes, and Columns 9 and 10 report wage and labor cost metrics. Panel A reports results with a form of speci…cation (1) that only includes initial values of the outcome variable as a control variable. The …rst row shows increases in nodal district activity for all metrics. The higher standard errors of these estimates, compared to the rows beneath them, re ‡ect the fact that there are only nine nodal districts. Yet, many of these changes in activity are so substantial in size that one can still reject that the e¤ect is zero. We do not emphasize these results much given that the upgrades were built around the connectivity of the nodal cities. Because the d coe¢ cients are being measured for each band relative to districts more than 50 km from the GQ network, the inclusion or exclusion of the nodal districts does not impact results regarding non-nodal districts.
Our primary emphasis is on the highlighted row where we consider non-nodal districts that are 0-10 km from the GQ network. To some degree, the upgrades of the GQ network can be taken as exogenous for these districts.
Columns 1-3 …nd increases in the aggregate activity of these districts. The coe¢ cient on output is particularly strong and suggests a 0.4 log point increase in output levels for districts within 10 km of the GQ network in 2007/9 compared to 2000, relative to districts more than 50 km from the GQ system. As foreshadowed in Table 1, estimates for establishment counts and output in districts 0-10 km from the GQ network exceed the employment responses. These employment e¤ects fall short of being statistically signi…cant at a 10% level, and this is not due to small sample size as we have 76 districts within this range. Generally, the response around the GQ changes favored output over employment, which we trace out further below with industry-level analyses.
Columns 4-6 examine the entry margin by quantifying levels of young establishments and their activity. We …nd much sharper entry e¤ects than the aggregate e¤ects in Columns 1-3, and these entry results are very precisely measured. The districts within 0-10 km of GQ have a 0.8-1.1 log point increase in entry activity after the GQ upgrade compared to districts more than 50 km away.
Columns 7 and 8 report results for the average labor productivity and TFP in the districts 0-10 km from the GQ network. These average values are weighted and thus primarily driven by the incumbent establishments.
Labor productivity for the district increases (also evident in a comparison of Columns 2 and 3). On the other hand, we do not observe TFP growth using the Sivadasan (2009) approach, and unreported estimations …nd limited di¤erences between the TFP growth of younger and older plants (relative to plants of similar ages in the pre-period). This general theme is repeated below with continued evidence of limited TFP impact but a strong association of the GQ upgrades with higher labor productivity. Columns 9-10 …nally show an increase in wages and average labor costs per employee in these districts.
For comparison, the third row of Panel A provides the interactions for the districts that are 10-50 km from the GQ network. None of the e¤ects on the allocation of economic activity that we observe in Columns 1-6 for the 0-10 km districts are observed at this spatial band. This isolated spatial impact provides a …rst assurance that these e¤ects can be linked to the GQ upgrades rather than other features like regional growth di¤erences.
By contrast, Columns 7-10 suggest we should be cautious about placing too much emphasis on the productivity and wage outcomes as being special for districts neighboring the GQ network, since the patterns look pretty similar for all plants within 50 km of the GQ network. On the other hand, it is important to recognize that the productivity/wage growth in Columns 7-10 for the districts 10-50 km are coming from relative declines in activity that are evident in Columns 1-6. That is, the labor productivity of 0-10 km districts is increasing because output is expanding more than employment, but in the 10-50 km districts the labor productivity is increasing due to employment contracting more than output. The di¤erent foundations for the productivity and wage changes suggest that we should not reject the potential bene…ts of the GQ network on these dimensions, and we return to this issue below with a detailed analysis of productivity distributions for entrants and incumbents.
The remaining panels of Table 2 test variations on these themes. Panel B next introduces the longer battery of district traits described above. The inclusion of these controls substantially reduces the coe¢ cients for the nodal districts. More important, they also diminish somewhat the coe¢ cients for the 0-10 km districts, yet these results remain quite statistically and economically important. The controls, moreover, do not explain the di¤erences that we observe between districts 0-10 km from the GQ network and those that are 10-50 km away. Appendix Table 2 reports the coe¢ cients for these controls for the estimation in Panel B. From hereon, this speci…cation becomes our baseline estimate, with future analyses also controlling for these district covariates.
Panel C further adds in state …xed e¤ects. This is a much more aggressive empirical approach than the baseline estimations as it only considers variation within states (and thus we need to have districts located on the GQ network and those farther away together in individual states). This reduces the economic signi…cance of most variables, and raises the standard errors. Yet, we continue to see evidence suggestive of the GQ upgrades boosting manufacturing activity.
Panel D presents results about the di¤erences in the types of GQ work undertaken. Prior to the GQ project, there existed some infrastructure linking these cities. In a minority of cases, the GQ project built highways where none existed before. In other cases, however, a basic highway existed that could be upgraded. Of the 70 districts lying near the GQ network, new highway stretches comprised some or all of the construction for 33 districts, while 37 districts experienced purely upgrade work. In Panel D, we split the 0-10 km interaction variable for these two types of interventions. The entry results are slightly stronger in the new construction districts, while the labor productivity results favor the road upgrades. This latter e¤ect is strong enough that the total output level grows the most in the road upgrade districts. Despite these intriguing di¤erences, the bigger message from the breakout exercise is the degree to which these two groups are comparable overall.
Panel E extends the spatial horizons studied in Panel B to include two additional distance bands for districts 50-125 km and 125-200 km from the GQ network. These two bands have 48 and 51 districts, respectively. In this extended framework, we measure e¤ects relative to the 97 districts that are more than 200 km from the GQ network. Two key observations can be made. First, the results for districts 0-10 km are very similar when using the new baseline. Second, the null results generally found for districts 10-50 km from the GQ network mostly extend to districts 50-200 km from the GQ network. Even from a simple association perspective, the manufacturing growth in the period surrounding the GQ upgrades is localized in districts along the GQ network.
It is tempting to speculate that the steeper negative point estimates in Columns 4-6 suggest a "hollowingout" of new entry towards districts more proximate to the GQ system after the upgrades. This pattern would be similar to Chandra and Thompson's (2000) …nding that U.S. counties that were next to counties through which U.S. highways were constructed were adversely a¤ected. Chandra and Thompson (2000) described their results within a theoretical model of spatial competition whereby regional highway investments aid the nationallyoriented manufacturing industries and lead to the reallocation of economic activity in more regionally-oriented industries like retail trade. Unreported estimations suggest that this local reallocation is not happening for Indian manufacturing, at least in a very tight geographic sense. 14 For India, the evidence is more consistent with potential diversion of entry coming from more distant points. Either way, the lack of statistical precision for these estimations prevents strong conclusions in this regard.
Appendix Table 3 provides several robustness checks on these results. We …rst show very similar results when not weighting districts and including dropped outlier observations. We obtain even stronger results on most dimensions when just comparing the 0-10 km band to all districts more than 10 km apart from the GQ network, which is to be expected given the many negative coe¢ cients observed for the 10-50 km band. We also show results that include an additional 10-30 km band. These estimations con…rm a very rapid attenuation in e¤ects.
The appendix also shows similar (inverted) …ndings when using a linear distance measure over the 0-50 km range.
Appendix Table 4 documents alternative approaches to calculating labor productivity and TFP consequences.

Comparison of GQ Upgrades to NS-EW Highway
The stability of the results in Table 2 is encouraging, especially to the degree to which they suggest that proximity to the GQ network is not re ‡ecting other traits of districts that could have in ‡uenced their economic development.
There remains some concern, however, that we may not be able to observe all of the factors that policy makers would have known or used when choosing to upgrade the GQ network and designing the speci…c layout of the highway system. For example, policy makers might have known about the latent growth potential of regions and attempted to aid that potential through highway development.
We examine this feature by comparing districts proximate to the GQ network to districts proximate to the NS-EW highway network that was not upgraded. The idea behind this comparison is that districts that are at some distance from the GQ network may not be a good control group if they have patterns of evolution that do not mirror what districts immediately on the GQ system would have experienced had the GQ upgrades not occurred. This comparison to the NS-EW corridor provides perhaps a stronger foundation in this regard, especially as its upgrades were planned to start close to those of the GQ network before being delayed. The identi…cation assumption is that unobserved conditions such as regional growth potential along the GQ network were similar to those for the NS-EW system (conditional on covariates).
The upgrades scheduled for the NS-EW project were to start contemporaneous to and after the GQ project.
To ensure that we are comparing apples to apples, we identi…ed the segments of the NS-EW project that were to begin with the GQ upgrades and those that were to follow in the next phase. We use separate indicator variables for these two groups so that we can compare against both. Of the 90 districts lying within 0-10 km of the NS-EW network, 40 districts were to be covered in the 48 NS-EW projects identi…ed for Phase I. The empirical appendix provides greater detail on this division. Table 3 repeats Panel B of Table 2 and adds in four additional indicator variables regarding proximity to the NS-EW system and the planned timing of upgrades. In these estimations, the coe¢ cients are compared to districts more than 50 km from both networks. None of the long-di¤erenced outcomes evident for districts in close proximity to the GQ network are evident for districts in close proximity to the NS-EW network, even if these latter districts were scheduled for a contemporaneous upgrade. The placebo-like coe¢ cients along the NS-EW highway are small and never statistically signi…cant. The lack of precision is not due to too few districts along the NS-EW system, as the district counts are comparable to the distance bands along the GQ network and the standard errors are of very similar magnitude. The null results continue to hold when we combine the NS-EW indicator variables. Said di¤erently, with the precision that we estimate the positive responses along the GQ network, we estimate a lack of change along the NS-EW corridor.

Straight-Line Instrumental Variables Estimations
Continuing with potential identi…cation challenges, a related worry is that perhaps the GQ planners were better able to shape the layout of the network to touch upon India's growing regions (and maybe the NS-EW planners were not as good at this or had a reduced choice set). Tables 4a and 4b consider this problem using IV techniques.
Rather than use the actual layout of the GQ network, we instrument for being 0-10 km from the GQ network with being 0-10 km from a (mostly) straight line between the nodal districts of the GQ network.
The identifying assumption in this IV approach is that endogenous placement choices in terms of weaving the highway towards promising districts (or struggling districts 15 ) can be overcome by focusing on what the layout would have been if the network was established based upon minimal distances only. This approach relies on the positions of the nodal cities not being established as a consequence of the transportation network, as the network may have then been developed due to the intervening districts. This is a reverse causality concern, and an intuitive example is the development of cities at low-cost points near to mineral reserves that are accessed by railroad lines. Similar to the straight-line IV used in Banerjee et al. (2012), the four nodal cities of the GQ network were established hundreds or thousands of years ago, making this concern less worrisome in our context.
The exclusion restriction embedded in the straight-line IV is that proximity to the minimum-distance line only a¤ects districts in the post-2000 period due to the likelihood of the district being on the GQ network and experiencing the highway upgrade. This restriction could be violated if the districts along these lines possess characteristics that are otherwise connected to growth during the post-2000 period. For example, these districts could generally have had more-skilled workforces than other districts, and perhaps these educational qualities became more important after 2000. The districts may also have possessed more favorable spatial positions. To guard against these concerns, we will estimate the IV with and without the battery of covariates for district traits in 2000. 16 Panel B of Figure 1 shows the implementation. IV Route 1 is the simplest approach, connecting the four nodal districts outlined in the original Datta (2011) study. We allow one kink in the segment between Chennai and Kolkata to keep the straight line on dry land. IV Route 1 overlaps with the GQ layout and is distinct in places. We earlier mentioned the question of Bangalore's treatment, which is not listed as a nodal city in the Datta (2011) work. Yet, as IV Route 2 shows, thinking of Bangalore as a nodal city is visually compelling. We thus test two versions of the IV speci…cation, with and without the second kink for Bangalore.
Panel A of Table 4a provides a baseline OLS estimation similar to Panel A of Table 2. For these IV estimations, we drop nodal districts (sample size of 302 districts) and measure all e¤ects relative to districts more than 10 km from the GQ network. This approach only requires us to instrument for a single variable-being within 10 km of the GQ network. Panel B shows the reduced-form estimates, with the coe¢ cient for each route being estimated from a separate regression. The reduced-form estimates resemble the OLS estimates for many outcomes.
The …rst-stage relationships are quite strong. IV Route 1, which does not connect Bangalore directly, has a …rst-stage elasticity of 0.43 (0.05) and an associated F-statistic of 74.5. IV Route 2, which treats Bangalore as a connection point, has a …rst-stage elasticity of 0.54 (0.05) and an associated F-statistic of 138.1. Panel C presents the second-stage results. Not surprisingly, given the strong …t of the …rst-stage relationships and the directionally similar reduced-form estimates, the IV speci…cations generally con…rm the OLS …ndings. In most cases, we do not statistically reject the null hypothesis that the OLS and IV results are the same. Wage and labor productivity are the two exceptions, where the IV indicates that OLS underestimates the true impact.
In Table 4b, we repeat this analysis and further introduce the district covariates measured in 2000 that we modelled in Panel B of Table 2. 17 When doing so, the …rst-stage retains reasonable strength. IV Routes 1 and 2 have associated F-statistics of 13.9 and 20.9, respectively. The covariates have an ambiguous e¤ect on the reduced-form estimates, being very similar for aggregate outcomes, generally lower for entry growth, and generally higher for productivity and wage e¤ects. Most results carry through, although the second-stage coe¢ cients for employment and output entry are substantially lower. Among the controls added, the inclusion of the total population control is the most important for explaining di¤erences between Table 4a and 4b. We again do not statistically reject the null hypothesis that the OLS and IV results are the same for most outcomes. We obtain similar outcomes when also including controls for distance from India's 10 largest cities.
On the whole, we …nd general con…rmation of the OLS …ndings with these IV estimates, which help with particular concerns about the endogenous weaving of the network towards certain districts with promising potential. IV estimates indicate that there may be an upward bias in the entry …ndings, perhaps due to endogenous placement towards districts that could support signi…cant new plants in terms of output. A second alternative is that the GQ upgrades themselves had a particular feature that accentuates these metrics (e.g., high output levels of contracted plants to support the actual construction of the road). This latter scenario seems unlikely, however, given the industry-level patterns documented below.

Dynamic Speci…cations
Dynamic patterns around these reforms provide additional assurance about the role of the GQ upgrades in these economic outcomes and insight into their timing. A …rst step is to estimate our basic …ndings in a pre-post format. We estimate this panel regression using non-nodal districts within 50 km of the GQ network. We thus estimate e¤ects for 0-10 km districts compared to those 10-50 km apart from the GQ highways. 18 Indexing districts with i and time with t, the panel speci…cation takes the form: The distance indicator variable takes unit value if a district is within 10 km of the GQ network, and the P ostGQ t indicator variable takes unit value in the years 2001 and afterwards. The panel estimations include a vector of district …xed e¤ects i and a vector of year …xed e¤ects t . The district …xed e¤ects control for the main e¤ects of distance from the GQ network, and the year …xed e¤ects control for the main e¤ects of the post-GQ upgrades period. Thus, the coe¢ cient quanti…es di¤erences in outcomes after the GQ upgrades for those districts within 10 km of the GQ network compared to those 10-50 km away. in all …ve surveys. The results are quite similar to the earlier work, especially for the entry variables. The total activity variables in Columns 1-3 are somewhat diminished, however, and we will later describe the time path of the e¤ects that is responsible for this deviation. The productivity and wage estimations show weaker patterns, which is to be expected given how close the two bands looked in Table 2's analysis. We report them for completeness, but we do not discuss their dynamics further. 19 Panel B studies the actual completion dates of the GQ upgrades. Due to the size of the GQ project, some sections were completed earlier than other sections. We model this by extending the 0-10 km indicator variable to also re ‡ect whether the district's work was completed by March 2003 (27 districts (2) to take a non-parametric dynamic format: Rather than introduce post-GQ upgrades variables, we introduce separate indicator variables for every year starting with 1999. We interact these year indicator variables with the indicator variable for proximity to the GQ network. The vectors of district and year …xed e¤ects continue to absorb the main e¤ects of the interaction terms. Thus, the t coe¢ cients in speci…cation (3) quantify annual di¤erences in outcomes for 0-10 km districts compared to those 10-50 km away, with 1994 serving as the reference period. These estimations include 1188 observations as the cross of 12 years with the 99 non-nodal districts within 50 km of the GQ network for which we can always observe their activity.
By separately estimating e¤ects for each year, we can observe whether the growth patterns appear to follow the GQ upgrades hypothesized to cause them. Conceptually, we also believe this dynamic approach is a better way of characterizing the impact of the GQ upgrades than the speci…c completion dates of segments. Once the upgrades started, work began all along the GQ network and proceeded in parallel. Every state along the GQ network had at least one segment completed within the …rst two years of the program. Work continued thereafter across all states, with the average spread of completion times between the …rst and last segments for states being 6.4 years. Since manufacturing activity and location choice decisions can easily be in ‡uenced by upgrades on nearby segments (and even anticipation of future upgrades to a segment), we believe it more appropriate to model the GQ event as a whole, timing the impact of all segments from 2001.
Panels A and B of Figure  Panels C and D show the series for log total plant counts and output. Aggregate plant counts are very stable before the upgrades start. There is some measure of a downward trend in output levels for 0-10 km districts before the reform, but these pre-results are not statistically di¤erent from each other nor from 1994's levels.
After the GQ upgrades start, total plant counts and output also climb and then stabilize, before climbing again as the sample period closes. At all points during this post period, the coe¢ cient values are positive, indicating an increase over 1994 levels, but the di¤erences are not statistically signi…cant until the end.
The paths depicted in these …gures provide important insights. The young entrant measures in Panels A and B are in essence ‡ow variables into the district. Thus, comparing the post-2006 period to 2004, it is not that the earlier cohort of young …rms is shrinking. Instead, a surge of entry occurred as the GQ upgrades made areas more accessible, and with time this surge abated into a lower sustained entry rate that still exceeded pre-reform levels. By contrast, the metrics in Panels C and D are stock variables. Thus, their gradual development over time as more entrants come in and the local base of …rms expands makes intuitive sense.
We began in Table 2 by considering long-di¤erenced speci…cations that compare activity in 2000 with 2007/9. period would generally lead to larger e¤ects due to the dip in some variables. To this end, the appendix shows that the downward shift in output in Figure 2 is by far the largest pre-movement among the outcomes considered.
Encouragingly, there is no evidence of a pre-trend that upward biases our work with any outcome variable.

Entrants and Incumbents
Plant-level data o¤er the opportunity to examine the roles of entrants and incumbents in aggregate growthwhether the growth is mainly through the displacement of older plants by new entrants, within-plant productivity growth, or some combination of the two. The ideal scenario for this analysis is to have panel data on plants (e.g., Glaeser et al. 2013). While we unfortunately lack this panel structure, we can use information on the ages of plants to consider cohorts over time.
Columns 1-3 of Table 6 consider the role of incumbents in this growth by estimating the log activity in 2007/9 due to plants that have been alive at least 10 years compared to the total initial activity of the district in 2000.
The positive coe¢ cient in Column 1 for the 0-10 km group suggests that a greater fraction of the …rms already present in the 0-10 km districts by 2000 (i.e., before the GQ upgrades began) survived to 2007/9 than …rms in their peer cohorts in districts farther away from the GQ network. Columns 2 and 3 further show that employment and output increased disproportionately for these incumbent …rms. Moreover, the relative magnitudes of Columns 1-3 emphasize a point made earlier about the productivity results. For the 0-10 km districts, output is rising at a faster pace than employment, leading towards higher labor productivity at the same time that plant survival is also growing. By contrast, incumbents in the 10-50 km districts are closing at a similar rate or even faster than the control group. These more-distant plants are also shedding employment faster than output. As a result, their labor productivity is also rising, but the origin of this productivity growth is very di¤erent from the districts near to the GQ network.
Columns 4-6 quantify the role of entrants by considering as the outcome variable the log activity in 2007/9 due to plants that have been alive fewer than 10 years compared to the total initial activity of the district in 2000.
This young …rm activity is measured against the same baseline as in Columns 1-3, but is important to note that the relative coe¢ cient sizes measure proportionate e¤ects and thus do not directly rank order aggregate e¤ects (further discussed in the closing section). The outcome measures are all very strong for the 0-10 km districts.
There is also some evidence suggestive of larger entrants being less likely to locate in the 10-50 km band.
Thus, both entrants and incumbents contribute to the aggregate growth evident in Table 2. The last three columns consider as an outcome variable the share of activity in each district in 2007/9 contained in …rms that have been alive ten or more years. Despite their better survival rates and growth compared to distant incumbent peers, the share of activity accounted for by incumbent …rms in districts along the GQ network declines. 21 Table 7 further analyzes the productivity distributions and selection margins in districts by distance from the GQ network. We …rst normalize the plant-level productivity estimates developed using the Sivadasan (2009) methodology by dividing by the employment-weighted average productivity estimate for an industry-year. We then calculate in Column 1 the average normalized TFP in 2000 for plants within districts by distance from the GQ network. These entries sum over all industries and plants within each district group, weighting individual observations by employment levels. Normalized productivity levels are naturally centered around one and are somewhat higher in nodal districts-a typical …nding in urban productivity patterns-with the further initial di¤erences over the other distance bands being marginal.
Column 2 provides a similar calculation in 2007/9. The normalization process again centers values around one, such that aggregate TFP growth is removed at the industry level. The percentage listed next to each entry in Column 2 is the average value in 2007/9 compared to that in 2000. Overall, there is limited movement for any of the groups; the 0-10 km range increases slightly, while the other three ranges show very small declines. This pattern is possible because the 0-10 km group is becoming larger during the period of study.
The more interesting tabulations are in Columns 3 and 4. In Column 3, we calculate these TFP averages for plants that are at least ten years old in 2007/9, while Column 4 presents the comparable …gures for plants less than ten years old. Productivity rises with plant age, such that the values in Column 3 are higher than in Column 4. This may be due to di¤erences in technical e¢ ciency. As pointed out in Foster et al. (2008), these di¤erences could also result from using revenues for TFP calculations rather than physical products, if, for example, young …rms have lower prices to build demand. Either way, our focus is on the relative comparisons back to the initial 2000 period that are expressed in the accompanying percentages.
Column 3 shows that surviving incumbent plants in the 10-50 km range from the GQ network have substantially higher TFP compared to initial values than in the other district ranges. These districts have reduced entry rates, and the entrants have lower TFPs compared to the other bands. By contrast, the TFP distributions for the 0-10 districts have a more homogeneous adjustment over entrants and incumbents. The stability in entrant TFP in these districts is important given the massive increase in entry rates associated with the GQ upgrades.
Despite these surges, the TFP positions are not weakening compared to districts that are 50 km or more from the GQ network. This comparison shows again the very di¤erent sources of productivity development in the 0-10 km versus 10-50 km ranges surrounding the GQ network with the upgrades.

District and Industry Heterogeneity
The remaining analyses quantify heterogeneity in e¤ects for districts and industries. Table 8 considers district heterogeneity using the long-di¤erenced speci…cation (1). Articulating this heterogeneity is challenging empirically because the data variation becomes very thin as one begins to partition the sample by additional traits beyond proximity to the GQ network. We take a simple approach by allowing the coe¢ cient on 0-10 km districts to vary by whether the district is above or below the median value for a trait. Panel A reports the baseline estimation, and we include unreported main e¤ects for interactions in Panels B-E.
Panels B and C document the two key dimensions that we have identi…ed. Districts along the GQ network with higher population density and literacy rates show a stronger response. Given that these density levels are less than in nodal cities that are excluded from the analysis, this response provides some support for the hypothesis that intermediate-sized districts were particularly aided by the GQ infrastructure. 22 By contrast, Panels D and E do not …nd prominent di¤erences when looking at within-district infrastructure levels or distances along the GQ network from nodal cities. 23 Finally, unreported analyses investigate whether labor regulations play a role in these adjustment patterns. Using the employment protection and industrial dispute resolution laws from Ahsan and Pages (2008), we do not …nd evidence that districts located in states with above or below average stringency to their labor regulations respond di¤erently to each other. At least on this widely discussed policy dimension (Besley and Burgess 2004), local policy conditions display a weaker connection than workforce factors like literacy rates, which could perhaps themselves be seen as barriers of adjustment. We do not push this interpretation strongly given that we are unable to assess other possible dimensions like entry regulations or corruption levels. Table 9a describes a key feature of the industry heterogeneity in entry that occurred after the GQ upgrades.
We focus speci…cally on the land and building intensity of industries. We select this intensity due to the intuitive inter-relationship that non-nodal districts may have with nodal cities along the GQ network due to the general greater availability of land outside of urban centers and its cheaper prices. This general urban-rural or coreperiphery pattern is evident in many countries and is associated with e¢ cient sorting of industry placement.
Moreover, this feature has particular importance in India due to government control over land and building rights, leading some observers to state that India has transitioned from its "license Raj" to a "rents Raj" (e.g., Subramanian 2012a,b). Given India's distorted land markets, the heightened connectivity brought about by the GQ upgrades may be particularly important for e¢ cient sorting of industry across spatial locations. We measure land and building intensity at the national level in the year 2000 through the industry's land and building value per unit of output (listed in Appendix Table 7).
In Table 9a, we repeat the entry speci…cations isolating activity observed for industries in three bins: those with low land intensity (the bottom quartile of intensity), medium intensity (the second quartile), and high intensity (the top two quartiles). These estimations use the long-di¤erenced approach in speci…cation (1). The 0-10 km districts show a pronounced growth in entry by industries that are land and building intensive. Especially for young …rm establishments and output, the adjustment is weaker among plants with limited land and building intensities compared to the top half (there are no important di¤erences between the two quartiles in the top half). As remarkable, the opposite pattern is generally observed in the top row for nodal districts-where nodal districts are experiencing heightened entry of industries that are less land and building intensive after the GQ upgrades-and no consistent patterns are observed for districts 10-50 km from the GQ network. Table 9b shows a similar picture after including district controls and state …xed e¤ects, and Appendix Tables 8a and 8b show instead a weak or opposite relationship is evident with labor and materials intensity. Using capital intensity to group industries not surprisingly gives similar results to land and building intensity. These patterns suggest that the GQ upgrades aided the e¢ cient sorting of industries across locations along the network.

Changes in Allocative E¢ ciency
Our …nal exercise takes up directly the allocative e¢ ciency of the Indian economy. In a very in ‡uential paper, Hsieh and Klenow (2009) describe the degree to which India and China have a misallocation of activity toward unproductive manufacturing plants. That is, India has too much employment in plants that have low e¢ ciency, and it has too little employment in plants with high e¢ ciency levels. We evaluate whether the GQ upgrades are connected with improvements in allocative e¢ ciency for industries that were mostly located near the GQ network in 2000, compared to those that were mostly o¤ of the GQ network. The hypothesis is that allocative e¢ ciency will improve most in industries that were initially positioned around the GQ network. This could be due to internal plant improvements in operations, increases in competition and the entry/exit of plants, and adjustments in price distortions.
Quantifying improvements in allocative e¢ ciency is quite di¤erent than the district-level empirics undertaken thus far as we must look at the industry's production structure as a whole. We thus calculate for the 55 three-digit  to the NS-EW network. These results are robust to controlling for the land and building intensity of an industry, calculating proximity to the GQ network using a 50 km range, and similar exercises. With 55 data points, there are natural limits on the extensions and robustness checks that can be undertaken, but these exercises provide some con…dence in the conclusion that the GQ upgrades had a positive impact for the allocative e¢ ciency of India's manufacturing sector. This impact may have been particularly strong for industries where new activity was already occurring in a modest band around the network. 25

Discussion and Conclusions
This study …nds that the GQ upgrades led to a substantial increase in manufacturing activity. This growth included higher entry rates, incumbent productivity expansion, adjustments in the spatial sorting of industries, and improved allocative e¢ ciency in the manufacturing industries initially located along the GQ network. We close this paper by further discussing the economic magnitudes of the GQ reforms by the end of the sample period and what might lie ahead. These discussions, by their nature, require stepping beyond the econometric analyses conducted, and the assumptions made below are important for the insights derived. We focus on output for this analysis. The organized sector accounts for over 80% of Indian manufacturing output, and much of the work in this paper points to a central connection of the GQ upgrades to output growth.
We start with the impact of the upgrades for the 0-10 km districts. Our preferred estimates of output impacts to the end of the sample period are the 0.43 (0.16) and 0.37 (0.21) coe¢ cients in Panels B and C of Table 2, respectively. These estimates are quite robust to speci…cation checks, and they grow somewhat in magnitude in the IV analyses. Taking a mid-point of 0.4 suggests a 49% overall output increase from initial values for the average district located near the GQ network. Compared to pre-period levels, this would be an increase of output levels for the average district from USD 1.8 billion to USD 2.6 billion. The actual increase, for reference, was to USD 3.8 billion. The estimates would thus credit about 43% of the observed increase to the GQ upgrades, with the rest due to general expansion of Indian manufacturing and the accentuated development of manufacturing along other dimensions that these districts possess (e.g., education levels).
Looking forward, it seems that the scope for further increase in output is relatively modest. While Panel D of Figure 2 shows sustained growth from 2006-2009 in total output levels, we expect this upward trend to ‡atten and stabilize for the long-term. This forecast comes from consideration of the drivers of growth: 1) continued new entry along the GQ network, 2) the growth to scale of past entrants to these districts since the upgrades, and 3) the growth of incumbents present from before 2000. The second and third e¤ects combine multiple elements-selection and exit among …rms, output expansion among survivors and reallocation, and within-…rm 2 5 Unreported estimations suggest that increases in price competition may have played a role. We generally …nd accentuated declines in output price dispersion over the 2000 to 2007/9 period for industries located closer to the GQ network. As one example, every 10% increase in initial industry employment within 200 km of the GQ network is associated with a -0.14 (0.09) change in output price dispersion, with the latter measured in unit standard deviations. These results, however, are not statistically signi…cant, and we do not observe quality dispersions nor consumer prices by region. As such, we do not strongly emphasize this channel but note suggestive evidence in this regard. productivity growth. Repeated cross-sectional data cannot perfectly separate these e¤ects, but we can provide some informative calculations.
The …rst e¤ect-the ongoing heightened rate of new output formation-appears to have reached its long-term level. The dynamic estimation in Panel B of Figure 2 suggests that the increment of new entry over initial levels for districts along the GQ network stabilized by 2006-2009. The heightened output formation at this time is 40% less than at its 2004 peak. Similarly, the share of establishments that are young …rms is stable in these districts at 17%-18% in 2007-2009, after an increase from 14% in 2000. This e¤ect added about 11% output overall to the average district compared to initial levels.
The second driver-growth to scale of young cohorts of entrants-is often thought to be weak in the Indian context (e.g., Hsieh and Klenow 2014). In this context, however, we observe some growth due to scaling in the 0-10 km districts. Yet, there does not appear to be much scope for further growth in this regard either. The average size of all entrants since 2000 is 86% of all remaining pre-2000 incumbents by 2009; for the most recent cohort, the size ratio is 93%. Thus, while some further growth stimulus may occur through growth of recent entrants, it seems unlikely from these conditions that such increments will be very large.
The third factor-growth of the surviving incumbents-is also unlikely to be a source of further growth. By 2009, this group accounts for about 68% of output in these districts. While their output level increased, it has not kept pace with the district as a whole. This surviving group has been strongly selected during the 2000-2009 period, so we do not anticipate a mass exodus nor rapid growth for them.
Having reviewed these three components, we believe GQ's output growth may increase beyond 49%, but it is di¢ cult to project it being much higher. Most of the long-term impact of the GQ upgrades on the 0-10 km districts appears to have been achieved. One uncertainty in this forecast is the role of entrants and reallocation.
The data suggest stability is setting in on the rate of reallocations, but it is possible that reallocations could accelerate again in the future if some industries or plants face very large moving costs (e.g., recently built plants, agglomeration economies) that further weaken. Another source of uncertainty is how continued development of other infrastructure projects in India will cut into the connectivity advantage of districts along the GQ network (recall that the GQ upgrades were the …rst of seven phases planned).
With repeated cross-sectional data, we are not able to precisely decompose the aggregate growth gains into the three channels described above. Some rough calculations suggest a balance between entrants (the …rst two channels) and incumbents (the third channel). The 2009 output level of pre-2000 incumbents is 148% of the district's size in 2000 for the 0-10 km districts; as a comparison point, the similar ratio for the 50+ km districts is 138%. These percentages include initial levels, and thus the net growth among pre-2000 incumbents is 48% and 38%, respectively. The 2009 output level of post-2000 entrants is 70% of the district's size in 2000 for the 0-10 km districts, with the comparison point now being 67%. Thus, in both distance bands, entrants account for a majority of net output growth, and the increment for the 0-10 km districts is roughly comparable for both incumbents and entrants, suggestive of mostly balanced growth roles that somewhat favor entrants.
With this long-term portrait, we can also make a cautious ballpark estimate of the overall e¤ect by considering the four spatial bands. We …rst make the important assumption that the GQ upgrades had no e¤ect on the 236 districts that are 50 km beyond the GQ network. This group, which accounted for 43% of India's output in the initial period, has served as a frequent control group and does not appear to be impacted substantially by the GQ network. Some evidence for this assumption is found in the NS-EW placebo that compares distant districts from the GQ network by their proximity to the NS-EW network. Finding no major di¤erences across sub-groups of districts in this analysis suggests that aggregate e¤ects in the distant spatial range were unlikely to be material.
With this assumption in place, we next conclude from our analytical work that weak e¤ects accrued to districts 10-50 km from the GQ network, which accounted for a little under 10% of initial output levels. There may have been some declines, as evidenced by the small negative coe¢ cients in Table 2, but these results are not statistically signi…cant. Even taking the point estimates as true values, the overall magnitude would still be small (1%-2% of total Indian manufacturing) given that the declines would be applied against a small share of the total manufacturing output in India.
The biggest growth comes from the 0-10 km districts, which accounted for 34% of initial levels. If long-term output development is modest from its current level, as argued above, we can use the 49% growth estimate. This would suggest a 17% output growth contribution overall for India's manufacturing sector.
Finally, we come to the question of the nodal districts, which represented 15% of initial output. Strictly speaking, the estimations would suggest that this group added a further 9% output expansion. Some portion of this growth may be due to GQ upgrades, but it is likely the case that a majority of the credit belongs elsewhere.
For representative purposes, a value of 2% is assigned.
Thus, on a whole, our ballpark calculation would suggest that the GQ upgrades increased manufacturing output by 15%-19%. For reference, Indian manufacturing output doubled during this period, and 37% of this growth occurred in the non-nodal 0-10 km districts along the GQ network. The estimates would thus credit something less than a …fth of the organized sector growth to better connectivity provided by the enhanced GQ network, with all of that impact concentrated on adjacent districts. This calculation makes clear the key assumption underlying the analysis-the only factors pushing forward these nearby districts, where the majority of this manufacturing growth occurred, were the GQ upgrades or the other traits explicitly modeled. 26 These powerful e¤ects may be localized to organized manufacturing. Khanna (2014) examines changes in night-time luminosity around the GQ upgrades. He …nds evidence that the upgrades yielded weaker spatial di¤erentials in terms of night-time lights, suggestive of a spreading out of economic development. There are su¢ cient overlaps in our methodologies that the di¤erences between the papers are not likely to be due to choices regarding empirical technique. Instead, it seems more likely that the organized manufacturing sector reacted di¤erently from other forms of economic activity, and Section 3 noted the signi…cantly weaker responses observed for unorganized manufacturing. Our …ndings thus point to a particular connection of the highway upgrades to the spatial development, sorting, and allocative e¢ ciency of large-scale manufacturing plants that ship inputs and outputs at a distance. Improvements in this sector's e¢ ciency and spatial organization can generate substantial economic gains.
These outcomes are an important input into policy choices in India and other developing economies. This paper provides quantitative estimates of the likely economic growth associated with highway upgrades and the spatial impacts of these infrastructure projects. These estimates are most applicable to similar settings. For example, the GQ upgrades were the …rst phase of an overhaul to the inadequate transportation infrastructure of India. Thus, the impact may be larger than what would be observed from marginal investments into more developed settings. Similarly, the projects were undertaken during a period of economic growth, possibly allowing for greater rates of entry and more adjustment in the location choice decisions of new entrants than in economies with stagnant growth (where infrastructure investment can also be seen as a way to boost the economy). On the whole, our project speaks to the severe constraints that inadequate infrastructure can have for the development of manufacturing in emerging economies and the potential growth that may follow from alleviating that constraint.  Notes: Panels A and C illustrate the dynamics of young entrant and total plant count growth for non-nodal districts located 0-10 km from the GQ network relative to districts 10-50 km from the GQ network. The solid line quantifies the differential effect for the GQ upgrades by year, with 1994 as the reference year. Dashed lines present 90% confidence intervals, with standard errors clustered by district. Panels B and D consider comparable output estimations. Appendix Table 6a reports complete regression results.      Table 2. Long-differenced estimations compare results from proximity to the GQ network to the NS-EW highway network that was planned for partial upgrade at the same time as the GQ project but was then delayed. Phase I portions of the NS-EW upgrade were planned to overlap with the GQ upgrades but were postponed. The regressions control for the initial district conditions listed in Table 2.
Effects for districts based upon distance from the NS-EW network: Effects for districts based upon distance from the GQ network: DV: Change in manufacturing trait listed in column header  Table 2. Panel A modifies the base OLS estimation to exclude nodal districts and measure effects relative to districts 10+ km from the GQ network. This sample contains 302 districts. Panel B reports reduced-form estimations of whether or not a district edge is within 10 km of a straight line between nodal districts. Panel C reports IV estimations that instrument being within 10 km from the GQ network with being within 10 km of the straight line between nodal districts. Route 1 does not connect Bangalore directly, with the first-stage elasticity of 0.43 (0.05) and the associated F-statistic of 74.5. Route 2 treats Bangalore as a connection point, with the first-stage elasticity of 0.54 (0.05) and the associated F-statistic of 138.1. The null hypothesis in the exogeneity tests is that the instrumented regressor is exogenous.
B. Reduced-form estimates for distance from a straight-line between nodal districts C. IV estimates using distance from a straight-line between nodal districts A. Base OLS estimation that excludes nodal districts and measures effects relative to districts 10+ km from the GQ network DV: Change in manufacturing trait listed in column header   Table 2 other than road and railroad access variables. Route 1 does not connect Bangalore directly, with a firststage elasticity of 0.38 (.05) and associated F-statistic of 13.9. Route 2 treats Bangalore as a connection point, with a first-stage elasticity of 0.49 (.05) and associated F-statistic of 20.9.
A. Base OLS estimation that excludes nodal districts and measures effects relative to districts 10+ km from the GQ network B. Reduced-form estimates for distance from a straight-line between nodal districts (0,1) District 0-10 km from line ROUTE 1 (0,1) District 0-10 km from line ROUTE 2 C. IV estimates using distance from a straight-line between nodal districts     Table 6: Long-differenced estimations of the relative role of incumbents versus entrants in districts Log levels of activity in 2007/9 due to incumbents alive for at least 10 years compared to total initial district activity in 2000 Log levels of activity in 2007/9 due to firms less than 10 years old compared to total initial district activity in 2000 Share of activity in 2007/9 in district that is contained in incumbent plants at least 10 years in age Notes: See Table 2

Empirical Appendix
This appendix describes some key resources and data preparation steps regarding highway segments.
Data Source on GQ Details: As per the National Highway Authority of India (NHAI), the GQ highway venture was a collection of 128 projects implemented all over the country. By March 2011, 120 of these projects had been completed, while eight were either pending completion or terminated. We compiled information on each of these 120 projects from the annual reports of NHAI from 1998-99 to 2010-2011 as well as from the Ministry of Roads, Transport and Highways. These annual reports identi…ed the project name for the highway stretch, the length of the highway stretch, the national highway number, the start date for the project, and target and actual completion dates.
Mapping GQ Projects to Districts: In most cases, the name of the project indicated the start and end towns on a highway stretch. This information was used to identify districts lying within 10 km of the highway stretch.
The start and end points of the segment along the highway were located in the shape …le and that segment was then selected to query all of the districts located within 10 km of that segment in the GIS software. In some cases, the project name was not clear or the town name could not be located using the shape …le, Google maps, or open street maps. In such cases, we used information on the NHAI website for the highway project chainage and mapped the preceding or succeeding highway stretch. We then traced back kilometer by kilometer on the speci…ed highway number to get the names of the towns that would lie close to the highway stretch.
New Construction versus Upgrades: We obtained information on whether each GQ project focused on new highway construction or the renovation and upgrade of an existing highway using a supplier database called Process Register. Process Register is a comprehensive online reference database of suppliers of products and services used in the process, energy, and greater manufacturing industries. NHAI is a listed supplier on Process Register and most NHAI projects are listed in the database. Of the 70 districts lying near the GQ network, 37 districts experienced purely upgrade work. For 33 districts, some or all of the work was new construction.
Implementation Date: We grouped projects by whether they were completed by March 2003, March 2006 or later. We then matched districts to individual highway projects and their completion dates. Several districts touch two or more stretches of highway. For such districts, we allocated them to the earliest completion date as they had access to some connectivity before other districts did. However, we did not allocate a district into an earlier bin if the earlier GQ project was River Over Bridge (ROB), a bridge section, or a short bypass, as most of these constructions were small in terms of kilometers of length.
There were some highway projects, such as Vijayawada-Chilkaluripet, which came in several packages, and we could not distinguish among packages in terms of identifying districts because the start and end points were the same in all packages. Fortunately, in most cases, such projects were implemented in the same time period, so we could combine projects to get the set of districts lying within 10 km of the project in question.    Table 4: Alternative definitions of district-level labor and total factor productivity Appendix Notes: Dynamic estimations consider the location and productivity of organized-sector manufacturing activity in non-nodal Indian districts within 50 km of the GQ network for 1994-2009 from the Annual Survey of Industries. The interaction terms quantify the differential effect for GQ upgrades for non-nodal districts within 10 km of the GQ network compared to districts that are 10-50 km from the GQ network by year, with 1994 as the reference year. The GQ upgrades commenced in 2001. Estimations report standard errors clustered by district, include district and year fixed effects, have 1188 observations, and weight observations by log total district population in 2001. App.