Journal of Development Economics 145 (2020) 102442

Contents lists available at ScienceDirect

Journal of Development Economics

journal homepage: www.elsevier.com/locate/devec

On the road: Access to transportation infrastructure and economic growth

in China☆
Abhijit Banerjee a , Esther Duflo a , Nancy Qian b, ∗
a
MIT, J-PAL, BREAD, NBER, CEPR, USA
b
Northwestern University, BREAD, NBER, CEPR, USA

A R T I C L E I N F O A B S T R A C T

JEL classification: This paper estimates the effect of access to transportation networks on regional economic outcomes in China over
D2 a twenty-year period of rapid income growth. It addresses the problem of the endogenous placement of networks
O4 by exploiting the fact that these networks tend to connect historical cities. Our results show that proximity to
R4
transportation networks have a moderately sized positive causal effect on per capita GDP levels across sectors,
but no effect on per capita GDP growth. We provide a simple theoretical framework with empirically testable
Keywords: predictions to interpret our results. We argue that our results are consistent with factor mobility playing an
Infrastructure important role in determining the economic benefits of infrastructure development.
Growth
Inequality
Firms

“A key issue [on whether railroads benefit economic development], have dramatically better transportation infrastructure than poorer ones.
however, is whether such railroad influence was primarily exoge- However, policymakers considering the trade-offs of investing in infras-
nous or endogenous, whether railroads first set in motion the forces tructure must consider several related questions. First, they must con-
culminating in the economic development of the decade, or whether sider the question of causality: is infrastructure development a worth-
arising in response to profitable situations, they played a more pas- while object of policy, or is it better to rely on the natural forces of
sive role.” – Albert Fishlow, American Railroads and the Transforma- the market and/or competition between local jurisdictions to endoge-
tion of the Ante-bellum Economy, 1965 pp. 203 nously provide the necessary infrastructure in response to demand? For
example, Fogel (1962, 1964) famously argues that one of the most fre-
quently mentioned historical innovations in transportation infrastruc-
1. Introduction ture, the railroad, was less effective for economic development in the
United States than the pre-existing river networks and that this misdi-
Transportation infrastructure is often mentioned as a key to pro- rected investment was a result of government policies for promoting
moting growth and development. The argument relies on the simple railroads.
logic that one first needs to have access to markets and ideas before Second, policymakers are typically concerned about the distribu-
one can benefit from them. This belief is supported by the observation tional effects of infrastructure, which are by no means obvious. On the
that the historical construction of infrastructure, such as railroads, coin- one hand, for fixed factor endowments, the increased access to markets
cided with periods of rapid economic growth in Western Europe, Japan and ideas should benefit all regions. For example, in the historical con-
and the United States. Today, it is indisputable that richer countries text of the United States, it has been argued that transportation infras-
tructure gives rise to more cities, which then turned into “engines” of

☆
This paper updates and supersedes “The Railroad to Success: The Effect of Access to Transportation Infrastructure on Economic Growth in China” ( Banerjee et
al., 2004), which used the same basic empirical strategy, but substantially less data. We are grateful to Naomi Lamoreaux, Tom Rawski, Thomas Piketty and the
participants at the 2004 MacArthur Network for Inequality Conference in Beijing, the China Summer Institute, and the 2011 IGC Conference in London for very
helpful comments. We thank Zhichao Wei, Gongwen Xu and the large team they assembled to collect the data; Ricardo Dahis, Zhentao Jiang and Joris Mueller for
excellent research assistance; and Giovanni Zambotti and Ceren Baysan for invaluable assistance with GIS. We acknowledge financial support from the IGC.
∗
Corresponding author.,
E-mail addresses: [email protected] (A. Banerjee), edufl[email protected] (E. Duflo), [email protected] (N. Qian).

https://doi.org/10.1016/j.jdeveco.2020.102442
Received 31 May 2018; Received in revised form 5 January 2020; Accepted 22 January 2020
Available online 15 February 2020
0304-3878/© 2020 Elsevier B.V. All rights reserved.
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

growth for the country as a whole.1 On the other hand, transporta- exploiting these new opportunities, exporting to the rest of the world,
tion infrastructure increases the access of rural regions to cities, and using their access, although they could also export their resources to
the well-known agglomeration effects of cities may cause productive the terminal cities, which would have the opposite effect. We there-
capital and skilled labor to move from rural regions to cities over time, fore also study growth effects of being close to the line over the period
with the result that those who remain in rural areas receive very limited 1986–2003.
benefits from urbanization or even become impoverished. Along similar The results show that being close to the line had a positive level
lines, it has been argued that the expansion of motor road networks in effect. Per capita GDP was higher in places closer to the line. However,
the United States promoted large-scale suburbanization and left many the effect is not large. The elasticity of per capita GDP with respect
cities without a viable economic model (Glaeser and Gottlieb, 2009). to distance from historical transportation networks is approximately
This paper makes progress in understanding the impact of access −0.07. The small level effect is consistent with independent data from a
to transportation infrastructure by examining the causal effect of access higher-quality household survey for rural areas, the National Fixed Point
on economic performance in different regions in China during a twenty- Survey (NFS) collected by the Chinese Ministry of Agriculture, which
year period of rapid growth. We ask the straightforward question: do shows that distance has no significant effect on household income.
areas that are “quasi-randomly” assigned to have better access to trans- For the estimates of the effect of proximity on growth, we find a pre-
portation networks consequently have better economic outcomes in the cisely estimated zero effect. The estimated elasticity between distance
long run? Specifically, we attempt to empirically examine two closely to the line and annual per capita GDP growth is −0.002 and statistically
related questions. First, we ask whether access to better transporta- insignificant (the standard error is 0.003). Places close to the line grew
tion enriches the average region that is affected (because it draws in exactly as fast as places further away.
or generates more new economic activities) or impoverishes it (because Our finding that better access to transportation networks does not
it becomes easier for human and physical capital to exit). Second, we have a large impact on the (relative) economic performance of those
ask whether areas that have better access to transportation networks areas is consistent with the Fogelian view that transportation infrastruc-
benefit much more and serve as engines of growth when new economic ture by itself does not contribute much to growth, excepting perhaps
opportunities arise and growth becomes possible after 1979. where there was already a demand for it. Based on similar logic, China
For our discussion, it is important to keep three points in mind. scholars have criticized the tremendous amount of public investment in
First, our focus is on long term effects. We are not only interested in the domestic transportation infrastructure after 1990 (Huang, 2008).
impact on trade and prices that result from greater access, which tend However, there is an alternative and complementary interpretation
to be relatively quick, but also in the subsequent changes in the pat- under which the measured benefits of infrastructure are small even if
terns of localization of economic activity as people and factories relo- better transportation causes substantial gains to overall GDP. The basic
cate. Second, the emphasis on understanding the effect of infrastructure idea emphasizes the role of factor mobility. Under full labor and cap-
for the average location is crucial to our study since it is entirely pos- ital mobility, wages and incomes would be equalized in our treatment
sible that some of the largest cities benefit from infrastructure through and control areas even if there are large macro effects and we would
greater concentration of resources while most other places lose out. see no difference in their outcomes. Of course, the assumption of full
Finally, there cannot be one definitive answer to these questions, since factor mobility is clearly inconsistent with the institutional conditions
the answer will clearly depend on the starting point – i.e., the first road in China. In the paper, we present a simple model which shows that
to connect the agricultural hinterland to a port is very different from we can observe similar patterns with limited factor mobility. Specif-
the fifth such road. ically, if labor mobility is very limited, but capital is also relatively
We use county-level economic data from China to try to answer immobile compared to goods, and its mobility depends on the distance
these questions. In many ways, China offers an ideal setting for our to transportation infrastructure, more remote areas may actually retain
work. In the late 19th and early 20th century, the Chinese govern- more of their capital compared to better connected areas (where all
ment and a set of Western colonial powers built railroads connecting the resources move to the nearest metropolitan center). For the latter
the historical cities of China to each other and to the newly constructed reason, GDP per capita may not dramatically decline in remote areas.
“Treaty Ports”.2 We identify our average “treated” areas to be those that Moreover, this effect only tells us about the level of GDP. As far growth
were close to the straight line connecting this set of cities. Our analy- is concerned, since even the more remote locations retain a substantial
sis excludes the terminal cities, where there are obviously confounding part of their resources, they also retain the possibility of participating
effects. Our strategy is to first compare areas closer to the lines to areas in and benefitting from the exposure to global markets that raised the
further away and show that they have on average better infrastructure. growth rate everywhere in China. Therefore, the impact on the growth
We then compare various measures of economic outcomes further and rate, starting at a lower level, can be similar in remote and less remote
closer to the line and interpret any difference in economic outcomes as areas. We summarize this intuition in the body of the paper and provide
the overall effect of any transportation infrastructure – the original rail- a formal discussion in the Appendix.
roads and any other infrastructure later added – along these historical In assessing what general lessons one can learn from our results, one
transportation corridors. should consider whether our results are driven by conditions specific
This strategy has a number of advantages. First, it provides us with to the Chinese context. For example, one may be concerned that the
an exogenous source of variation in access to transportation networks. marginal effect of access to infrastructure is especially low in China
Second, this variation goes back to at least fifty years before our study due to the massive public investment in infrastructure during recent
begins in 1986, by which time the patterns of economic activity would years.3 We believe that this is highly unlikely to be driving our results
have had ample chance to relocate. We can therefore ask what the long because our data show large variation in access to infrastructure. The
run level effect of being close to the line (and hence to transporta- distance to the railroad for the counties that are the nearest the line
tion) was, say around 1986. Third, our study period, 1986–2006, coin- (defined as the nearest decile) are only one-third of the distance of those
cides with China’s opening up and subsequent growth acceleration. Our furthest (defined as the furthest decile) from the line. Similarly, the
treatment areas were plausibly in a good position to take the lead in nearest counties have more than twice the length of highway relative
to the furthest counties (despite the fact that the latter are almost eight
times as large).
1
For example, see the World Bank’s World Development Report 2009 on
Reshaping Economic Geography by Aoyama and Horner (2009) for a nuanced
3
statement of this view. For example, see Huang (2008) for a discussion on infrastructure investment
2
For example, see Pong (1973). in recent years in China.

2
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Another concern for external validity is that the lack of factor mobil- which finds evidence that increased market access from the historical
ity stems from the Chinese government’s attempts to control labor expansion of U.S. railway networks increases land values; and Sequeira
mobility and that the empirical findings are not easily generalizable et al. (2017), which finds that access to railroads increased immigra-
to the context of other developing countries.4 We acknowledge that the tion during America’s Age of Mass Migration (1850–1920), which had
Chinese government may be unique in implementing an explicit pol- long-run consequences for economic prosperity.8
icy for controlling migration for so long. However, it is important to Moreover, our paper provides a potential interpretation of the lack
note that the actual patterns of low levels of migration are not unique of infrastructure effects which is of some independent interest. The idea
to China. In particular, the main policy effort has focused on unskilled that the lack of factor mobility might limit the measured impact of better
low-wage rural workers (Meng, 2005), a group that has been found infrastructure is of considerable relevance to many developing countries
to be relatively immobile in other contexts such as in India (Munshi that are currently investing in improving their infrastructure.
and Rosenzweig, 2009). Moreover, our simple theory predicts similar Our findings also add to understanding the long-run effects of Euro-
effects regardless of which factor (e.g., labor, capital) is immobile and pean Imperialism on China’s economic development. In using Treaty
the immobility of capital has been documented in several developing Ports to construct the lines, we are closely related to Jia (2014), which
country contexts. For example, Chan et al. (2011) documents the immo- documents that Treaty Ports experienced better long-run economic
bility of capital in China during the period of our study.5 development than other regions in China. We add to this by show-
Since we began our study in 2004, a growing number of recent ing that through railroads, areas outside of the Treaty Ports were also
papers have developed compelling identification strategies to evalu- affected.9
ate the impact of transportation infrastructure. Most existing studies The rest of the paper is organized as follows: We start with a brief
examine the effect of transportation infrastructure from the point of review of the literature in Section 2. Section 3 presents the theoreti-
view of market integration and the focus is on price convergence and cal framework that we use to think about our results including a sim-
changes in the relative price of factors along the lines predicted by trade ple model of industrial location choice. Section 4 provides the back-
models. The results suggest that transportation infrastructure favors ground and the empirical strategy. Section 5 describes the data. Section
greater price convergence and that factor prices shift in the direction 6 presents the results. Section 7 offers concluding remarks.
predicted by trade theory (e.g., Michaels, 2008; Donaldson, Forthcom-
ing; Keller and Shiue, 2008).6 Recent studies also provide evidence 2. Growth, capital and mobility
that better transportation can have adverse local effects. For example,
Faber (2014) finds that China’s new highway system adversely affects This section briefly discusses factor mobility in China in relation to
productivity in newly connected regions and Baum-Snow (2007) and the simple model we present in the next section. We aim to make three
Baum-Snow et al. (2017) find that better transportation infrastructure points. First, central planning policies caused the endowment of human
shifts populations and economic activities away from city centers in the and physical capital to be higher in urban areas relative to rural areas in
United States and China. the pre-reform era (1949-76). However, to promote rural industrializa-
Our study differs from these studies in its focus on the longer-run tion, the pre-reform government also invested substantial amounts of
and more macro question: do areas that benefit from access to the capital in rural areas (Unger, 2002). Second, restrictions on migration
reduction in trade costs and perhaps other costs become wealthier as largely prohibited the mobility of unskilled labor during the post-reform
a consequence? This is by no means obvious even if there is clear evi- period of our study and limited financial development probably inhib-
dence that trade and other flows such as migration increased when ited capital mobility (West and Zhao, 2000). Finally, the post-reform
infrastructure became available. Our estimates provide a much more era was characterized by very high growth rates.10
reduced form effect, which presumably includes not just the possible Chinese central planners have always focused on economic growth
gains from more efficient trade but also the effects of greater factor and industrialization. In the early 1950s, this meant moving skilled
mobility, better access to education, health care and finance, and other, workers and machines into cities. During this period, the percentage
more diffuse effects coming from the diffusion of ideas, technologies, of government revenues used to fund industrial development increased
etc. In this sense, our study is more closely related to recent studies from 32% in 1952 to 57% in 1957 (Eckstein, 1977). Much emphasis was
that examine the effects on population growth, land values and city also put into improving human capital in cities. In addition to moving
size. For example, Atack et al. (2010) finds that access to railroads has skilled workers into cities, a special emphasis was put on secondary
a strong positive effect on urbanization but a small effect on population and higher education. All secondary and higher education institutions
growth in the United States.7 Other recent studies examining long-run in China are located in cities which naturally causes human capital to
effects using U.S. data include Donaldson and Hornbeck (forthcoming), be drawn into cities even if some of the students were born in rural
areas.
Rural areas also received investment, albeit less than the cities. An
4 For example, see West and Zhao (2000) for a review of studies on labor
enormous number of primary schools were established so that all rural
migration.
5
children could have access to a basic education. Literacy rates in China
For example, see Banerjee and Duflo (2005); Duflo (2004) for evidence of
reportedly improved from less than 20% in 1949 to 68% by 1982, even
limited capital mobility within Indonesia.
6
Michaels (2008) examines the effect of highway construction in the United though almost 80% of the population was still rural (Jowett, 1989).
States in the 1950s, using both a difference-in-difference (DD) approach and Rural areas also received investments in physical capital: villages were
an instrumental variables approach, where he exploits the variation in access collectivized and physical capital was owned and managed by collec-
caused by the fact that highways tended to be built in either a North-South tives. When China de-collectivized during the early 1980s, collective
direction or an East-West direction starting from big cities. Donaldson (Forth- assets were inherited by villages, and were often used to form Town
coming) studies the effects of railroad construction in 19th century India using and Village Enterprises (TVE).
a DD approach. Keller and Shiue (2008) uses a similar strategy to examine the
opening up of railways between regions within Germany. Also, Chandra and
8
Thompson (2000) use historical U.S. data to find that connections to highways Outside of the U.S. context, see Alder (2015), which uses a model-based
have heterogeneous effects across industries. approach to demonstrate the benefits of China’s current transportation system
7
While Atack, Bateman, Haines, and Margo (2010) primarily uses a DD relative to India’s system; and Storeygard (2016), which finds that higher trans-
approach, it also constructs an instrument for the distance to the railroad based portation costs reduce city size in sub-Saharan Africa.
on the straight line between the start and end points of a railway line. The 9
Note that the exclusion of termini cities means that Treaty Ports are
authors generously credit the straight-line instrument to an earlier version of excluded from our analysis.
10
our paper (Banerjee et al., 2004). See, for example, Hu et al. (1997) for an overview of Chinese growth.

3
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

For our study, it is important to note the following facts. First, a of physical capital.
significant proportion of industrial output in China during our study The premise of this model is that goods move more easily than
period came from TVEs. As a percentage of national industrial output, capital. Unfortunately, we cannot directly observe the relative mobil-
output from TVEs grew from 9% in 1978 to 36% in 1993.11 Second, ity of goods and capital. However, our model tells us that relatively low
TVE assets are jointly owned by all community residents, which were mobility of capital is likely to be associated with a situation where there
approximately 400 households in an average village and 3500 house- is higher inequality in better connected areas, as long as the direction
holds in an average township. Households owned equal shares in TVEs of capital movement is from less connected areas towards better con-
and it was illegal to sell or transfer their shares to non-community mem- nected ones. Using regional income inequality data computed from the
bers. Third, the law required that at least 60% of the profits be retained National Fixed Point Survey collected by China’s Ministry of Agriculture
in the village.12 The data show that over half of the profits were re- (1987–2005), we do find that inequality is higher in better connected
invested.13 These three facts together suggest that a significant amount areas.
of productive capital was in rural areas, and policy both prevented their
mobility to cities and promoted further capital accumulation in rural 4. Historical background and empirical strategy
areas.
Labor mobility was also restricted. If a worker moved without offi- 4.1. The birth of modern infrastructure
cial permission, she lost access to all public goods. For urban residents,
this meant losing access to schools, healthcare, and during the 1980s As explained above, the basic idea behind our empirical strategy
and early 90s, it also meant the loss of food rations and housing. For is to examine the correlation between the distance to the nearest
rural residents, this meant the loss of farmland. Government permission straight line connecting two historical cities and the outcomes of inter-
was easier to obtain for skilled workers such as college graduates who est. Throughout the paper, we assert that these lines capture major
could obtain jobs that assisted them in getting the permission to relo- transportation networks during the 1980s because they capture the first
cate or workers with skills that were needed in specific industries such modern infrastructure (e.g. railroads) built in China and much of the
as construction during the mid- and late- 1990s. But for the rest of the infrastructure development afterwards began by initially building along
population, permission was extremely difficult to obtain (e.g., Meng, these routes. Later in Section 6.4.1, we will provide evidence for our
2005; Meng and Kidd, 1997). Therefore, while the number of migrant assertion.
workers increased greatly during this period, most of them were tem- To draw the lines, we start with the set of important historical cities
porary migrants who maintained their original residences.14 in China circa 1860: Beijing, Chengdu, Guiyang, Kunming, Lanzhou,
Finally, it is important to point out the differences in growth rates Nanchang, Taiyuan and Xian. These were urban centers that were politi-
between cities and rural areas and how they changed over time in China cally and economically important and which never became Treaty Ports
during the post-Mao reform era, when income increased rapidly for the at any time (Murphey, 1970).15 To these we add the four Treaty Ports
country. During the first years of the period, 1978-84, the real income that were set up by the League of Eight Nations after they defeated the
of rural residents grew at 17.7% per year while it was only 7.9% for Qing government in the First Opium War in 1842 (Shanghai, Ningbo,
urban residents. This pattern was reversed in the mid-1980s and the Fuzhou and Guangzhou). These four cities were chosen for their strate-
urban advantage increased steadily for the remainder of the reform era. gic locations. The “unequal treaties” that were signed between China
On average, rural real income growth rates declined to only 4.1% while and the League of Eight Nations after the Opium Wars allowed the
urban real income growth was approximately 6.6% (Cai, 2010). Western countries to house their military in the Treaty Ports but not
beyond. Therefore, these ports were chosen to be easily accessible by
3. Conceptual framework European ships and also to be strategically advantageous for reaching
Chinese cities in case of an uprising or war. Later waves of Treaty Ports
In the Appendix, we present a simple model where labor is immo- were chosen more for economic reasons, and therefore are more likely
bile and capital is less mobile than goods, though even goods are costly to be correlated to factors that can affect our outcomes of interest.
to move. As a result, even remote areas continue to hold onto a part The four Treaty Ports in our sample are all along the coast or a major
of their capital and produce exportables and benefit from globalization. navigable river. Shanghai and Ningbo are on the northern and southern
Moreover, if the mobility of capital is more limited out of relatively mouth of the Yangtze River, Fuzhou was on the southern coast of the
remote areas, the effect on GDP per capita of being well connected will Yellow Sea, and Guangzhou was on the Xi River, near its mouth on the
be the result of two contending forces. On the one hand, distance is South China Sea. All of these ports were easily accessible by the naval
costly and makes exports less profitable. On the other hand, if remote gunships of the Western countries and therefore allowed them to both
areas retain more capital per head than better connected areas, this impose their military presence as well as control international trade
would boost the productivity of labor in those areas. As a result, the with China.16 With the exception of Guangzhou, these locations were
variation in per capita GDP between near and far places may be rel- villages and not prominent historical urban centers prior to becoming
atively small and both be involved in the production of exports. As a Treaty Ports (such as Nanchang or Xian). Therefore, the lines that we
result, the boost to TFP and growth resulting from the opening of the draw between these Treaty Ports and the historical Chinese cities have
economy for global trade may be the same in proportional terms in all no reason to go through regions of particular importance to the Chinese.
locations. Therefore, even though better transportation helps China as
a whole to gain more from trade, GDP level differences between well
and poorly connected areas can be small and there may be no differ-
15
ences in growth rates between the two areas. One could make the same See Appendix Fig. A1 for a map of all Treaty Ports and historically important
argument if the relatively immobile factor was human capital instead cities taken from Murphey (1970) page 35.
16
The Treaty Ports were established in Article 2 of The Treaty of Nanjing,
which was signed between the British and the Qing government. Article 2
requested the four cities we mention and Xiamen to be established as Treaty
11 Ports. But in practice, Xiamen did not receive significant investment from
See the Statistical Material of Township and Enterprises, 1992.
12 the West and only became a Treaty Port during the second wave of Treaty
See Articles 18 and 32 in The Regulation on Township and Village Collective
Enterprises of the People’s Republic of China (1990). Port Relinquishment by the Qing in 1865. Therefore, in our line construction,
13 we omit Xiamen. The other Treaty Ports of the second wave were Tianjin,
See Statistical Survey of China, 1992: pp. 67.
14
There have been numerous studies on migration in China. Zhao (1999) pro- Niuzhang, Yantai, Zhenjiang, Hankou, Shantou, Taibei and Tainan (e.g., Pong,
vides a survey of recent evidence. 1973; Spence, 1990).

4
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Moreover, it is important to point out that the Chinese were signifi- they do not match well are North-Western China (Xinjiang, Tibet, parts
cantly behind the Europeans in terms of naval technology in 1842, and of Inner Mongolia), where construction occurred under the Communist
did not possess a fleet of similar ocean-going naval gunships for which government after the 1970s, partly as an attempt to politically inte-
the Treaty Ports were chosen. More generally, China had conducted a grate these areas into China; and North Eastern China (Manchuria),
very limited amount of international trade since the 16th Century dur- where most of the construction was done by a de facto colonial Japanese
ing the Ming and Qing Dynasties. It did not have an outgoing navy for government during the 1920–30s (Fig. 1 shows that in Manchuria,
several centuries leading up to the Opium Wars.17 Therefore, places most counties have a railroad). For this reason, our estimating sam-
such as Shanghai, Ningbo and Fuzhou, while not entirely uninhabited ple will exclude Xinjiang, Tibet, Inner Mongolia and the provinces in
prior to 1842, were rural areas with small stations for domestic naval Manchuria.
patrol boats. Their insignificance before 1842 is shown by the fact that Our main source of plausibly exogenous variation for access to
none of the four cities were connected to the Grand Canal, which was infrastructure is the nearest distance from the center of each county to
a north-south canal built to connect Beijing to the important Southern this straight line. Both the centroids and the nearest distance are com-
cities.18 It follows that when we draw lines to connect the Treaty Ports puted by ArcGIS using the Asia Conical Projection. We use geographic
and historical Chinese cities, we are unlikely to be systematically cap- distance rather than travel distance measured as kilometers. This line is
turing important routes from before 1842. Instead, the lines will capture also our proxy for transportation infrastructure. We deliberately make
modern transportation networks built afterwards. no use of information about changes in infrastructure.
The first and perhaps most important transportation infrastructure To check that the lines do indeed proxy for transportation infras-
are railroads. They were mostly built during the early 20th Century by tructure, we estimate the correlation between distance to the line and
the Qing government and Western countries. The latter provided much various measures of infrastructure using the following equation:
of the financing and had substantial influence over the placement of
Icpt = 𝛿 lnLcp + 𝜌p + 𝛾t + 𝜀cpt . (1)
the railroads. They were largely built to promote Western economic
and military interests in China and connected Treaty Ports to historical Transportation infrastructure in county c in province p and year t,
cities, and also connected historical cities to Colonial cities outside of Icpt , is a function of: the natural logarithm of the distance to the nearest
China. For example, the British planned and financed railways to con- line connecting Treaty Ports and historical cities illustrated in Fig. 1,
nect the Yangtze River valley as well as a north-south railway to con- Lcp ; province fixed effects, 𝜌p ; and year fixed effects 𝛾 t . Note that the
nect Wuhan to Guangzhou against the protest of the Qing government, fact that the line is likely to be connected with many different types of
who feared that this would facilitate fast British troop deployment from transportation infrastructure means that Lcp is not an excludable instru-
Shanghai and Ningbo to important Chinese cities. The French planned ment for any given infrastructure.
and financed a railway to connect Kunming to Hanoi, an important city Our main estimating equation is the following:
in French Indochina. The Russians planned a railway that was almost
a straight line from Beijing to Vladivostok through Liaoning, Jilin and ycpt = 𝛽 lnLcp + ΓZcp + 𝜌p + 𝛾t + 𝜀cpt . (2)
Heilongjiang provinces (Spence, 1990, pp. 249-56). The outcome for county c, province p and year t, ycpt , is a function
of: the natural logarithm of the shortest distance to the line for county
4.2. Straight lines c in province p, lnLcp ; a vector of county-specific controls, Zcp ; province
fixed effects, 𝜌p ; and, year fixed effects, 𝛾 t . The standard errors are
We construct our independent variable using a simple algorithm. We clustered at the county level. If proximity to the line is beneficial, then
draw a straight line from each historically important city to the nearest 𝛽̂ < 0.
Treaty Port and/or to the nearest other historically important city. If Interpreting 𝛽 as the causal effect of proximity to the line assumes
there are two cities (or ports) where the difference in distances is less that the only difference between places near the line and places further
than 100 km, we draw a line to both. The line is continued past the away is the distance to the line. This obviously relies on the termini
city until it hits a natural barrier (e.g. Tibetan Plateau, coast line), or a cities not being chosen so that the straight line between them would
border to another country. The lines are shown in Fig. 1.19 run through economically important regions. This is the reason why
As expected, the lines drawn this way coincide well with railroads we focus on the ancient cities of China and the Treaty Ports – i.e., the
constructed during the early 20th century.20 The three places where historical cities are both sufficiently far from each other and clearly
more important than any place between them in the historical era that
it is easier to be comfortable with the identification assumption in this
17
See (Spence, 1990, Ch. 2) for a detailed discussion of China during the 19th context. Similarly, the Treaty Ports were chosen for their suitability for
Century. European gunships rather than what laid between them and the histori-
18
A large flood in 1855 permanently changed the course of the Yellow River,
cal cities. Note that we restrict our attention to the first four Treaty Ports
causing the canal to significantly decline in importance.
19 to avoid the potentially endogenous influences of later Treaty Ports,
The goal of the lines is to provide a measure of proximity to infrastruc-
ture that is exogenous to local potential for economic growth (conditional on
which may have been chosen for economic reasons (e.g., proximity to
the baseline controls). For our purposes, the ideal identification would be pro- economically viable or prosperous regions).
vided by randomly assigning infrastructure in 1840 and estimating the effect of There are two caveats. First, being closer to the line will, by con-
proximity to that infrastructure on outcomes in the 20th century. In practice, struction, mean that a county is also closer to the terminal cities. There-
since infrastructure was not randomly constructed, we exploit variation in the fore, our baseline specification will control for distance to the termi-
proximity to our constructed lines, controlling for things such as distance to nal cities. Second, the line from some historically important cities to
the terminus city. The more randomly assigned is the line, the better for causal a Treaty Port might follow a river, an important traditional means for
inference. Stopping the lines in the cities where the historical railways end in transportation as well as an important input for agriculture (e.g., river
practice would introduce endogeneity relative to extending the lines to the bor- beds provide fertile soils). In this case, distance from our line will also
der. The concern is that the historical railways ended at point X because the
capture the distance from the river, which presumably captures many
regions beyond it had less potential to grow. Thus, we precisely want to extend
other effects. To address this, our baseline specifications always control
the lines to the border to avoid endogeneity.
20
While the railroads suffered much damage during World War II, after the
for distance to the nearest navigable river.
war, the Guomingtang (KMT) and then the Communist (post −1949) govern- The baseline estimation also controls for other potentially influential
ments undertook extensive repairs and construction focused on upgrading the factors, which we will discuss and motivate later in the paper.
physical structure. A comparison of maps from the 1930s to maps from the Note that it is not clear that we can expand the set of cities being
1950s indicate that they mostly did not alter the course of the railroads. connected (and therefore use more of the data) without running into

5
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. 1. Lines and transportation infrastructure.

potential problems. One issue is that of endogeneity raised earlier. and historically important cities) using the algorithm described earlier.
Another equally important issue comes from the very nature of the con- We compute the nearest distance from each centroid to the straight
struction of lines. We compare places that are close to a line with those line, railroads, navigable rivers, the coastline, the country border and
that are further away. The implicit assumption is that moving further segment termini.21 Fig. 1 displays a map of county boundaries, our con-
away from one line does not bring us closer to a different line, a problem structed lines, railways and major navigable rivers.
that occurs when there are too many lines. We ensure this by having The first outcome measure we examine is county-level per capita
relatively few lines and using a sample of counties that are not too dis- GDP. These data are from the Provincial Statistical Yearbooks from China
tant from any line. The maximum distance of any county in our sample from 1986 to 2003 stored in the Archives National Library in Beijing,
from the nearest line will be 366 km. Fig. 1 shows that there are only China. We manually collected and digitized data from all published
ten lines. We will return to discuss this further in Section 6.4.1. yearbooks that reported county-level statistics on GDP. These data are
interesting because they measure production whereas previous studies
have mainly focused on prices. However, there are several problems
5. Data
with these data. First, GDP may have been measured using different
techniques across provinces and over time. To the extent that these
This paper uses data from multiple sources. All raw maps are
changes are documented or obvious (e.g., changes in the units of mea-
obtained in digital format from the Michigan China Data Center. Geo-
graphic measures are constructed using ArcGIS software, assuming a
Conical Projection. We define centroids of cities and counties. The lines
21
are constructed to connect the centroids of segment cities (Treaty Ports The county boundaries are based on a 1990 map of China.

6
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. 2. Counties with GDP data from the Provincial Statistical Yearbooks.

surement), we have corrected for them. But this is clearly still imperfect. are consistent with our theory. The first of these are firm-level data
Second, not all counties report GDP and those that are reported are not from the Census of Industrial Plants in 1993 and the Census of Manufac-
a random sample of Chinese counties. Third, many counties do not con- turing Firms during 2004–2006. We are able to geocode these data to
sistently report over time, which means that we have an unbalanced the county level.23 The first survey includes all industrial plants. The
panel where attrition is non-random. There is little documentation on second survey samples all state-owned manufacturing firms and all pri-
the logic behind the decision of which counties report GDP and we can vately owned manufacturing firms with revenues of five million RMBs
do little to correct for it. Our final sample is an unbalanced panel of 295 or more. We will examine two outcomes, the number of firms and their
counties within sixteen provinces.22 In addition to the data on GDP, we profits. The data are aggregated to the county and year level and form
collected data on county population so that we can calculate per capita an unbalanced panel of counties. Fig. 3 maps the counties for which we
GDP. Fig. 2 maps the counties for which we have GDP data. have firm data.
To address these measurement difficulties, we supplement the anal- The second additional data are village-level data for rural house-
ysis with two additional data sets of higher quality. While they cannot hold incomes from the National Fixed Point Survey (NFS) for the years
allow us to directly correct for the county-level GDP data, they do allow 1987–1991, 1993, 1995–2005. There were no surveys in 1992 and 1994
us to check that the estimated effects in these two alternative data sets

23
These data are in principle available for other years. However, we only use
the four years for which we could geographically identify the location of the
22
Beijing, Hebei, Jiangsu, Anhui, Fujian, Jiangxi, Shandong, Henan, Hubei, firm at the county level. This data has been used by many studies, the most
Hunan, Guangdong, Guangxi, Guizhou, Gansu, Qinghai, and Ningxia. well-known of which is probably Hsieh and Klenow (2009).

7
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. 2. Counties with GDP data from the Provincial Statistical Yearbooks.

for administrative reasons. The NFS is a longitudinal survey of about Our income variable measures total net income – i.e., the sum of house-
320 villages and 24,000 households distributed across all continental hold income (e.g., home production, agricultural production, wages)
Chinese provinces conducted by the research arm of the Ministry of minus the sum of production costs, excluding labor costs for home pro-
Agriculture (RCRE). The villages were chosen in 1987 to be nation- duction and agriculture. The data are aggregated to the county and
ally representative. There is very little attrition. To maintain its rep- year level. The RCRE provided us with income for each decile of the
resentativeness, villages and households are added over time. There- village income distribution and the Gini coefficient for the within vil-
fore, the panel of villages is not perfectly balanced. For this study, lage income distribution each year and did not provide us with average
we use household-level data on income. Each village contains on aver- income across all households. Therefore, in the analysis, we will focus
age 400 households and approximately one-third of them are surveyed on income of the 10th, 50th, 90th percentiles and the Gini coefficient.
by the NFS. The large number of households surveyed in each village Fig. 4 maps the counties for which we have NFS data. Note that the
means that we can examine the within village income distribution.24 exact location of these villages are confidential. Therefore, our distance
variables measure the distance from the centroid of the county that con-
tains the village to the object of interest. This introduces measurement
error to the right-hand-side of our estimates for household income that
24
Villages and households are surveyed every year. The survey uses a strati- is most likely classical in nature.
fied sampling approach. For each province, it first randomly selects a number For all samples, we exclude the autonomous regions of Tibet, Xin-
of counties, and then randomly selects a number of villages within each county. jiang and Inner Mongolia both because these provinces are predomi-
Households are then randomly selected from each village. See Martinez-Bravo
nantly non-Han ethnic minorities, and therefore faced different policies,
et al. (2017) for a description of these data.

8
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. 3. Counties with Firm data from the Censuses for Manufacturing Firms.

and because the railroads constructed in these regions were the results capture the distance to major transportation networks and to avoid the
of very different imperatives. For the latter reason, we also excluded the problem of having too many lines that we discussed earlier. During
three Manchurian provinces of Heilongjiang, Liaoning and Jilin. The our study period, there are very few highways with median dividers in
large cities that are on the segment termini are also excluded to avoid China. On average, a county has only approximately 6 km of divided
the results being driven by the end-points, which are on the line and highways. Most motor traffic occurred on paved motor roads without
were chosen because they were important to begin with. It is important dividers. An average county has approximately 84 km of such undivided
to note that other cities on the line (that are not the termini of line seg- paved roads.25 The average county is far from a navigable river, the
ments) are included in our sample so that our estimates will capture any coastline and the country border.
effects that transportation infrastructure may have on the formation or Note that the data show significant variation in access to transporta-
growth of cities. tion infrastructure. This alleviates any concerns that our study can-
not detect significant marginal effects of access because high levels of
infrastructure investment by the Chinese government causes there to be
5.1. Descriptive statistics
too little variation in access.
The average population of a county is approximately 201,347. Per
Table 1 describes the data. Panel A describes the sample with GDP
capita GDP is 6834 RMB. The nominal GDP reported in the statistical
data. On average, these counties are approximately 71 km from the
line and 39 km from railroads. The fact that the average distance to
railroads is less than the average distance to the line reflects the fact 25
The GIS data on roads are produced by Harvard CHGIS and presumed to
that we constructed many fewer lines than there are railroads to only reflect 1990 conditions.

9
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. 4. Counties with Income data from the National Fixed Point Survey.

yearbooks are adjusted by the national CPI. GDP from primary, sec- Since there are approximately three people per household in these
ondary and tertiary sectors are roughly similar in size. Average per data, the household income here implies a slightly lower income than
capita GDP growth is 8% in this sample, which is similar to the national the per capita GDP from the sample in panel A. This is not surprising
average during the study period. Most of the income growth comes from since the earlier sample includes urban and rural areas, while the house-
the secondary and tertiary sectors. Note that the number of observations hold income data in panel B only includes rural households, which are
differ across the GDP variables because not every county is engaged in on average poorer than urban households. Similarly, income growth is
economic activity in every sector. slower in rural areas. Recall that there are no data for 1992 and 94.
Panel B displays the descriptive statistics for the sample of house- Therefore, we interpolate the annual growth rates between 1991 and
hold income data. For the sake of brevity, we focus our discussion on 93, and 1993 and 95 as the growth rates for each two-year interval
the economic variables. The average within-village Gini coefficient is divided by two.
0.28. The net household income for the median household is on aver- Panel C describes the firm data. Again, for the sake of brevity, we
age 5460 RMB (constant), which is almost twice as much as the income focus on the economic variables. On average, there are 82 manufactur-
of the 10th percentile household and approximately half of the income ing firms in a county. We can divide these firms into three ownership
of the 90th percentile. Inequality is growing over time. The Gini coeffi- types: firms owned by the state, firms of mixed ownership, and firms
cient increases by 0.001 per year on average. This is driven by a higher owned by private individuals. State-owned firms are directly controlled
income growth rate for richer households, although the level of income by the state. Mixed-ownership firms are typically privatized state firms
increases across all parts of the income distribution. for which the state owns most of the equity. Individually owned firms

10
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 1
Descriptive statistics.
Variable Obs Mean Std.Dev.
A. Sample 1: County Level GDP (1986–2003)
Distance to Historical Line (km) 2744 71.31 64.16
Distance to Railroad (km) 2744 38.81 39.20
Length of Highway (km) 2744 6.05 12.77
Length of Paved Roads (km) 2744 83.54 47.16
Distance to River (km) 2744 183.23 137.38
Distance to Coastline (km) 2744 424.98 372.41
Distance to Country Border (km) 2744 865.14 205.50
Distance to Segment City (km) 2744 144.17 81.27
County Area (Sqkm) 2744 1918 1134
County Population (Individuals) 2744 201,347 211,227
Per Capita GDP (Constant RMB) 2744 6834 8076
Primary 2266 1968 2424
Secondary 2266 2192 3579
Tertiary 2199 1672 2295
Per Capita GDP Growth 1879 0.08 0.14
Primary 1644 0.04 0.17
Secondary 1644 0.10 0.19
Tertiary 1551 0.11 0.17

B. Sample 2: Household Income (1987-91, 93, 95–2005)

Distance to Historical Line (km) 1897 64.79 67.66
Distance to Railroad (km) 1897 37.82 42.36
Length of Highway (km) 1897 7.69 14.79
Length of Paved Roads (km) 1897 89.10 50.94
Distance to River (km) 1897 129.93 104.50
Distance to Coastline (km) 1897 475.19 318.46
Distance to Country Border (km) 1897 739.17 299.00
Distance to Segment City (km) 1897 133.01 78.90
County Area (Sqkm) 1897 2058 1261
Gini 1897 0.28 0.07
10th Percentile HH Income (Constant RMB) 1897 2729 1322
50th Percentile HH Income (Constant RMB) 1897 5460 2837
90th Percentile HH Income (Constant RMB) 1897 11,681 11,407
Gini Growth 1782 0.001 0.049
10th Percentile HH Income Growth 1782 0.018 0.221
50th Percentile HH Income Growth 1782 0.028 0.162
90th Percentile HH Income Growth 1782 0.031 0.199

C. Sample 3: Firms (1993, 2004-06)

Distance to Historical Line (km) 3663 81.63 87.83
Distance to Railroad (km) 3663 49.70 65.26
Length of Highway (km) 3663 5.72 13.80
Length of Paved Roads (km) 3663 92.57 78.33
Distance to River (km) 3663 148.43 111.98
Distance to Coastline (km) 3663 544.66 405.04
Distance to Country Border (km) 3663 705.24 309.89
Distance to Segment City (km) 3663 156.13 102.49
County Area (Sqkm) 3663 2405 3716
Number of Firms 3663 81.68 145.83
Public Ownership 3663 39.75 79.39
Mixed Ownership 3663 7.24 16.22
Individual Ownership 3663 24.76 80.28
Aggregate (Sum) Profits (Constant 10,000 RMB) 3321 227,411 1,176,004
Average Profits 3321 2889 28,211
Public Ownership 3233 1997 18,066
Mixed Ownership 1909 12,753 88,431
Individual Ownership 1785 1629 4441

Notes: Variables are observed at the county and year level. Sample 1 in panel A uses data
from Provincial Statistical Yearbooks. Sample 2 in panel B uses data from the National
Fixed Point Survey. Sample 3 in panel C uses data from the Censuses of Manufacturing
Firms. The geographic data in all samples are computed by the authors using ArcGIS.

are truly private enterprises that have little connection to the state. The five million in revenues).
data show that most firms are owned by the state and individuals. There Table 2 shows the outcome variables of interest for different dis-
are only a few firms that are owned by a mix of state and private parties. tances to the line. These data show that most of the economic measures
Next, we describe the data on firm profits. These only report profits on of interest decline with distance from the line. Most importantly, we do
counties with at least one firm. Therefore, the number of observations not observe systematic upticks in these measures as we approach the
will differ across variables because not every county has a manufac- furthest deciles, which is reassuring for the concern that distance from
turing firm of a particular type. The high level of reported profits is our line bring us closer towards another transportation network.
consistent with the fact that these data sample large firms (more than

11
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

5.2. Correlation between pre-treatment era measures and the distance to

the line

Notes: Columns (1)–(4) use data from the Provincial Statistical Yearbooks. Columns (5)–(8) use data from the National Fixed Point Survey. Columns (9)–(12) use data from the Census of Manufacturing
Aggregate Firm Profits
Before the main analysis, we can investigate the validity of our iden-

Std. Dev.

85,967
16,706
tification strategy by examining the relationship between the distance

5202
7681
6235
6170
5613
4549
6407
5943
(12)
to the historical line and three variables for which we have data in the
pre-treatment period that are likely to be correlated with the potential
for economic development. We describe them below.26
Mean

2451
2700
1169
2571
1555
2014
2639
1833
9150
2800
(11)

The first is a measure of population from 1850.27 Table 3 column

(1) shows that there is no relationship between log population in 1850
and log distance to the historical lines, conditional on the same baseline
Std. Dev.

184.24
171.65
214.98
149.72
135.23

166.48
101.47
Number of Firms

controls that we use for the main analysis. The only difference to the
99.54

76.66
26.42
(10)

baseline specification is that we exclude year dummies in this estimate

since the data are one cross section.28 We recognize that 1850 is after
the establishment of the first Treaty Ports. However, it has the advan-
111.85

115.95
97.40

96.31
92.99
73.43
93.32
75.57
44.37
16.07
Mean

tage of being prior to the Taiping Rebellion (1851-64), which caused

(9)

tremendous population losses and displacement.29

The second measure is an official political-economic rating of impor-
Std. Dev.

tance given by the Qing Dynasty, Chong Fan Pi Nan. These four indica-
0.08
0.06
0.08
0.05
0.06
0.06
0.06
0.07
0.07
0.07

tors were used to determine policies like military presence, conscrip-

(8)
Gini

tion, tax rates, tax enforcement, etc. The rating system was established
in 1731 (Zhang, 2017). Western historians, such as, McMahon (2014),
Mean

0.30
0.28
0.27
0.29
0.28
0.26
0.26
0.26
0.27
0.28

have translated Chong Fan Pi Nan as “frequented, troublesome, diffi-

(7)

cult, and fatiguing” administrative posts for Qing bureaucrats. Bai and
Jia (2016) uses these measures at the higher prefecture level, and inter-
Median Village Household Income

prets the designations in the following way: “chong (important in trans-

portation/communication), fan (important in business), pi (difficult to
gather taxes) and nan (high in crimes)”.30
Chong, Fan, Pi and Nan provide four indicators that are not mutu-
.Std. Dev.

ally exclusive, i.e., a county can be rated as any one or all four – Chong,
4243
1995
2428
3734
1505
1819
1726
3561
2429
2587

Fan, Pi and Nan. While crude, these measures are interesting because
(6)

they are potentially correlated to factors that could affect later eco-
nomic development (e.g., local cultural norms, administrative capacity,
Mean

6732
5472
5010
6736
4995
4475
4349
5756
5383
5774

connectedness to the central government, local political stability). How-

(5)

ever, since it is hard to predict whether the relationship between devel-

opment and any one of these factors would be positive or negative, and
Annual Per Capita GDP Growth

because these four variables are highly correlated, we will examine the
principal component in our analysis.
Table 3 column (2) shows that the first principal component is
Std. Dev.

uncorrelated with the distance to the historical line, conditional on the

0.16
0.14
0.12
0.13
0.14
0.14
0.16
0.12
0.13
0.19

baseline controls.
(4)

Finally, we obtain data on whether a county had a Buddhist temple

in 1820.31 The presence of a Buddhist temple was likely to be correlated
Mean

0.07
0.08
0.10
0.07
0.08
0.06
0.09
0.08
0.07
0.05

to population density and the location being politically and economi-

(3)

cally important. Bai and Jia (2016) also argues that it is associated with
Firms. Distance is calculated by the authors using ArcGIS.

social capital.
Distance from the historical line and economic indicators.

.Std. Dev.

978 counties have at least one Buddhist temple. Amongst these

Per Capita GDP

10,613
7742
9968
8365
6823
8334
6789
6771
6775
7505

counties, the average is two temples. The maximum number of temples

(2)

is 45 in a county. We merge them with our infrastructure data. Referee

Table 3 column (3) shows that the presence of any temple is uncorre-
Mean

7035
6684
7119
7079
6241
7507
6958
5972
5734
7994

lated with distance to the historical lines. In column (4) we examine

(1)

the number of temples with a Poisson regression. Similarly, we find no

relationship.
Distance to Historical Line by Decile

These results are consistent with our identification assumption,

which implies that there should be no correlation (conditional on the
baseline controls).

26
See the Appendix for more discussion of the data.
27
Table 2

Appendix Fig. A.2 displays these data.

28
We present robust standard errors (instead of clustering them at the county
0
1
2
3
4
5
6
7
8
9

level).
29
See Section 6.4 for more discussion.
30
Appendix Figs. A.3a - A.3d display these data.
31
Appendix Fig. A.4 displays these data.

12
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 3
Correlates of baseline characteristics and distance to the historical lines.
Dependent Variable
(1) (2) (3) (4)
Ln Pop 1850 Qing Rating PCA Buddhist Temple Dummy # of Buddhist Temples
Ln Dist Historical Line −0.0226 0.0166 −0.00929 −0.0184
(0.0273) (0.0348) (0.0102) (0.0368)
Observations 588 1117 2220 2220
R-squared 0.434 0.194 0.236

Notes: All regressions are cross-sectional estimates. They control for the full set of baseline controls in Table 5
column (6), except for the year fixed effects. Columns (1)–(3) presents OLS estimates. Column (4) presents Poisson
estimates. Robust standard errors are presented. Data for 1850 population are reported by Ge (2000). Qing Dynasty
ratings are reported by Zhao (1976). The number of Buddhist temples are reported by Harvard CHGIS.

6. Results tistically significant at the 5% level. It shows that the elasticity between
the distance to the line and per capita GDP is-0.0681. Note that because
6.1. Lines, railroads and transportation networks the data indicate that the relationship between the distance from his-
torical lines and per capita GDP is log-linear, our main specification in
Table 4 shows the estimates of the correlation between the distance column (6) does not control for the quadratic of the distance from the
to the nearest transportation infrastructure and the distance to our con- line.
structed lines based on equation (1). Distance is measured in terms of One way to assess the magnitude of our results is to benchmark
kilometers. Panel A shows that distance from the historical lines is pos- our estimates of the effect of distance on GDP across space to the total
itively correlated with distances from railroads, the coastline and seg- increase of GDP over time in our sample. In our sample, the 75th-
ment cities; negatively correlated with the distances to the country bor- percentile county in terms of distance is 3.8 times further away from
der and the length of highways within a county; and uncorrelated with the line than the 25th-percentile county. Our estimates imply that dis-
whether a county is on the coastline or near a navigable river, and the tance will cause the 75th-percentile county to have almost 26 percent
length of paved roads. The correlations shown in Panels B and C will be (−0.0681 × 3.8 = −0.258) lower per capita GDP. During the eigh-
discussed later in this section. teen years covered by our data, per capita GDP growth in our sample
grew from approximately 2744 to 9916 RMB (e.g., the annual growth
6.2. The effect of distance from the line on GDP rate was approximately 7.5 percent), which is approximately a 242%
increase. Therefore, a comparison of the effect of distance across space
To illustrate the effects of our baseline controls, we first estimate the to the increase in GDP over time suggests that the spatial difference
effects of distance to the line on the log of GDP per capita. In Table 5, we attributable to distance from the line is relatively moderate in size.33
begin with a specification that only controls for province and year fixed For the remaining results, we will show only the baseline specifica-
effects (see column (1)). In columns (2)–(6), we gradually introduce the tion for the sake of brevity. All regressions will control for the full set of
baseline controls. The distances to the segment city control for the effect baseline controls shown in column (6) of Table 5: the distances to seg-
of proximity to a large urban terminus. The distances to the nearest ment cities, the nearest navigable river, the coastline and the country
navigable river and coastline control for access to traditional methods of border; the total area of the county; the squared terms of each of the
transportation that existed before the lines of interest were constructed. aforementioned variables; and province and year fixed effects. As with
Controlling for the distance to the country border addresses the possible the results in Table 5, the estimated coefficients for the distance from
influences of a “border” effect.32 Finally, the control for the distance the line are very similar with different combinations of controls.34
to the coastline also addresses the fact that during the period of our In Table 6, we examine per capita GDP and annual growth in per
study, economic conditions diverged greatly between the coastal areas capita GDP by sector. We estimate the reduced form effect of the dis-
and the interior areas. Without this control, one could be concerned tance to the line from equation (1). The estimates for the full sample
that a positive correlation between economic outcomes and distance are shown in Panel A. Columns (1)–(4) show that distance to the line
to our lines is an outcome of faster growth in the coastal areas, which is negatively correlated with GDP levels across sectors. The estimates
may also be coincidentally closer to our lines on average. In addition to are statistically significant at the 5% level for per capita GDP in the
controlling for the log of the linear measure of these distance measures, secondary and tertiary sectors.
we also control for the quadratic terms to capture the idea that the costs Columns (5)–(8) show the estimates of the effect of distance from
of distance from transportation may be diminishing over distance (e.g., the line on per capita GDP growth. We calculate per capita GDP growth
there maybe increasing returns to profit). as the difference between log per capita GDP growth next year and this
The estimates show that the coefficient for the log distance to the year for each county, ln(pcgdpc,t +1 ) − ln(pcgdpc,t ). To control for the
historical line and its standard error is reasonably stable across specifi- possibility that poorer regions may experience different rates of growth
cations. The full baseline specification is shown in column (6). It is sta- relative to rich regions for reasons that are independent of access to
infrastructure (i.e. income may be mean-reverting), we control for two
lags of the level measures of the dependent variable: ln(pcgdpc,t −1 ) and

33
For another benchmark, consider the fact that the 25th and 75th percentile
counties grew at 3% and 13% per year, resulting in around 70% and 800%
growth in per capita income levels over eighteen years.
32 34
For example, see Feenstra (2002) and the studies referenced there within. These results are available upon request.

13
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

ln(pcgdpc,t −2 ).35 The estimates are statistically insignificant for all sec-

Notes: All regressions control for the logarithm of the area of the county, year and province fixed effects. Standard errors are clustered at the county level. These estimates use an unbalanced county-year
tors. They are also very small in magnitude, especially when we con-

Ln Distance to Country Border

sider that the mean growth rate in our sample is 4–8% percent per
year, depending on the sector. Therefore, we conclude that we find a
precisely estimated zero effect of the distance from the line on GDP per
capita growth.36
−0.0267(0.0151) Another way to assess the magnitude of the estimates is to make the

27440.043
extreme assumption that being near the line benefits production only
through a region’s access to railroads. Under this assumption, we can
(7)

estimate the upper-bound of the effect of the distance from railroads

by dividing our main estimates by the estimated correlation between
Ln Distance to Coastline

distance to the line and distance to the railroad (e.g., equation (2) with
the dependent variable being the log of distance to railroads). This esti-

−0.000817(0.0158)
mate is 0.133 with a standard error is 0.0628 (not shown in tables),
−0.00434(0.0151)

−0.0281(0.0165)
0.180(0.0658)

which means that conditional on all of the baseline controls, doubling

27440.099

a county’s distance from the line increases the distance to the near-
26050.032

26050.045
est railroad by approximately thirteen percent. Dividing the estimate in
(6)

Table 6 panel A column (1) by 0.133, we calculate that the maximum

elasticity of per capita GDP with respect to the distance to the railroad
Ln Distance to Navagable River

is 0.5 for all GDP. Dividing the estimate in column (5) by 0.133, we cal-
0.0860(0.0753)
0.105(0.0742)

0.151(0.0696)

culate that the maximum elasticity of growth with respect to distance

26050.102
26050.086

is −0.0156. As we pointed out earlier, the distance to the line is not an

Dependent Variables: Distance to Infrastructure

excludable instrument for the distance to the railroads because it is also

−0.0491(0.0601)

level panel. Distances to historical lines are computed by the authors. Distance to the expanded set of lines is taken from Faber (2009).

correlated with other forms of transportation infrastructure. However,

by interpreting these two-stage calculations as the upper-bound effects
27440.082

−0.0484(0.0642)

of railroads, we can starkly illustrate the small magnitude of the effect

0.0411(0.0550)

0.0472(0.0563)

of access to transportation on per capita growth relative to the effect on

(5)

26050.083
26050.081

the level of per capita GDP.

One potential issue for interpreting our finding that per capita GDP
Length of Paved Roads

levels are higher in regions near the line is the possibility of dis-
0.00104(0.0460)

placement. For example, the placement of transportation may cause a

−0.0618(0.0327)

−0.0632(0.0315)
0.0107(0.0485)

“crowding-in” effect such that firms relocate to be near the line. This
27440.329

could cause proximity to the line to be positively correlated with pro-

26050.336
26050.336

duction even if the investment in having a line does not increase aggre-
(4)

gate (provincial or national) production from when there is no line. To

investigate this issue, we repeat the estimation on a sample where the
Length of Highway

10% nearest counties are excluded, and then again on samples where
−0.180(0.0814)

−0.453(0.0809)

the 20% are excluded. If the full sample results are caused by produc-
27440.049

26050.158

tive firms relocating to be very near the railroad, then the estimated
effect should decrease in magnitude when we omit those groups (since
−0.0895(0.0752)
−0.441(0.0829)
(3)

one would expect firms that choose to relocate to be close to the line to
26050.162

relocate as close as possible to the line).

Ln Distance to RRLn Distance to Segment City

Table 6 panels B and C provide little support for the crowding-in

hypothesizes. For example, a comparison of the estimates in columns
(1)–(4) between the full sample estimates in panel A to panel C, where
the 20% nearest counties are omitted shows that per capita GDP, if
0.310(0.0398)
0.136(0.0349)

anything, slightly larger in magnitude as we move further away from

26050.432
(0.0360)

(0.0429)
Distance from the line and transportation infrastructure.

0.327
0.189

0.390
0.178

2744

2605
(2)

35
To check that our results are not driven by the particular lag structure of
0.348(0.0699)
0.369(0.0708)

0.169(0.0670)
0.202(0.0656)

the controls, we alternatively control for 3, 4 or 5 year moving averages of

26050.182
27440.084

26050.163

lag per capita GDP. Our results are robust and we find no effect of distance to
the line on growth. The estimated coefficients are similarly small in magnitude
(1)

and statistically insignificant. The sample size becomes smaller as we introduce

longer lags and the estimates become more imprecise. These estimates are not
C. Historical and Expanded Lines

Ln Distance to Expanded Lines

Ln Distance to Expanded Lines

shown for the sake of brevity, but are available upon request. Note that one
could alternatively control for per capita GDP in the first year of the panel. We
Observations R-squared
Observations R-squared

Observations R-squared

do not do this because the unbalanced nature of our panel means that we would
Ln Dist Historical Line
Ln Dist Historical Line

lose too many observations.

A. Historical Lines

B. Expanded Lines

36 Note that the estimates above avoid the Nickell (1981) bias as we do not

control for lag growth. To check that our results are not driven by this choice
of specification, we also estimate the growth regression using the more tradi-
Table 4

tional method of the Arrellano-Bond System Dynamic Panel Estimation, where we

control for the lag of per capita growth rate (Arellano and Bond, 1991). The
estimates are presented in Appendix Table A.1 panel A. The estimates are small
in magnitude and statistically insignificant. Thus, they are consistent with our
main estimates that distance from the lines have little effect on growth.

14
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 5
The effect of distance to the line on production levels.
Dependent Variable Ln Per Capita GDP
(1) (2) (3) (4) (5) (6)
Ln Distance to Historical Lines −0.0617–0.0434(0.0286) −0.0491 −0.0581 −0.0699 −0.0681
(0.0281) (0.0277) (0.0265) (0.0270) (0.0272)
Ln Distance to Segment City 0.065 −0.061 −0.063 −0.178 −0.208
(0.232) (0.266) (0.282) (0.325) (0.298)
Ln Distance to Segment City2 −0.0276 −0.0089 −0.0071 0.0072 0.0101
(0.0277) (0.0308) (0.0324) (0.0372) (0.0344)
Ln Distance to Navigable River 0.318 0.321 0.366 0.385
(0.153) (0.141) (0.140) (0.138)
Ln Distance to Navigable River2 −0.0517 −0.0481 −0.0550 −0.0554
(0.0200) (0.0186) (0.0188) (0.0183)
Ln Area −1.572 −1.442 −1.441
(0.681) (0.635) (0.642)
Ln Area2 0.0983 0.0911 0.0894
(0.0486) (0.0459) (0.0464)
Ln Distance to Coastline −0.243 −0.207
(0.224) (0.219)
Ln Distance to Coastline2 0.0168 0.0141
(0.0265) (0.0259)
Ln Distance to Country Border −16.49
(6.097)
Ln Distance to Country Border2 1.241
(0.459)
Observations R-squared 2744 2744 2744 2744 2744 2744
0.818 0.826 0.833 0.845 0.849 0.852

Notes: All regressions control for year and province fixed effects. Standard errors are clustered at the county level. These
estimates use an unbalanced county-year level panel. GDP data are from Provincial Statistical Yearbooks. All geographic
variables are computed by the authors.

Table 6
The effect of distance to the line on production levels and growth.
Dependent Variables: Per Capita GDP
Ln(Per Capita GDP) Annual Growth in Ln(Per Capita GDP)
All Primary Secondary Tertiary All Primary Secondary Tertiary
(1) (2) (3) (4) (1) (2) (3) (4)
A. Full Sample
Ln Dist −0.0681 −0.0353 −0.0944 −0.0773 −0.00229 −0.00025 −0.00787 −0.00104
Historical (0.0272) (0.0216) (0.0458) (0.0324) (0.00339) (0.00523) (0.00521) (0.00357)
Line
Observations 2744 2266 2266 2199 1110 1018 1018 899
R-squared 0.852 0.870 0.722 0.842 0.204 0.225 0.182 0.181

B. Omit 10% Nearest

Ln Dist −0.0677 −0.0492 −0.0634 Dependent −0.00180 0.00942 −0.00763 −0.00437
Historical (0.0396) (0.0408) (0.0643) Variables: (0.00541) (0.00754) (0.00889) (0.00709)
Line Per Capita
GDP
Observations 2480 2031 2031 Ln(Per 1004 914 914 808
R-squared 0.854 0.870 0.722 Capita GDP) 0.226 0.285 0.170 0.175

C. Omit 20% Nearest

Ln Dist −0.104 −0.117 −0.109 −0.077 −0.00102 0.00998 −0.00325 −0.00374
Historical (0.050) (0.049) (0.082) (0.054) (0.00556) (0.01130) (0.00998) (0.00610)
Line
Observations 2200 1789 1789 1728 882 798 798 692
R-squared 0.860 0.879 0.721 0.849 0.256 0.306 0.163 0.173

Notes: All regressions control for the full set of baseline controls in Table 5 column (6). The growth regressions in columns (5)–(8) also
control for the one and two year lags of sector-specific per capita GDP levels. Standard errors are clustered at the county level. The
sample is an unbalanced county-year level panel. The GDP data are from Provincial Statistical Yearbooks. All geographic variables are
computed by the authors.

the line. As with the full sample, we find no effect on per capita GDP the number and average profits of manufacturing firms. Panel A shows
growth. the estimates for the full sample. Columns (1)–(4) show that distance
from the line results in fewer firms. The estimates are statistically sig-
6.3. The effect on firm placement and household income nificant at the 1% level for all firm ownership types. The coefficient in
column (1) indicates that increasing the distance by 1% will result in a
Table 7 shows the estimated effects of the distance from the line on 0.09% reduction in the number of firms. In columns (5)–(8), we exam-

15
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 7
The effect of distance to the line on firm location and profits.
Dependent Variables Per Capita GDP
Ln Number of Firms Average Firm Profit
(1) (2) (3) (4) (5) (6) (7) (8)
All Public Mixed Individual All Public Mixed Individual

A. Full Sample
Ln Distance −0.091 −0.062 −0.089 −0.122 −0.105 −0.100 −0.072 −0.028
to Historical (0.022) (0.025) (0.033) (0.032) (0.040) (0.048) (0.067) (0.043)
Lines
Observations 3321 3321 3321 3321 2763 2416 1642 1503
R-squared 0.639 0.680 0.700 0.785 0.449 0.319 0.203 0.342

B. Omit Nearest 10%

Ln Distance −0.127 −0.095 −0.101 −0.146 −0.150 −0.167 −0.095 −0.046
to Historical (0.035) (0.036) (0.054) (0.054) (0.065) (0.076) (0.095) (0.070)
Lines
Observations 2994 2994 2994 2994 2477 2151 1460 1343
R-squared 0.647 0.693 0.691 0.776 0.443 0.322 0.211 0.348

C. Omit Nearest 20%

Ln Distance −0.152 −0.076 −0.138 −0.158 −0.174 −0.063 −0.011 −0.018
to Historical (0.046) (0.046) (0.068) (0.073) (0.085) (0.090) (0.122) (0.099)
Lines
Observations 2663 2663 2663 2663 2196 1897 1288 1182
R-squared 0.657 0.695 0.701 0.772 0.449 0.332 0.223 0.337

Notes: All regressions control for the full set of baseline controls in Table 5 column (6). Standard errors are clustered at the county level.
The sample is an unbalanced county-year level panel. The data for the dependent variables are from the Censuses of Manufacturing
Firms. All geographic variables are computed by the authors.

ine log average firm profits. The estimates show that amongst counties (−0.105 × 3.68 = −0.386). In contrast, average firm profits grew
that have at least one firm (of the relevant type), distance results in at approximately sixty percent per year during 2004-6. If this was sus-
lower profits. The estimates are statistically significant for all firms and tained for eighteen years, the cumulative growth in firm profits would
publicly owned firms at the 1% and 5% levels. Column (5) shows that be 2950%. While this crude estimate of cumulative profit growth is
a 1% increase in distance results in a 0.1% reduction in average firm likely to be significantly higher than actual firm profit growth over the
profits. The estimates for mix ownership and individually owned firms eighteen year period, it nevertheless illustrates the fact that the implied
are negative but imprecisely estimated. elasticities between distance from the line and firm profits is relatively
As with our earlier exercise, we assess the magnitude of our esti- small in size.
mates by comparing our estimates of the effect of distance on the In panels B and C, we repeat the estimates on samples where the
number of firms across space to the total increase in the number of nearest 10% and 20% counties to the line are excluded. The estimates
firms over time. Since the 75th-percentile county in terms of distance do not decline monotonically as we examine more distant firms. Thus,
from the line is approximately 4.68 times further away than the 25th- our finding that more firms locate nearer the line is unlikely to reflect
percentile county, our estimate in column (1) implies that it should have a crowding-in effect.
approximately 42.6 percent fewer firms (−0.091 × 4.68 = −0.426).37 We also examined the effect of distance on the growth of the num-
During the three years for which our data use a consistent sampling ber of firms, the growth of average firm profits and the returns to cap-
frame (2004-06), the average number of firms per county grew by ital as measured by profits divided by the value of total capital. These
twelve percent from approximately 83 to 93 firms per county.38 Rel- estimates were negative, small in magnitude and statistically insignif-
ative to the change over time, our estimate of the distance therefore icant. We do not report them in the paper for the sake of brevity and
implies a large effect. However, this is mostly an artifact of the short because of concerns over the quality of the data for returns to capital.
time horizon of the firm panel data. For example, if the number of Specifically, it is unclear how capital is valued by these firms. Much
firms had grown at the same rate (approximately five percent per year) of the capital is inherited from the state or collectives and one would
for eighteen years (which is the sample length of our GDP data), then only know the market value if she observed the market transaction of
the number of firms would have grown by approximately 130% from another similar piece of capital. If further away regions have fewer mar-
approximately 40 to 91 firms per county. Relative to the cumulative ket transactions such that firms there are more likely to under-value the
growth over the longer time horizon, the implied effect of distance on capital, then our estimate of the returns to capital will systematically
the number of firms appear to be higher. over-state the effect of the line as we move further away from the line.
Repeating the same calculation for average firm profits, the esti- This measurement issue is a generic problem in the Chinese data on
mate in column (5) implies that firms in the 75th-percentile county firm assets.
in terms of distance from the line should have approximately 39% Table 8 shows the estimated effects of distance on average house-
lower profits than firms in the 25th-percentile county on average hold income for agricultural households at the village level. Panel A
column (1) shows that distance from the line is negatively correlated
with the Gini coefficient for village household incomes. The estimate
37
Note that the geographic coverage varies across the samples in Tables 6–8, is statistically significant at the 5% level. In Column (5), we estimate
which means that the distance from the line for the 75th and 25th percentile the effect of distance on the annual change of the Gini coefficient. It
counties will also differ. shows that distance from the line is correlated with slower growth in
38
Recall that the 1993 firm data is from a census of all industrial plants and inequality. The estimate is statistically significant at the 1% level. The
has a different sampling frame relative to the Census of Manufacturing firms estimated effects on income growth are statistically indistinguishable
which includes the former and, in addition, privately owned manufacturing
firms with five million RMB or more in revenues.

16
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 8
The effect of distance to the line on income inequality.
Dependent Variables: Household Income Distribution
Ln (HH Income) by Percentile Annual Growth Rate
(1) (2) (3) (4) (5) (6) (7) (8)
Gini 10th 50th 90th Gini 10th 50th 90th

A. Full Sample
Ln Distance −0.0071 −0.0285 −0.0120 −0.0158 −0.00195 −0.00018 −0.00182 −0.00257
to Historical (0.0031) (0.0255) (0.0236) (0.0276) (0.00058) (0.00330) 0.00193) (0.00270)
Lines
Observations 1897 1897 1897 1897 1533 1533 1533 1533
R-squared 0.272 0.454 0.590 0.582 0.036 0.124 0.171 0.095

B. Omit Nearest 10%

Ln Distance −0.0050 0.0176 0.0291 0.0198 −0.00184 0.00003 0.00191 −0.00287
to Historical (0.0054) (0.0393) (0.0328) (0.0418) (0.00095) (0.00454) (0.00342) (0.00463)
Lines
Observations 1722 1722 1722 1722 1392 1392 1392 1392
R-squared 0.253 0.478 0.596 0.583 0.041 0.144 0.200 0.098

C. Omit Nearest 20%

Ln Distance −0.0025 0.0582 0.0696 0.0625 −0.00079 0.00340 0.00523 0.00394
o Historical (0.0075) (0.0477) (0.0403) (0.0556) (0.00113) (0.00574) (0.00445) (0.00508)
Lines
Observations 1531 1531 1531 1531 1238 1238 1238 1238
R-squared 0.262 0.499 0.619 0.606 0.043 0.134 0.185 0.088

Notes: All regressions control for the full set of baseline controls in Table 5 column (6). The regression in column (5) also controls for
the 1 and 2 year lag of the gini, and the regressions in columns (6)–(8) control for the 1 and 2 year lags of the relevant income levels.
Standard errors are clustered at the county level. The sample is an unbalanced panel of villages. Income data are from the National
Fixed Point Survey. All geographic data are computed by the authors.

Table 9
Historical lines and expanded lines.
Dependent Variables
Ln PC GDP Gini Ln Total # Firms
(1) (2) (3) (4) (5) (6)
Ln Distance to Hist Line −0.0823 −0.0826 −0.00715 −0.00795 −0.102 −0.101
(0.0276) (0.0279) (0.00333) (0.00383) (0.025) (0.025)

Ln Distance to Expanded Lines −0.0224 −0.00363 −0.029

(0.0326) (0.00589) (0.028)
Observations 2605 2605 1070 1070 2704 2704
R-squared 0.860 0.860 0.319 0.320 0.665 0.666

Notes: All regressions control for the full set of baseline controls in Table 5 column (6). Standard errors are clustered at
the county level. The sample is an unbalanced county-year level panel. The GDP data used in column (1)–(2) are from
Provincial Statistical Yearbooks. The income data used in columns (3)–(4) are from the National Fixed Point Surveys.
The firm data used in columns (5)–(6) are from the Censuses of Manufacturing Firms. The distance to the historical line
is calculated by the authors. The distance to the expanded set of lines is taken from Faber (2009).

from zero.39 lines is positively correlated with proximity to transportation infrastruc-

In panels B and C, we present the results for restricted samples ture such as railroads and coastal routes. However, our estimates also
where we omit the 10% and 20% counties nearest the line. The esti- suggested that our lines are uncorrelated with the more recently con-
mates show that the effect on household income inequality is mainly structed paved motorways, which have been found by Faber (2014) to
driven by the nearest counties. This might reflect that these areas are also affect production and growth. In this section, we test that our main
the ones that both gain the most in terms of trade opportunities, but results are robust to controlling for access to such recent transporta-
also lose the most from capital (and, though it is not in our model, tion infrastructure. Specifically, we directly control for distance to the
human capital) mobility. expanded set of lines constructed by an earlier working paper version
of Faber (2014), Faber (2009), which the author kindly shared with us.
6.4. Robustness This expanded set of lines includes our historical lines of transportation
and adds many additional lines to capture recently constructed road
6.4.1. Additional lines networks.
One obvious concern with our strategy regards the relevance of our First, we investigate the difference between the historical and
historical lines. Earlier in this section, we showed that proximity to our expanded lines in terms of how each correlates to transportation net-
works. Table 4 Panel B shows that on average, the distance from the
expanded set of lines is positively correlated to distance from rail-
39
Arrellano-Bond estimates are presented in Appendix Table A1 panel B. They roads and the distance from the segment city; negatively correlated
are consistent with the main results in showing that distance from the line with the length of highways and roads; and uncorrelated with distance
has little effect on income growth. The estimates are small in magnitude and from rivers, coastline and country borders. Therefore the key difference
statistically insignificant.

17
 A. Banerjee et al.
Table 10
Correlation between distance to the historical lines and geo-climatic determinants of growth.
Dependent Variable
Natural Conditions Man-made Factors
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Buckwheat Maize Rice (Wet) Sorghum Soy Sweet Potato Wheat White Potato Log Spring Log Spring Log Dist to Exposure to Taiping
Temp Rain Grand Canal Tianguo

A. County GDP Sample

Ln Dist −11.59 −83.36 −48.89 −87.26 −50.35 107.8 −87.41 −118.3 −0.00013 −0.0148 −0.0334 −0.0260
Historical (11.11) (57.67) (39.50) (46.55) (20.25) (48.24) (41.98) (60.53) (0.00152 (0.0073 (0.0389) (0.0259)
Line
Observations 295 295 295 295 295 295 295 295 177 177 295 295
R-squared 0.638 0.800 0.907 0.865 0.752 0.949 0.627 0.334 0.828 0.928 0.791 0.222

B. Firms Sample
18

Ln Dist −18.05 5.929 28.95 −23.53 −12.80 68.18 −114.0 −105.3 −0.00119 −0.00548 −0.0416 −0.0099
Historical (5.648) (26.15) (24.87) (20.97) (1037). (30.35) (17.71) (22.80) (0.00069) (0.00312) (0.0141) (0.0085)
Line
ObservationsR- 1817 1817 1817 1817 1817 1817 1817 1817 1098 1098 1817 1817
squared 0.706 0.693 0.647 0.808 0.647 0.806 0.720 0.578 0.784 0.905 0.822 0.347

C. Household Income Sample

Ln Dist −5.565 60.48 23.21 −16.56 −9.501 59.92 −47.01 −82.65 −0.00183 0.0119 −0.0471 −0.0304
Historical (12.81) (64.92) (59.89) (68.39) (16.98) (142.3) (37.19) (74.34) (0.00211) (0.0175) (0.0635) (0.0331)
Line
Observations 122 122 122 122 122 122 122 122 86 86 122 122
R-squared 0.784 0.885 0.848 0.925 0.934 0.874 0.906 0.812 0.854 0.912 0.827 0.484

Notes: All regressions are cross-sectional estimates. They control for the full set of baseline controls in Table 5 column (6) in the manuscript, except for the year fixed effects. Robust standard errors are presented.

Journal of Development Economics 145 (2020) 102442

Suitability data are computed from the FAO’s GAEZ database. Weather data are computed from The Terrestrial Air Temperature and Precipitation: Monthly and Annual Time Series (1950–1996). In column
(11), the distance to the Grand Canal is computed from the ArcGIS database provided by the Harvard Yenching Insitute (2016), CHGIS, Version 6, Cambridge. In column (12), exposure is a dummy variable
that equals one if a county was part of the Taiping Kingdom or ever experienced a Taiping-related battle. The data are coded by the authors based on Hua (1991).
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 11
Robustness to additional controls – agro-climatic suitability for the cultivation of staple crops.
Dependent Variables: Per Capita GDP
Ln(Per Capita GDP) Annual Growth in Ln(Per Capita GDP)
All Primary Secondary Tertiary All Primary Secondary Tertiary
(1) (2) (3) (4) (5) (6) (7) (8)
Ln Dist −0.0681 −0.0353 −0.0944 −0.0773 −0.00229 −0.00025 −0.00787 −0.00104
Historical (0.0272) (0.0216) (0.0458) (0.0324) (0.00339) (0.00523) (0.00521) (0.00357)
Line

A. Baseline
Observations 2744 2266 2266 2199 1110 1018 1018 899
R-squared 0.852 0.870 0.722 0.842 0.204 0.225 0.182 0.181
Observations 2744 2266 2266 2199 1110 1018 1018 899
R-squared 0.852 0.870 0.722 0.842 0.204 0.225 0.182 0.181

B. Control for Agro-climatic Suitability∗

Ln Dist −0.0537 −0.0363 −0.0728 −0.0587 −0.00041 0.00156 −0.00568 −0.00099
Historical (0.0268) (0.0211) (0.0457) (0.0319) (0.00333) (0.00496) (0.00507) (0.00344)
Line
Observations 2744 2266 2266 2199 1110 1018 1018 899
R-squared 0.876 0.887 0.761 0.868 0.212 0.233 0.188 0.187

Ln Number of Firms Average Firm Profits

All Public Mixed Individual All Public Mixed Individual

C.Baseline
Ln Dist −0.091 −0.062 −0.089 −0.122 −0.105 −0.100 −0.072 −0.028
Historical (0.022) (0.025) (0.033) (0.032) (0.040) (0.048) (0.067) (0.043)
Line
Observations 3321 3321 3321 3321 2763 2416 1642 1503
R-squared 0.639 0.680 0.700 0.785 0.449 0.319 0.203 0.342

D. Control for Agro-climatic Suitability∗

Ln Dist −0.0816 −0.0649 −0.0614 −0.1010 −0.0913 −0.0808 −0.0822 −0.0238
Historical (0.0210) (0.0260) (0.0324) (0.0312) (0.0391) (0.0475) (0.0665) (0.0428)
Line
Observations 3321 3321 3321 3321 2763 2416 1642 1503
R-squared 0.691 0.690 0.725 0.813 0.455 0.335 0.219 0.359

Ln(HH Income) by Percentile Annual Growth Rate

Gini 10th 50th 90th Gini 10th 50th 90th

E Baseline
Ln Dist −0.0071 −0.0285 −0.0120 −0.0158 −0.00195 −0.00018 −0.00182 −0.00257
Historical (0.0031) (0.0255) (0.0236) (0.0276) (0.00058) (0.00330) (0.00193) (0.00270)
Line
Observations 1897 1897 1897 1897 1533 1533 1533 1533
R-squared 0.272 0.454 0.590 0.582 0.036 0.124 0.171 0.095

F. Control for Agro-climatic Suitability∗

Ln Dist −0.0046 −0.0448 −0.0233 −0.0198 −0.00153 −0.00286 −0.00262 −0.00280
Historical (0.0032) (0.0242) (0.0224) (0.0274) (0.00061) (0.00326) (0.00195) (0.00281)
Line
Observations 1897 1897 1897 1897 1533 1533 1533 1533
R-squared 0.310 0.481 0.613 0.595 0.038 0.126 0.172 0.097

Notes: All regressions control for agro-climatic suitability for the cultivation of buckwheat, maize, wet rice, sorghum, soy, sweet potato,
wheat, white potato, and the full set of baseline controls in Table 5 column (6) in the manuscript. Suitability data are computed from
the FAO’s GAEZ database. All standard errors are clustered at the county level. In Panels A–D, the samples are unbalanced county-year
level panels. In Panels A and B, the data for the dependent variables are from from the Provincial Statistical Yearbooks. In Panels C
and D, the data for the dependent variables are from the Censuses of Manufacturing Firms. In Panels E and F, the regression in column
(5) also controls for the 1 and 2 year lag of the gini, and the regressions in columns (6)–(8) control for the 1 and 2 year lags of the
relevant income levels. The sample is an unbalanced panel of villages. Income data are from the National Fixed Point Survey.

between the expanded set of lines and our historical lines is that the dis- In Table 9, we test the robustness of our main estimates by running
tance to the former are negatively correlated with road density (recall a “horse race” between the historical and expanded sets of lines. For
that we control for area of the county), while the distance to the latter brevity, we focus on the main outcomes of interest.40 Note that the
are uncorrelated with road density. This is consistent with the fact that sample size is smaller than the one for our main estimates because the
the new lines capture new road networks built away from the railroads. data in (Faber, 2009) do not exactly match to ours. Nevertheless, our
In Panel C, we examine the correlations of our historical and expanded baseline estimates from using this restricted sample are similar to those
set of lines with transportation infrastructure in one regression. The cor-
relation between our historical lines and transportation infrastructure
are robust to controlling for the additional lines. 40
Results using other outcomes are consistent in showing that our main spec-
ification is very robust. They are available upon request.

19
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 12
Robustness to additional controls – weather conditions.
Dependent Variables: Per Capita GDP
Ln(Per Capita GDP) Annual Growth in Ln(Per Capita (GDP)
All Primary Secondary Tertiary All Primary Secondary Tertiary
(1) (2) (3) (4) (5) (6) (7) (8)

A. Baseline

Ln Dist Historical Line −0.0681 −0.0353 −0.0944 −0.0773 −0.00229 −0.00025 −0.00787 −0.00104
(0.0272) (0.0216) (0.0458) (0.0324) (0.00339) (0.00523) (0.00521) (0.00357)

Observations R-squared 2744 2266 2266 2199 1110 1018 1018 899
0.852 0.870 0.722 0.842 0.204 0.225 0.182 0.181

B. Control for Spring Temperature and Precipitation

Ln Dist −0.0657 −0.0040 −0.1070 −0.0821 −0.00041 0.00156 −0.00568 −0.00099
Historical (0.0376) (0.0313) (0.0638) (0.0446) (0.00333) (0.00496) (0.00507) (0.00344)
Line
Observations 1745 1450 1450 1389 1110 1018 1018 899
R-squared 0.870 0.885 0.752 0.861 0.212 0.233 0.188 0.187

Ln Number of Firms Average Firm Profits

All Public Mixed Individual All Public Mixed Individual

C.Baseline
Ln Dist −0.091 −0.062(0.025) −0.089 −0.122 −0.105 −0.100 −0.072 −0.028
Historical (0.022) (0.033) (0.032) (0.040) (0.048) (0.067) (0.043)
Line
Observations 3321 33210.680 3321 3321 2763 2416 1642 1503
R-squared 0.639 0.700 0.785 0.449 0.319 0.203 0.342

D. Control for Spring Temperature and Precipitation

Ln Dist −0.0638 −0.0618 −0.0892 −0.0823 −0.0891 −0.0843 −0.0325 −0.0247
Historical (0.0291) (0.0320) (0.0450) (0.0431) (0.0528) (0.0636) (0.0912) (0.0663)
Line
Observations 2128 2128 2128 2128 1731 1522 1002 905
R-squared 0.692 0.698 0.712 0.775 0.407 0.294 0.236 0.313

Ln(HH Income) by Percentile Annual Growth Rate

Gini 10th 50th 90th Gini 10th 50th 90th

E.Baseline
Ln Dist −0.0071 −0.0285 −0.0120 −0.0158 −0.00195 −0.00018 −0.00182 −0.00257
Historical (0.0031) (0.0255) (0.0236) (0.0276) (0.00058) (0.00330) (0.00193) (0.00270)
Line
Observations 1897 1897 1897 1897 1533 1533 1533 1533
R-squared 0.272 0.454 0.590 0.582 0.036 0.124 0.171 0.095

F. Control for Spring Temperature and Precipitation

Ln Dist −0.0099 −0.0032 0.0056 −0.0003 −0.00405 0.00191 −0.00178 −0.00565
Historical (0.0041) (0.0363) (0.0291) (0.0294) (0.00093) (0.00470) (0.00230) (0.00376)
Line
Observations 1344 1344 1344 1344 1087 1087 1087 1087
R-squared 0.336 0.504 0.634 0.649 0.048 0.134 0.181 0.130

Notes: All regressions control for log average spring temperature and log average spring precipitation, and the full set of baseline controls in
Table 5 column (6) in the manuscript. Weather data is reported by the Terrestrial Air Temperature and Precipitation: Monthly and Annual Time
Series (1950–1996) data set. All standard errors are clustered at the county level. In Panels A–D, the samples are unbalanced county-year level
panels. In Panels A and B, the data for the dependent variables are from from the Provincial Statistical Yearbooks. In Panels C and D, the data
for the dependent variables are from the Censuses of Manufacturing Firms. In Panels E and F, the regression in column (5) also controls for the
1 and 2 year lag of the gini, and the regressions in columns (6)–(8) control for the 1 and 2 year lags of the relevant income levels. The sample
is an unbalanced panel of villages. Income data are from the National Fixed Point Survey.

from using our full sample. The estimates show that our baseline esti- than we use to capture the historical network (see Faber (2014) Fig. 3).
mates of the effect of historical lines are very robust to controlling for There are tradeoffs to increasing the number of lines. One the one
the additional lines and suggest that the historical lines are indeed the hand, it follows the modern network more closely, which increases the
relevant lines to study in our context. strength of the first stage if used as an instrumental variable as in Faber
Faber (2014) finds that Chinese counties that are along the way (2014). On the other hand, it introduces the concern that increased dis-
between two cities connected by the modern trunk network have lower tance from one line could reflect proximity to another line. Because of
GDP growth than unconnected cities that are not along the way. For the latter concern, we choose to have a more parsimonious set of lines.
comparison purposes, note two important differences between his anal-
ysis and ours. The first is the sample. As we note above, the geographic
coverage is different. Also, his study focuses on 1997–2006, whereas 6.4.2. Additional controls
our data coverage begins in 1986. The second is that his study uses Table 10 examines the relationship between the distance to the
many more lines to capture the modern trunk work laid out in 1997 line and potential drivers of growth. We categorize these factors into

20
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table 13
Robustness to additional controls – log distance to the grand canal.
Dependent Variables: Per Capita GDP
Ln(Per Capita GDP) Annual Growth in Ln(Per Capita GDP)
All Primary Secondary Tertiary All Primary Secondary Tertiary
(1) (2) (3) (4) (5) (6) (7) (8)

A.Baseline
Ln Dist Historical Line −0.0681 −0.0353 −0.0944 −0.0773 −0.00229 −0.00025 −0.00787 −0.00104
(0.0272) (0.0216) (0.0458) (0.0324) (0.00339) (0.00523) (0.00521) (0.00357)
Observations R-squared 2744 2266 2266 2199 1110 1018 1018 899
0.852 0.870 0.722 0.842 0.204 0.225 0.182 0.181

B. Control for Ln Dist to Grand Canal

Ln Dist Historical Line −0.0671 −0.0345 −0.0904 −0.0755 −0.00284 −0.00165 −0.00846 −0.00145
(0.0267) (0.0215) (0.0452) (0.0319) (0.00332) (0.00504) (0.00510) (0.00352)
Observations 2744 2266 2266 2199 1110 1018 1018 899
R-squared 0.854 0.870 0.725 0.843 0.206 0.229 0.183 0.182

Ln Number of Firms Average Firm Profits

All Public Mixed Individual All Public Mixed Individual

C Baseline
Ln Dist Historical Line −0.091 −0.062 −0.089 −0.122 −0.105 −0.100 −0.072 −0.028
(0.022) (0.025) (0.033) (0.032) (0.040) (0.048) (0.067) (0.043)
Observations R-squared 3321 3321 3321 3321 2763 2416 1642 1503
0.639 0.680 0.700 0.785 0.449 0.319 0.203 0.342

D. Control for Ln Dist to Grand Canal

Ln Dist Historical Line −0.092 −0.065 −0.083 −0.121 −0.106 −0.104 −0.073 −0.034
(0.022) (0.025) (0.033) (0.032) (0.040) (0.048) (0.067) (0.043)
Observations R-squared 3321 3321 3321 3321 2763 2416 1642 1503
0.639 0.680 0.701 0.785 0.449 0.319 0.203 0.343

Ln(HH Income) by Percentile Annual Growth Rate

Gini 10th 50th 90th Gini 10th 50th 90th

E.Baseline
Ln Dist Historical Line −0.0071 −0.0285 −0.0120 −0.0158 −0.00195 −0.00018 −0.00182 −0.00257
(0.0031) (0.0255) (0.0236) (0.0276) (0.00058) (0.00330) (0.00193) (0.00270)
Observations R-squared 1897 1897 1897 1897 1533 1533 1533 1533
0.272 0.454 0.590 0.582 0.036 0.124 0.171 0.095

F. Control for Ln Dist to Grand Canal

Ln Dist Historical Line −0.0072 −0.0299 −0.0124 −0.0157 −0.00196 −0.00014 −0.00177 −0.00239
(0.0032) (0.0257) (0.0237) (0.0277) (0.00058) (0.00334) (0.00194) (0.00270)
Observations R-squared 1897 1897 1897 1897 1533 1533 1533 1533
0.272 0.455 0.590 0.582 0.036 0.124 0.171 0.095

Notes: All regressions control for log distance to the Grand Canal, and the full set of baseline controls in Table 5 column (6) in the manuscript.
The distance to the Grand Canal is computed from the ArcGIS database provided by the Harvard Yenching Insitute (2016), CHGIS, Version 6,
Cambridge. In Panels A and B, the data for the dependent variables are from from the Provincial Statistical Yearbooks. In Panels C and D, the
data for the dependent variables are from the Censuses of Manufacturing Firms. In Panels E and F, the regression in column (5) also controls
for the 1 and 2 year lag of the gini, and the regressions in columns (6)–(8) control for the 1 and 2 year lags of the relevant income levels. The
sample is an unbalanced panel of villages. Income data are from the National Fixed Point Survey.

two groups: geo-climatic factors and man-made factors. The first group only difference is that because there is only one cross-section of data,
includes agro-climatic suitability to Chinese staple crops (buckwheat, we do not control for year fixed effects.42 Table 10 Panels A, B and C
maize, wet rice, sorghum, soy, sweet potato, wheat, white potato) and columns (1)–(10) present the correlations for the three samples used in
weather. For the latter, we examine the log of spring temperature and our paper. We see that distance to the line is correlated with some of
the log of spring rainfall, because higher rainfall and temperature dur- the measures. However, there is no obvious pattern – some signs are
ing the spring months is the most important predictor of agricultural positive, while others are negative.
production in China on average (Meng et al., 2015).41 To investigate whether our results are confounded by omitted vari-
To examine whether our baseline specification suffers from omitted ables bias, we include all of the suitability measures into the baseline.
variable bias, we regress each suitability and weather variable on the Table 11 shows that the results are very similar. A few of the coefficients
distance to the line while including the full set of baseline controls. The become slightly less precisely estimated when we add the large number
of controls, but are statistically similar in magnitude to the baseline.
In Table 12, we include the two weather measures in the baseline.
41
The data for suitability are computed from the FAO’s GAEZ database. We We examine weather separately from suitability for two reasons. First,
choose irrigation as an input. The results are nearly identical if we choose rain- the sample size for the weather examination is slightly smaller. Second,
fed irrigation as an input. The weather data is reported by the Terrestrial Air
Temperature and Precipitation: Monthly and Annual Time Series (1950–1996) data
42
set. The latter does not cover all counties in China. Thus, the sample size will We present robust standard errors (instead of clustering them at the county
be slightly smaller when we include whether. Appendix F a display the data. level).

21
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

weather conditions are already internalized by the suitability calcula- additional control will not affect our results.
tions and having them both in the same regression would lead to over
controlling. Again, the results are very similar to the baseline.
6.4.3. Omit Border Regions
The second group of omitted variables that we consider is man-
In the last robustness exercise, we omit the border provinces: Hei-
made: the distance to the Grand Canal and exposure to the Taiping
longjiang, Liaoning, Jilin, Inner Mongolia, Qinghai, Xinjiang, Tibet,
Rebellion. The Grand Canal was an important transportation route his-
Guangxi, Gansu and Yunnan. This effectively excludes all of the coun-
torically as it connected the major rivers. In 1855, the course of Yel-
ties that lie beyond the last historical city in the network.44 Appendix
low River, one of the most important in China, changed due to natural
Table A2 presents the results with the full sample from the paper and
causes, which led to the closing of important sections of the Canal.
the restricted sample. They are very similar.
Together with the decline in Qing government administrative capacity,
the Grand Canal became obsolete (until a major revival effort in the
1990s). Our distance measure is computed from the ArcGIS database 7. Conclusion
provided by the Harvard Yenching Insitute (2016).43
Taiping Tianguo (Taiping Rebellion) was a Christian anti-Qing state In this paper, we investigate the effects of access to transportation
which formed in parts of Southern China during 1851–1864. The rebel infrastructure on economic development during the two decades after
kingdom was laid under siege by the Qing government for most of its China opened up to trade and market reforms, when it experienced
existence. The rebellion was eventually crushed by the Qing with the rapid GDP growth. We find that regions closer to historical transporta-
help of Western Imperial powers. The fighting continued for seven years tion networks have higher levels of GDP per capita, higher income
after the fall of the kingdom in 1871. Because the first wave of Treaty inequality, a higher number of firms and higher average firm profits.
Ports that we use to construct the lines were established in 1842-44, the However, these level differences are relatively small in magnitude and
intensity of the Taiping Rebellion should arguably be an outcome vari- we find no evidence that distance affected income growth during this
able rather than a control. One can make this case given the historical period.
evidence (Spence, 1997). Our results do not contradict the Fogelian (Fogel, 1962, 1964) inter-
For data, we rely on a Chinese language source, Hua (1991). We pretation or the view of (Huang, 2008) that during this period of fast
hand-code exposure as a dummy variable that equals one if a county growth, the Chinese government should not have focused so much on
was part of the Taiping rebel kingdom or if a Taiping-connected bat- building transportation infrastructure. However, they are also consis-
tle ever took place in that county. Appendix Figure A.9 maps the data. tent with an alternative explanation where the infrastructure might
The darker regions indicate areas where at least one Taiping-connected have brought sizable benefits for the economy as a whole, but the local-
battle took place. The lighter regions indicated parts of the Taiping ization of the gains (and the overall level of the gains) was limited by
Kingdom. the lack of factor mobility. The fact that we do not see a strong diver-
Table 10 columns (11)–(12) examines the correlations between gence between well and poorly connected areas does not rule out the
these two variables and log distance to the historical lines, controlling possibility that infrastructure had benefits for all of them, but the lack
for the same baseline controls as we did for the natural condition vari- of factor mobility prevented the gains from being concentrated in rela-
ables. The only statistically significant correlation is between the dis- tively better connected areas.
tance to the canal and the distance to the line for the firm sample in These results should not discourage those who believe that invest-
Panel B column (11). ment in transportation infrastructure can promote economic develop-
To ensure that we do not have an omitted variables problem, we ment. Rather, they highlight the importance of other factors which
re-estimate the baseline for all of the main results including log dis- determine the economic impact of infrastructure. Moreover, as we
tance to the Grand Canal. Table 13 shows that our results are entirely noted in the introduction of this paper, without knowing the returns
robust to its inclusion. We do not attempt to control for exposure to the of such investment, one cannot say whether investments in transporta-
Taiping Rebellion since it may be an outcome of the European inva- tion infrastructure ought to be made. Finding credible ways to estimate
sions. However, note that the our finding no correlation between the or even bound the social returns remains a very important next step in
Taiping Rebellion and the distance to the line means that including this this research agenda.

Appendix.

A. Model
There are a number of reasons why good transportation infrastructure can be advantageous for economic development. First, it plausibly reduces
trade costs and promotes market integration. This should lead to a convergence in prices, reduce price volatility and reallocate resources along the
lines of comparative advantage. It also increases market size, which allows firms to capture gains from specialization, increasing returns, and
promotes more intense competitio. Second, it promotes factor mobility – e.g., it is easier to migrate to the city if one can return easily whenever
needed; easier to lend to a borrower whose project one can visit; and easier to deposit one’s savings in a bank if the bank is more accessible. Third,
it is easier to take advantage of opportunities for investment in human capital – e.g., one can send her child to a better school or take her to a better
doctor. Finally, there are intangible benefits. For example, freer movement of people and goods may bring with it new aspirations, new ideas and
information about new technologies.

B. A Simple Model of Trade and Factor Mobility

The goal of the model is to look at the effects of distance in a setting where distance affects both the mobility of goods and that of factors of
production. The model will illustrate how access to infrastructure can produce very different results depending on which of the two is more affected
by distance. In order to get at these issues in the most direct possible ways, we shut down many of the standard dynamic effects coming from capital
accumulation and population growth. We recognize that excluding capital accumulation would be an especially bad assumption if we were trying

43
CHGIS, Version 6, Cambridge, MA. Appendix Fig. A.5 shows the map of the
Grand Canal.
44
Appendix Fig. A.8 shows a map of these regions.

22
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

to quantitatively match the performance of the Chinese economy. That is not our goal here. Instead, we simply aim to qualitatively understand the
consequences of there being multiple types of mobility, and bringing in the accumulation of factors is unlikely to add important new insights.

B.1. Building Blocks

There are M + N + 1 regions in this economy: M distant regions, N connected regions and 1 metropolis. Each region produces one good
exclusively for export which could be the same as or different from the goods that it imports (e.g. food), and another good which it consumes. These
goods could be either identical or differentiated. The key assumption is that the relative price of the exportable good in terms of the importable
good in the “world market” is the same, p. However, distance to the market adds to the cost of exporting. We model this by assuming that this
transportation cost is increasing in distance from the market such that the price received by the exporters is p in the metropolis, p(1 − d1 ) in the
connected region and p(1 − d2 ) in the distant regions, where d2 > d1 .
Production is carried out by a population of firms of identical size in each region. Production requires two inputs which we will call labor and
capital, but could also be labor and human capital with small adjustments in the arguments. Output of the exportable is given by AK 𝛼 L1−𝛼 (K )𝛽
everywhere, where K is the average level of K in firms in that region.45 In other words, in the urban economics tradition, we allow for spillovers
from co-location. However, we assume that the spillovers are not so large as to swamp diminishing returns entirely: 𝛼 + 𝛽 < 1.46 Assume for the
time being that there is no other technology of production.
The key assumption is with respect to factor mobility. We assume that labor does not move: The metropolis has an endowment of labor of L∗
while all other regions have an endowment of L′ . Capital, on the other hand, does move, but moving is costly. We assume that in equilibrium, the
direction of movement that would be needed is from the various regions to the metropolis. This is consistent with the view that in the initial years
of Chinese growth after 1978, much of the growth and capital accumulation occurred in rural areas, and it was only later that economic freedoms
were extended to urban areas and the urban growth rate crossed its rural counterpart. Therefore, when the rental rate for capital in the metropolis
is r, we assume that the opportunity cost of capital in the connected regions is r(1 − 𝜌d1 ) and that in the distant regions is r(1 − 𝜌d2 ). In other
words, the further one is, the more it costs her to send capital to the metropolis. Therefore, she is willing to accept a lower return on capital if it is
invested in her own region (e.g., because she can monitor the borrower more easily).47 We assume that there are no other constraints on mobility
(e.g., no within-region credit constraints).

B.2. Analysis of the Basic Model

Analysis of this model is straightforward. Profit maximization with respect to the inputs yields the generic conditions:
K
w = p(1 − d)A(1 − 𝛼)( )𝛼 (K)𝛽 and (3)
L
L
r (1 − 𝜌d) = p(1 − d)A𝛼( )1−𝛼 (K )𝛽 ,
K
where w is the wage rate in that type of region, L is the labor endowment, K is the equilibrium amount of capital invested in a firm in that region
and d is the corresponding distance variable (d = 0 for the metropolis, d = d1 for the connected regions and d = d2 for the distant regions). In
addition, there is the capital market clearing condition:

MKD + NKC + KM = K , (4)

where KD is the average amount of capital used in the distant region (per firm), KC is the same thing in a connected region and KM is that in the
metropolis. K is the total supply of capital in the economy.
Manipulating the capital demand condition and using the fact K = K and L = L′ outside the metropolis yields
p ( 1 − d)
K 1−𝛼−𝛽 = A𝛼(L′ )1−𝛼 , (5)
r (1 − 𝜌d)

which tells us that whether the distant regions or the connected ones have more capital per firm depends on whether the ratio ((11−𝜌 −d)
d)
is increasing
or decreasing in d. If 𝜌 > 1, which is the case where capital is less mobile than goods, then the distant region will actually have more capital per
worker. Using the wage-rental ratio as the measure of inequality, as is conventional in trade models, we see that

w (1 − 𝛼)( LK′ )
= . (6)
r (1 − 𝜌d) A𝛼
It follows directly that inequality is higher wherever K is lower. In other words, if capital is less mobile than goods, then the more distant region
would have less inequality because it is able to retain more of its capital. A similar result would hold if we replaced capital by human capital and
used the skill premium to measure inequality.
Finally, we compare outputs per worker/capita,
1
y = p(1 − d)A( ′ )𝛼 (K )𝛼+𝛽 , (7)
L
which can be written as
𝛼+𝛽
1 p ( 1 − d)
y = p ( 1 − d) A ( ′ ) 𝛼 ( A𝛼(L′ )1−𝛼 ) 1−𝛼−𝛽 . (8)
L r (1 − 𝜌d)

45
We could easily let A vary across the locations to capture differences in the flow of ideas.
46
See Duranton and Puga (2004) for a review of this literature.
47
The equivalent assumption for human capital would be that there is a cost to relocating from one’s home region to the city, but the cost is lower if she is better
connected (e.g., because it is easier to travel to and from).

23
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

In the case where 𝜌 < 1, this expression is clearly decreasing in d since both the p(1 − d) term and the rp((11−𝜌 −d)
d)
term decline with d. But when
𝛼+𝛽
𝜌 > 1, we might actually observe the reverse, especially when spillovers are large (1 − 𝛼 − 𝛽 is close to zero) and therefore 1−𝛼−𝛽 is large. Once
again this is because the better connected region loses more of its capital.
Result 1: In the basic model, output per capita will always be higher and inequality lower in the better connected region as long as capital is
more mobile than goods. However, when capital is less mobile than goods, the more distant area will have less inequality. The difference in per
capita output between the regions will tend to be small and per capita output may even be higher in the more distant region.
What is the effect of trade opening in this economy? If we model it as an increase in p, the price of the exportable, it increases incomes everywhere
at the same rate. The rate of growth will not depend on the location.
Result 2: In the basic model, the effect of trade opening would be to raise income levels everywhere in proportion and hence have no differential
growth effect.

B.3. A Simple Extension

The growth result here is somewhat of an artifact of the way the model is set up. Suppose we add an alternative production technology that uses
only labor and produces a perfect substitute for the importable good (e.g. agriculture) using the technology x = BL, where L is the labor input.
The good is consumed at the location and does not need transporting. The point is that now the wage in the exporting sector, w, needs to be bigger
than B for there to be production of exportable goods. In this model, there can be three types of equilibria: type A, where both close and distant
locations export; type B, where one of the locations exports and the other does not; and type C, where neither exports. As long as 𝜌 < 1, we know
that wages, which are proportional to output per capita, will be lower in the more distant location, and therefore, if we are in case B, the distant
location will not export. It follows that as long as 𝜌 < 1, the effect of trade opening will either be the same in both areas (types A and C), or the
more connected area will grow faster.
On the other hand, when 𝜌 > 1, it is not clear which of the two locations will have lower wages, and the gap between wages is likely to be
small. Therefore we are more likely to be in either type A or C equilibria. Given the high average growth rate of approximately 8% in both close
and far regions (see Table 2 column 3), scenario A seems more likely at least for the present. This is consistent with the fact that China now has
excellent infrastructure and both near and far places are relatively easily accessed.48 In any case, in both type A and type C equilibria, the effect of
trade opening on growth rates is the same both in close and distant places, unless the effect of the trade shock is just big enough to move one area
from not exporting to exporting but not the other.
Result 3: In the model with “agriculture”, trade opening is likely to benefit the closer area more in terms of growth rates as long as capital is
more mobile than goods. But in the reverse case, growth rates in both close and distant areas should react relatively similarly to trade opening.
Moreover, as long as China is in a type A equilibrium, Result 1 should continue to hold in this case.
To summarize, a pattern where inequality is higher in more connected areas, but output level differences are small and growth rate differences
are absent, is consistent with a setting where capital is less mobile than goods. The lack of a differential growth effect in this scenario is consistent
with an overall beneficial effect of transportation infrastructure, which is what allows both close and more distant areas to be exporting.

B.3.1. Summary. The point of the simple model is to underscore the fact that infrastructure, in theory, can lead to more or less divergence between
close and far areas, and that this depends on the relative mobility of goods and factors. There are, of course, other reasons that affect the extent of
divergence. On the one side, for example, there is a natural force of convergence: the government might reasonably plan to construct infrastructure
where it was previously absent, so that the connectivity in less-connected regions may improve faster than in the better-connected areas, reducing
the difference in trade costs. This is also why we do not focus on the change in infrastructure as an intervening variable, since it is potentially
correlated with the level of infrastructure. On the other side, there may be agglomeration effects that lead the places that were initially somewhat
well connected to get even better connected. For example, that may be where a new university, export processing zone or airport gets built.

C. Pre-1840 Data

1850 population data were originally printed in county- and prefecture-level Gazetteers of the Qing Dynasty, and later aggregated and translated
to modern Chinese by Ge (2000). We manually coded, digitized and geo-referenced the data from Ge (2000), which include 630 counties. Due to
changes in historical county boundaries, we ultimately geo-referenced 588 of these counties.49
Chong, Pi, Fan, Nan indicators are provided in a modern Chinese translation of the Provisional History of the Qing Dynasty by Zhao (1976). For
this revision, we read through the text discussion of each county, extracted and manually coded the measures, digitized and geo-coded them at the
county level. We have this measure for over 1199 counties. Because of changes in county boundaries over time, the sample for analysis is 1117
counties.50

48
In the model, it is possible for even a very poorly connected area to export because we place no lower bound on the interest rate. But if transportation is really
expensive, the interest rate will have to be very negative in the distant areas to permit exporting. It seems likely that capital owners will then prefer to hold cash or
gold and therefore, there will not be any exports.
49
They are mapped in Appendix Fig. A2.
50
They are mapped in Appendix Figs. A3d - Fig. A.3a.

24
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Buddhist temple locations in 1820 are digitized and geo-coded by Harvard CHGIS.51

Table A.1
The Effect of Distance to the Line on Production – Dynamic Panel Estimation
Dependent Variables Growth
All Sectors Primary Secondary Tertiary
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
1 Lag 2 Lag 3 Lag 1 Lag 2 Lag 3 Lag 1 Lag 2 Lag 3.Lag 1.Lag 2.Lag 3.Lag
A. Per Capital GDP Growth
Ln Distance 0.00386 0.00394 −0.00073 −0.00066 −0.00065 0.00571 0.00582 0.00584 0.00267 0.00266 0.00267
to Hist Line: (0.00331) (0.00334) (0.00397) (0.00401) (0.00401) (0.00507) (0.00513) (0.00514) (0.00333) (0.00334) (0.00334)
Observations 1373 1373 1281 1281 1281 1281 1281 1281 1175 1175 1175
Number of 258 258 258 258 258 258 258 258 245 245 245
Counties
B. Rural Household Income Growth
10th 50th 90th
Ln Distance 0.00169 0.00173 0.00173 0.00016 0.00017 0.00017 0.00285 0.00281 0.00281
to Hist Lines (0.00328) (0.00333) (0.00333) (0.00181) (0.00183) (0.00183) (0.00288) (0.00294) (0.00294)
Observations 1655 1655 1655 1655 1655 1655 1655 1655 1655
Number of 122 122 122 122 122 122 122 122 122
Counties
Note: All regressions control for the full set of baseline controls in Table 5 column (6). In panel A, the lagged dependent variable is instrumented using 1, 2 or 3 lags of the
per capita GDP level for the relevant sector. In panel B, the lagged dependent variable is instrumented with 1, 2, or 3 lags of the relevant income level. The standard errors are
clustered at the county level. The sample is an unbalanced panel of counties. The data in panel A are from Provincial Statistical Yearbooks.The data in panel B are from the
National Fixed Point Survey. All geographic data are calculated by the authors. Note that Panel A, Column (3) could not be computed due to a highly singular covariance matrix.

51
They are mapped in Appendix Fig. A4.

25
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Table A.2
Robustness to the Exclusion of Border Regions
Dependent Variables
Ln(Per Capita GDP) Annual Growth in Ln (Per Capita GDP)
All Primary Secondary Tertiary All Primary Secondary Tertiary
(1) (2) (3) (4) (5) (6) (7) (8)
A. Baseline
Ln Dist Historical Line 0.0681 −0.0353 −0.0944 −0.0773 −0.00229 −0.00025 −0.00787 −0.00104
(0.0272) (0.0216) (0.0458) (0.0324) (0.00339) (0.00523) (0.00521) (0.00357)
Observations R-squared 2744 2266 2266 2199 1110 1018 1018 899
0.852 0.870 0.722 0.842 0.204 0.225 0.182 0.181

B. Omit Counties in Border Provinces

Ln Dist Historical Line −0.0793 −0.0270 −0.1190 −0.0862 −0.00151 0.00322 −0.00570 −0.00205
(0.0280) (0.0216) (0.0471) (0.0355) (0.00364) (0.00520) (0.00546) (0.00323)
Observations R-squared 2434 1956 1956 1889 977 885 885 766
0.852 0.868 0.711 0.830 0.213 0.272 0.223 0.246

Ln Number of Firms Average Firm Profits

All Public Mixed Individual All Public Mixed Individual

C. Full Sample
Ln Distance to Historical Lines −0.091 −0.062 −0.089 −0.122 −0.105 −0.100 −0.072 −0.028
(0.022) (0.025) (0.033) (0.032) (0.040) (0.048) (0.067) (0.043)
Observations R-squared 3321 3321 3321 3321 2763 2416 1642 1503
0.639 0.680 0.700 0.785 0.449 0.319 0.203 0.342

D. Omit Counties in Border Provinces

Ln Distance to Historical Lines −0.086 −0.060 −0.080 −0.088 −0.110 −0.119 −0.078 −0.022
(0.023) (0.027) (0.036) (0.033) (0.040) (0.050) (0.069) (0.044)
Observations R-squared 2655 2655 2655 2655 2248 1982 1408 1306
0.604 0.685 0.710 0.826 0.501 0.359 0.220 0.398

Ln(HH Income) by Percentile Annual Growth Rate

Gini 10th 50th 90th Gini 10th 50th 90th

E. Full Sample
Ln Distance to Historical Lines −0.0071 −0.0285 −0.0120 −0.0158 −0.00195 −0.00018 −0.00182 −0.00257
(0.0031) (0.0255) (0.0236) (0.0276) (0.00058) (0.00330) (0.00193) (0.00270)
Observations R-squared 1897 1897 1897 1897 1533 1533 1533 1533
0.272 0.454 0.590 0.582 0.036 0.124 0.171 0.095

F. Omit Counties in Border Provinces

Ln Distance to Historical Lines −0.0062 −0.0205 0.0001 −0.0031 −0.00214 0.00138 −0.00077 −0.00215
(0.0031) (0.0250) (0.0229) (0.0268) (0.00056) (0.00335) (0.00198) (0.00280)
Observations R-squared 1625 1625 1625 1625 1317 1317 1317 1317
0.268 0.473 0.607 0.599 0.041 0.128 0.180 0.099
Notes: All regressions control for the full set of baseline controls in Table 5 column (6) in the manuscript. All standard errors are clustered at the county level.
In Panels A–D, the samples are unbalanced county-year level panel. The data for the dependent variables are from the Provincial Statistical Yearbooks in Panels
A and B, and from the Censuses of Manufacturing Firms in Panels C and D. In Panels E and F, the regression in column (5) also controls for the 1 and 2 year
lag of the gini, and the regressions in columns (6)–(8) control for the 1 and 2 year lags of the relevant income levels. The sample is an unbalanced panel of
villages. Income data are from the National Fixed Point Survey. Panels B, D and F exclude border provinces: Gansu, Yunnan, Heilongjiang, Liaoning, Jilin, Inner
Mongolia, Qinghai, Xinjiang, Tibet and Guangxi.

26
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.1 Historically Important Cities and Treaty Ports. Source: Murphey (1970) page 35.

27
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.2 1850 Population.

28
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.3 18th Century Qing Dynasty Political Economic Indicators

29
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.4 Buddhist Temples in 1820

30
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.5 The Grand Canal

31
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.6 Natural Conditions: Suitability for Cultivating Staple Crops and Weather

32
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.7 Suitability for Cultivating Staple Crops (cont.)

33
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.8 Omitted Border Provinces

34
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Fig. A.9 Taiping Rebellion

References Arellano, M., Bond, S., 1991. Some tests of specification for panel data: Monte Carlo
evidence and an application to employment equations. Rev. Econ. Stud. 58 (2),
Alder, S., 2015. Chinese Roads in India: the Effect of Transport Infrastructure on 277–297.
Economic Development. 2015 Meeting Papers 1447. Society for Economic Atack, J., Bateman, F., Haines, M., Margo, R.A., 2010. Did railroads induce or follow
Dynamics. economic growth? Urbanization and population growth in the American midwest,
Aoyama, Y., Horner, R., 2009. World development report 2009: reshaping economic 1850-60. Soc. Sci. Hist. (2), 171–197.
Geography, by the World Bank. J. Reg. Sci. 50 (5), 1467–9787. Bai, Y., Jia, R., 2016. “Elite recruitment and political stability: the impact of the
abolition of China’s civil service exam. Econometrica 84 (2), 677–733.

35
A. Banerjee et al. Journal of Development Economics 145 (2020) 102442

Banerjee, A., Duflo, E., Qian, N., 2004. The Railroad to Success: the Effect of Access to Jia, R., 2014. “The legacies of forced freedom: China’s treaty ports. Rev. Econ. Stat. 96
Transportation Infrastructure on Economic Growth in China. Mit working papers. (4), 596–608.
MIT. Jowett, A.J., 1989. Patterns of literacy in the People’s Republic of China. Geojournal 18,
Banerjee, A.V., Duflo, E., 2005. Growth theory through the lens of development 417–427, https://doi.org/10.1007/BF00772696.
economics. of Handbook of Economic Growth, chap. 7. In: Aghion, P., Durlauf, S. Keller, W., Shiue, C.H., 2008. Institutions, Technology, and Trade. NBER Working
(Eds.), Handbook of Economic Growth, vol. 1. Elsevier, pp. 473–552. Papers 13913. National Bureau of Economic Research, Inc.
Baum-Snow, N., 2007. Did highways cause suburbanization?∗. Q. J. Econ. 122 (2), Martinez-Bravo, M., Miquel, G.P.I., Qian, N., Yao, Y., 2017. “The Rise and Fall of Local
775–805. Elections in China: Theory and Empirical Evidence on the Autocrat’s Trade-Off,”
Baum-Snow, N., Brandt, L., Henderson, J.V., Turner, M.A., Zhang, Q., 2017. Roads, Working Paper 24032. National Bureau of Economic Research.
railroads, and decentralization of Chinese cities. Rev. Econ. Stat. 99 (3), 435–448. McMahon, D., 2014. Rethinking the Decline of China’s Qing Dynasty: Imperial Activism
Cai, F., 2010. Chinese rural reform in 30 years: an analysis in terms of institutional and Borderland Management at the Turn of the Nineteenth Century. Asian States
economics. Front. Econ. China 5, 150–168, https://doi.org/10.1007/s11459-010- and Empires. Taylor & Francis.
0008-1. Meng, X., 2005. Introduction: poverty and labor markets in China. J. Comp. Econ. 33
Chan, K.S., Dang, V.Q., Lai, J.T., Yan, I.K., 2011. Regional capital mobility in China: (4), 641–643.
1978 to 2006. J. Int. Money Finance 30 (7), 1506–1515. Meng, X., Kidd, M.P., 1997. “Labor market reform and the changing structure of wage
Chandra, A., Thompson, E., 2000. Does public infrastructure affect economic activity? determination in China’s state sector during the 1980s. J. Comp. Econ. 25 (3),
Evidence from the rural interstate highway system? Reg. Sci. Urban Econ. 457–490. 403–421.
Donaldson, D. (Forthcoming): “Railroads of the Raj: Estimating the Impact of Meng, X., Qian, N., Yared, P., 2015. “The institutional causes of China’s great famine,
Transportation Infrastructure,” American Economic Review. 1959 –1961. Rev. Econ. Stud. 82 (4), 1568–1611.
Donaldson, D., and R. Hornbeck (forthcoming): “Railroads and American economic Michaels, G., 2008. The effect of trade on the demand for skill: evidence from the
growth: a market access approach,” Q. J. Econ.. interstate highway system. Rev. Econ. Stat. 90 (4), 683–701.
Duflo, E., 2004. The medium run effects of educational expansion: evidence from a large Munshi, K., Rosenzweig, M., 2009. Why Is Mobility in India So Low? Social Insurance,
school construction program in Indonesia. J. Dev. Econ. 74 (1), 163–197. Inequality, and Growth. Working Paper 14850. National Bureau of Economic
Duranton, G., Puga, D., 2004. Micro-foundations of urban agglomeration economies. of Research.
Handbook of Regional and Urban Economics, chap. 48. In: Henderson, J.V., Thisse, Murphey, R., 1970. The Treaty Ports and China’s Modernization: what Went Wrong?,
J.F. (Eds.), Handbook of Regional and Urban Economics, vol. 4. Elsevier, pp. Michigan Papers in Chinese Studies. University of Michigan, Center for Chinese
2063–2117. Studies.
Eckstein, A., 1977. China’s Economic Revolution. Cambridge University Press. Nickell, S., 1981. Biases in dynamic models with fixed effects. Econometrica 49 (6),
Faber, B., 2009. Trade Integration, Market Size, and Industrialization: Evidence from 1417–1426.
China’s National Trunk Highway System. Working paper. London School of Pong, D., 1973. Confucian patriotism and the destruction of the woosung railway, 1877.
Economics. Mod. Asian Stud. 7, 647–676 04.
Faber, B., 2014. Trade integration, market size, and industrialization: evidence from Sequeira, S., Nunn, N., Qian, N., 2017. Migrants and the Making of America: the Short-
chinas national trunk highway system. Rev. Econ. Stud. 8 (13), 1046–1070. and Long-Run Effects of Immigration during the Age of Mass Migration. Working
Feenstra, R.C., 2002. Border effects and the gravity equation: consistent methods for Paper 23289. National Bureau of Economic Research.
estimation. Scot. J. Polit. Econ. 49 (5), 491–506. Spence, D.J., 1990. The Search for Modern China. Norton.
Fogel, R., 1962. A quantitative approach to the study of railroads in American economic Spence, J., 1997. God’s Chinese Son: the Taiping Heavenly Kingdom of Hong Xiuquan, God’s
growth: a report of some preliminary findings. J. Econ. Hist. 22 (2), 163–197. Chinese Son: the Taiping Heavenly Kingdom of Hong Xiuquan. W.W. Norton.
Fogel, R., 1964. Railroads and American Economic Growth: Essays in Econometric Storeygard, A., 2016. Farther on down the road: transport costs, trade and urban growth
History. Johns Hopkins Press, Baltimore. in sub-saharan Africa. Rev. Econ. Stud. 83 (3), 1263–1295.
Ge, J., 2000. Zhong Guo Ren Kou Shi. Population Annals of China. Fudan University Unger, J., 2002. The transformation of Rural China, Asia and the Pacific. M.E. Sharpe.
Press, Baltimore. West, L.A., Zhao, Y., 2000. Rural Labor Flows in China. Berkeley University Press,
Glaeser, E., Gottlieb, J., 2009. The wealth of cities: agglomeration economies and spatial Berkeley, California.
equilibrium in the United States. J. Econ. Lit. 47 (4), 983–1028. Zhang, Z.-G., 2017. Study on Adjustment of Chong-Fan-Pi-Nan System in the Qing
Hsieh, C.-T., Klenow, P.J., 2009. Misallocation and manufacturing TFP in China and Dynasty. Liaoning Province Social Sciences.
India. Q. J. Econ. 124 (4), 1403–1448. Zhao, E., 1976. Qing Shi Gao. Provisional History of the Qing Dynasty. Zhonghua shu ju:
Hua, Q., 1991. Tai ping tian guo di li zhi. Taiping Heavenly Kingdom Geography. Xinhua shu dian Beijing fa xing suo, Beijing.
Nanning Shi: Guangxi ren min chu ban she. Zhao, Y., 1999. Leaving the countryside: rural to urban migration decisions in China,”
Huang, Y., 2008. Capitalism with Chinese Characteristics: Entrepreneurship and the The American Economic Review Papers and Proceedings, vol. 89, pp. 281–286.
State, 1 edn. Cambridge University Press.

36