Friends from Afar: The Taiping Rebellion, Cultural
Proximity and Primary Schooling in the Lower Yangzi,
1850-1949∗
Yu Hao‡
Peking University, School of Economics
Melanie Meng Xue§
UCLA Anderson School of Management
This Version: June 2016
Abstract
This paper tests the hypothesis that the cultural distance between migrants and natives impedes public goods provision. The Taiping Rebellion was a shock that caused
groups without a history of shared governance to be relocated into the same region.
We use a unique historical dataset of surnames in the Lower Yangzi of China to construct a measure of the cultural distance between migrants and natives (MNCD).
We find an one-standard-deviation increase in MNCD is associated with a decrease
of over 0.19 public primary schools per 10,000 persons in the early 20th century.
Results survive various robustness checks and an instrumental analysis exploiting
pre-existing cultural distances between native and the nearby population. Evidence
from the timing of MNCD taking effect, suggests that the primary mechanism runs
from migrant-native cultural distance through quality of collective decision-making
to modern primary education.
Keywords: Cultural Distance; Primary Education; Local Public Goods; Quasi-Exogenous
Migration
JEL Codes: D72, J15, N45, N95, O15, Z1
∗
We thank the editor Debin Ma, two anonymous referees, as well as Ying Bai, Zhiwu Chen, Qiang Chen,
Gregory Clark, Christian Dippel, Mark Koyama, James Kung, Nan Li, Kris Mitchener, Jean-Laurent Rosenthal,
Tuan Hwee Sng, Romain Wacziarg, Noam Yuchtman and conference participants at the All-UC Conference on
“Frontiers in Chinese Economic History” , Chinese University of Hong Kong, Shandong University, Xiamen University and Peking University for helpful comments and suggestions. All remaining errors are the responsibility
of the authors.
‡
[email protected]
§
[email protected]
I. INTRODUCTION
An extensive literature documents the negative impact of population heterogeneity on public
goods provision (Easterly and Levine, 1997; Alesina, Baqir, and Easterly, 1999; Alesina and
Ferrara, 2005). However, recent research suggests the lack of history of shared and centralized
governance between groups is just as likely to be responsible for the adverse outcomes associated
with the coexistence of different ethnic groups(Gennaioli and Rainer, 2007; Michalopoulos and
Papaioannou, 2013). This raises the question whether ethnic cleavages or artificial jurisdictions
has caused poor economic performance. Dippel (2014) contributes to the debate by showing that
a lack of a history of shared governance can negatively affect even ethnically and linguistically
homogeneous populations. We instead show that even for previously detached groups, cultural
distance can matter for the coexistence of different ethnic groups.
We exploit variation in cultural distance between previously detached groups following an external shock: the Taiping Rebellion. We use a unique dataset of Chinese surnames of approximately
100,000 individuals over the course of 150 years. The same dataset also allows us to build common measures of population heterogeneity such as fractionalization and polarization.
The Taiping Rebellion (1850-1864) was a massive civil war in South China that constituted a
one-time shock to the population makeup. The rebellion led to the loss of 17 million lives in the
Lower Yangzi, or half of the native population (Cao, 1998). After the war, migrants flocked into
the region and began to coexist with natives. This shock created two groups without a history
of shared governance or prior interaction in a region. Cultural proximity between migrants and
natives varied. We hypothesize that cultural distance between migrants and natives (“MNCD”),
who lived in the same community after the rebellion, had a negative impact on public goods
provision.
We provide the historical context to show that migration was plausibly exogenous to the cultural
distance between migrants and natives. First, migrants moved into the area with little prior
contact with natives. Migrants were not selected based on their cultural proximity with the
natives (as is the case with chain migration). Ex-ante sorting was minimum. Second, in traditional China where ancestral land was of cultural prominence, natives were not able to move out
as freely in response to the arrival of migrants whose preferences differed. Hence ex-post selfsorting was not a concern, either.1 To further establish causality, we introduce an instrumental
variable approach exploiting variation in pre-existing native-nearby cultural distance.
We go on to test our hypothesis that a greater migrant-native cultural distance lowers public
goods provision. Our proxy for public goods provision is the number of public primary schools
at county level. In the baseline model, we find that a one-standard-deviation increase in MNCD
is associated with a decrease of 0.19 public primary schools per 10,000 persons between 1900
1
Compared to Ager and Brückner (2013), we use arguably more exogenous migration as a treatment, as
natives and migrants had few opportunities to engage in ex ante screening, or ex-post self-sorting.
1
and 1910. That is a fifth of the mean of the number of public primary schools by population, or
40% of the standard devision. We then include in the controls share of arable land, distance to
the Grand Canal, distance to the Yangtze River, distance to the provincial capital, and distance
to Shanghai. We show that MNCD wins horse races against alternative explanatory variables
including the traditional fractionalization index and the polarization index. For robustness,
we control for initial conditions, interventions in education (missionary activities and temple
conversion), and confront possible effects of war (battle exposure, demographic shock and human
capital shock) on schools.
While our key finding is that cultural distance has an independent effect on public goods provision outcomes, we find evidence that the negative effects of the cultural distance between
surname groups can be mitigated by the history of shared governance. Our finding is mainly
built off the horse race results of MNCD against other measures of population heterogeneity
that ignore the history of shared governance within native surname groups.
We provide suggestive evidence on the mechanisms through which MNCD prevented the establishment of public primary schools. First, we exploit institutional features of early 20th century
China to form a testable hypothesis : MNCD should have the strongest effect on lower-primary
and dual-primary schools, since (a.) MNCD should matter the most in an environment of selfgovernance, (b.) villages were traditionally self-governed, and (c.) villages (and townships) were
responsible for the building of lower-primary schools, and sometimes, dual-primary schools.2
Consistent with our prediction, we find that MNCD only affects schools at lower-primary and
dual-primary schools, not at upper-primary and secondary schools. Second, we exploit time
variation in institutions through the first half of the 20th century. Massive institutional changes
in 20th century China provides an excellent laboratory to observe the impact of MNCD. In our
sample, we have periods of decentralization and centralization of the education system, when
decisions to education children were made locally and when those decisions were made by the
national or provincial government, respectively. This allows us to use evidence from timing to
interpret our finding. We expect to see a larger effect of MNCD during the period featuring
more self-governance and more decentralization of fiscal authority. And consistent with this
prediction, we find the effect of MNCD on modern education is pronounced in the early 20th
century but is muted in China for much of the 20th century under autocratic rule and fiscal
centralization. Soon after the centralization of the educational system in 1927, MNCD no longer
had a significant effect on schools. We conclude that MNCD resulted in fewer primary schools
being build due to lower quality of collective decision-making in local communities.
Our study builds on the literature on the relationship between diversity of individual preferences
and public goods provision. Alesina, Baqir, and Easterly (1999) show theoretically that the
2
At the time, the entire phase of primary education was divided into two: upper- and lower-primary education.“Upper primary education” refers to the more advanced stage of primary education. Most schools specialized
in either upper- or lower-primary education. Those providing both upper and lower primary education were called
“dual-primary schools”.
2
median distance from the preference of the median voter can be considered as an indication of
how polarized preferences are. The model predicts that public goods provision will be adversely
affected in a polarized society characterized by two separate groups with relatively homogeneous
preferences within the group, but very distinct preferences across groups. More recent work
shows that in the process of decentralization and redistricting, the benefits of reduced diversity
can be undone if the newly governed population is highly polarized (Bazzi et al., 2015). In our
paper, we use the cultural distance between migrants and natives as a proxy for the difference
in preferences between these two groups. We find that cultural distances between groups indeed
matter for public goods provision, whereas the traditional fragmentation measure that assigns
the same distance to all groups does not produce the same effects.
Our study also contributes to the literature of the effect of genetic dissimilarity on economic
development. Ashraf and Galor (2013) find the beneficial and detrimental effects of diversity on
productivity, and conclude that an immediate level of diversity is the most conducive for economic development. Desmet, Le Breton, Ortuño-Ortı́n, and Weber (2011) link genetic distance
to the stability and breakup of nations, and provides empirical support for the use of genetic distance as a proxy of cultural heterogeneity. Spolaore and Wacziarg (2009) show genetic distance
affects income differences across countries through a barrier effect to the diffusion of development from the world technological frontier. Our paper similarly uses genetic distance as a proxy
for cultural distance, and focuses on the public goods provision consequences of greater genetic
and cultural distances between groups.
This paper is organized as follows. Section II explains the historical context. Section III discusses
data sources and the basis for constructing our measure of migrant-native cultural distance. Section IV summarizes my baseline results and the comparison of migrant-native cultural distance
to the fractionalization and polarization of the population. Section V introduces a number of
robustness checks, accounting for initial conditions, interventions in education and war-related
conditions. Section VI comprises an instrumental variable analysis. Section VII identifies the
quality of collective decision-making as a possible channel for migrant-native cultural distance
to influence the supply of modern primary education. Section VIII concludes the paper.
II. HISTORICAL CONTEXT
A The Taiping Rebellion
The Taiping Rebellion was a massive civil war in South China which lasted from 1850 to 1864.
At least 17 million people, or half of the populace, died in the lower Yangzi (Cao, 1998; Cao and
Li, 2000). Battles broke out throughout the Lower Yangzi and all counties, with the exception
of Shanghai, were occupied for at least 3 months. The area around Nanjing, the capital city
of Taiping regime since 1853, had lingering conflicts for over ten years. The most prosperous
and important cities in the Lower Yangzi, Hangzhou and Suzhou were occupied by the Taiping
3
army after 1860. Shanghai, protected by foreign powers, was the least affected, and it served
as a shelter for over 200,000 refugees (Ge, 2002a, pp.62–63). Famine and plague followed the
battles. So did mass migration.
A.1
In-Flow Migration
Migration occurred both during the Taiping Rebellion itself and in the aftermath of the rebellion.
While migration internal to the Lower Yangzi was certainly common, post-Taiping migration
was best characterized by long-distance migration from North China and from the Middle Yangzi
River. A crucial difference between pre- and post-Taiping migration is that pre-Taiping migration was largely driven by income differences, job opportunities, and based on ethnic bonds
and geographic proximity (Li, 2011), whereas post-Taping rebellion migration originated from
a very wide range of geographic areas and featured diverse economic and cultural backgrounds.
Another difference is the scale and pace of migration—post-Taiping migration was far more
rapid and broader in scale. For this reason, in this paper we focus on post-Taiping migration.3
The mass migration led to conflicts between natives and migrants, and between different migrant
groups. In villages and townships conflicts arose as a result of clashing preferences and interests,
different dialects, skills and social customs. Conflicts, as documented in local gazetteers, took
place over a wide range of issues such as usage of public water, property rights of ownerless land
and eligibility for imperial exams (Ge, 2002a, pp. 303-308).
A.2
The Economic and Political Consequences of the Taiping Rebellion
The Taiping Rebellion constituted a multi-dimensional shock to the region. Most likely, it had
more than one way to affect primary schools. Those effects could be at play on both the supply
and demand side of education. In Section V.C, we provide a quantitative analysis of how various
outcomes of the Taiping Rebellion might have affected the building of primary schools in the
early 20th century.
The rebellion damaged local infrastructure. In the Jiaxin prefecture of Zhejiang, 21% of Buddhist and Tao temples were destroyed by rebels affiliated with the God Worshiping Cult (Li,
2002). County public schools and private schools, where lower degree holders received further instruction to prepare for the higher level exams, were also destroyed or damaged in large numbers.
The destruction of local infrastructure could have arguably undermined the resources useful to
the launch of modern schools fifty years after the rebellion. That said, it is widely documented
that most temples and schools were restored shortly and even more were built in the late 19th
century. For example, Li (2002) found that 98 temples were destroyed by the rebels but 220
were built (or restored) within twenty years after the rebellion because living standards was
3
The provincial governments advertised all around China for migrants and depicted the Lower Yangzi as
a ‘kingdom of free land; and the ‘land of opportunities’. Farmers from Henan, Anhui, Hubei, Hunan, North
Jiangsu, and South Zhejiang came for a better living (Ge, 2002a, pp.100-106). After 1900, industrialization drew
massive immigrants into the urban area of Shanghai (Junya and Wright, 2010). Its population increased by
four-fold from 1907 to 1947.
4
even better than before the war and trade and commercial network was quickly restored. Kuhn
(2002), to the contrary, interpreted this trend (along with the rise of local charity organization)
as the rise of local gentry at provincial and county level overseeing local public affairs, which
eventually led to the formalization of local self-governance in the early 20th century.
The rebellion dismantled kinship networks. Clans used to provide financial aid for clan members
to receive education. During the rebellion rich families migrated to the urban area with their
less well-off relatives left behind in the countryside. Clans also lost land property to the war, the
rent from which were assigned as public funds for supporting education (Li, 1981). As discussed
by Xu and Yao (2015), kinship networks are an alternative to formal institutions in providing
public goods by effectively overcoming free-riding problems. However, in the context of education
reform in the 1900s, strong kinship networks can be a double-edged sword. Clans sometimes
would prefer the option of funding informal and private tutoring exclusively enjoyed by clan
members to establishing a school accessible by both clan and non-clan members. A weaker
kinship network, in that case, may have reduced within-kinship public goods but enhanced
cross-kinship public goods.
The rebellion led to huge population losses, which induced higher land-labor ratios and higher
wages (Cao and Chen, 2002). High wages forced war-stricken areas to abandon subsistence
agriculture and switch to labor-saving technologies and industries. In Wujin and Wuxi, the
silk industry superseded rice farming to be the largest employer in the rural area (Mickey and
Shiroyama, 2009). Lin and Li (2014) show that areas with a larger impact of war saw more
modern industrial enterprises in the late 19th century and had a higher level of urbanization
in the 1930s. In addition, the rebellion inadvertently created political room for institutions in
favor of modernization to set roots. Pro-reform officials were assigned to post-war provinces.
They established formal institutions to promote industrialization.
B Educational Reforms: From Traditional to Modern Education
Fifty years after the Taiping Rebellion, Qing Government put forward an educational reform.
The abolition of the imperial exam system went hand in hand with the attempt to establish a
western-style, modern school system.
Prior to 1905, education focused on Confucian classics and aimed at preparing students for the
imperial examinations. The traditional educational system included two stages: mass primary
education aimed for basic literacy and talent spotting, and more advanced education that drilled
candidates selected from the first stage to pass the exams (Leung, 1994). In the late 19th century
growing economic openness gave rise to higher demand for education in science, technology, and
other non-exam skills (Yuchtman, 2015). Attempts by missionaries to build modern schools
began in some coastal cities as early as the 1860s. But only until the abolishment of the
exam system in 1905 did modern education begin to expand. The Ministry of Education was
established, and Offices of Provincial Education was founded, along with county-level agencies
known as “Education Exhorting Offices” (quan xue suo).
5
Educational reform was not a smooth process. Despite ambitious political and educational
reforms, few things changed on the ground. For villages and townships, the process of building
modern schools was slow and painful. County officials often found it difficult to raise county
taxes, and make within-county transfers to ensure universal primary schools. Clans sometimes
would prefer to provide direct financial aid to clan members for them to take cheaper informal
and private tutorships, rather than establish a school open to both clan and non-clan members.
More details concerning both the institutional features of the traditional exam system and of
the modern education system are included in the appendix (Appendix D and Appendix E).
III. DATA AND MEASUREMENT
Table A-1 provides summary statistics of all the variables used in the paper and their data
sources. Below we focus on the underlying logic of our independent variable, migrant-native
cultural distance.
A Independent Variable: Migrant-Native Cultural Distance
Our independent variable is the cultural distance between migrants and natives. We rely on
surname data to construct our measure. To be specific, we use differences in the surname mix
to proxy for the cultural distance between migrants and natives:
MNCDi =
where normalized isonomyN,M N,i =
1
,
normalized isonomyN,M N,i
PS
P
P
qP k k,N,iPk,M N,i
.
S 2
S 2
P
k k,N,i
k Pk,M N,i
(1)
S is the number of surnames in the two
groups. Pk,N,i and Pk,M N,i are the relative frequencies of surname k within natives and within the
entire population including natives and migrants.4 The isonomy between the native population
P
and the entire population, Sk Pk,N,i Pk,M N,i , measures how likely any individual randomly drawn
from within natives bears the same surname as one drawn from within the entire population.
P 2
, and the isonomy of the
We normalize it with the isonomy of the native population, Sk Pk,N,i
PS 2
entire population, k Pk,M N,i . MNCD captures how culturally dissimilar natives and migrants
were. Figure 1 illustrates migrant-native cultural distance in the Lower Yangzi.
Our approach is in line with Bai and Kung (2011, 2015); Li (2011); Spolaore and Wacziarg
(2009). Li (2011) uses surname distances between pair of countries or regions at a given time to
measure multilateral genetic and cultural distance. Following Du et al. (1997), Bai and Kung
(2011) and Bai and Kung (2015) use isonomy (similarity in surname distribution between any
4
In practice, we extract migrant-native cultural distance from the distance in the surname distribution of
a county’s population before and after the rebellion, with the assumption that the surname mix of the natives
remained relatively stable during that period. It is clear that the total population of natives declined after the
rebellions, but as long as the proportion of each surname population to the total population did not change, our
assumption remains valid.
6
Figure 1: Migrant-Native Cultural Distance
pair of population) to approximate genetic and cultural distance across regions. They show that
this surname-based measure is strongly correlated to a measure of genetic distance based on the
frequency distribution of the A, B and O alleles of the ABO gene at the province level. They
also find that surname-based measure is strongly correlated to a measure of cultural distance
based on dialects. Similar to Spolaore and Wacziarg (2009), they find that the smaller genetic
or cultural distance to the technological frontier from a region, the faster the technology diffused
to that region and hence faster growth. In this paper, we adopt the same measure as in Bai
and Kung (2011); but instead of using isonomy between two regions, we use isonomy between
before-migration population (natives) and after-migration population (natives and migrants) to
proxy for the consanguinity between natives and migrants.
We do not have data on which surnames were associated with migrants but it is generally
the case that migrants as a group were different to natives in terms of surname distribution,
especially when they came from far away. To obtain the surname distribution of a county before
migration, we hand-collect data from county chronicles (a.) name lists of civilians who died
during the rebellion 1851-1865, (b.) exam degree holders (1645-1850) (c.), surnames of chaste
women, (d.) surnames of their husbands if they are also recorded. The number of records for
7
each county ranges from 800 to 3000. To obtain the surname distribution of a county after
migration, we use the following sources: (a.) surnames of dead soldiers (1927-1953) and (b.)
college students born in that county who graduated between 1900 and 1949. The number of
surnames from the above sources ranges from 500 to 2500 per county. More details can be found
in Appendix B.
We draw surnames of individuals from a wide range of social backgrounds to improve the representativeness of the surname sample. One concern might be that these samples seem too small
to correctly estimate the true surname distribution of population if each surname accounts for a
small fraction of population. However, this is not the case for China. In each county the largest
few surnames each accounts for more than 5% of the population. To correctly estimate surname
distribution of population at least for the largest few surnames, one needs a population sample
of as small as a few hundred. For the six counties we indeed have a small sample problem, we do
not include them in most of our regressions. As counties with a small sample are not random, we
impute our outcome variable for those six counties, and run regressions on the sample including
them as well.
One way to check the validity of our measure is to cross-check with measures reflecting cultural
or ethnolingustic devisions. One indicator of cultural devisions in the Lower Yangzi is linguistic
enclaves. Before the rebellion, almost the entire lower Yangzi spoke Wu. After the rebellion,
migrants settled in this area, giving rise to thousands of Mandarin-speaking villages and communities (Huang, 2004; Cao, 2006; Simmons et al., 2006). In Figure 2, we mark counties with
linguistic enclaves where Mandarin (guan hua) is used. As shown on the map, counties with
some of the highest values of MNCD harbor linguistic enclaves even today. This enhances our
confidence in the validity of our measure.
Ideally, we would also like to construct a weighted fragmentation index that takes into account
whether individuals with the same surname to be both natives, both migrants, or one native
and one migrant. The fractionalization of the population comes down to dissimilarity of options
preferred by each other. Rather than to simply stipulate that options are either similar or
dissimilar, (Bossert et al., 2003) propose alternative frameworks that permit more degrees of
similarity between options. A generalized index of fractionalization is described in (Bossert
et al., 2011). Drawing on the insights from Bossert et al. (2003), Alesina and Ferrara (2005)
and Caselli and Coleman (2013), we can assume the distance between two individuals with the
same surname, if one is a migrant and the other is a native, to be positive; and the distance
between two individuals with the same surname, both being natives or being migrants, to be
zero. Unfortunately, without individual-level or surname-level data on migrant-native status,
our ability to operationalize this index with our data is limited.
B Control Variables
To account for other factors shaping modern education, we include in the baseline controls primary schools rate in 1880, population size and urbanization rate. Culturally dissimilar migrants
8
Figure 2: Mandarin Linguistic Enclaves in the Lower Yangzi (Wu-Speaking) Sources: Cao
(2006)
might be selected into places with economic conditions that are not in favor of public goods
provision and schooling. So in the full set of controls, we include ruggedness, share of arable
land, agricultural suitability, distance to the Yangtze River, distance to the Grand Canal, distance to the provincial capital and distance to Shanghai to capture other differences in economic
conditions between counties. For robustness, we account for initial conditions—the number of
charitable organizations by 1850, population and population density in 1820, and measure of the
impact of war—battle exposure (months), % of elderly and youth (under 20 or over 40), % of
adult men (between 20 and 40) and a measure of differences in human capital between migrants
and natives which we call the “human capital shock”. We infer that natives who were able to set
up more charitable organizations have higher social capital, which is likely correlated with both
the type of migrants they admit and retain, as well as with their own ability to provide public
education. We also discuss other potential shocks to education, such as missionary activities,
measured by the log of one plus communicants per 10,000, and temple conversion, measured
by the log of one plus number of temple-converted schools. We expect both variables to be
positively associated with our dependent variable.
IV. BASELINE RESULTS
We use an OLS model to estimate the impact of migrant-native cultural distance on the number
9
of public primary schools:
#public primary schools per 10,000 personsi = α + βMNCDi + Xi Ω + i .
(2)
The dependent variable #public primary schools per 10,000 persons is the number of public
primary schools per 10,000 persons right after the educational reform.5 MNCDi is the migrantnative cultural distance in County i. MNCD exclusively focuses on the cultural distance between
migrants and natives—two groups with no history of shared governance.6 Xi Ω are a vector of
county-level controls. i is a disturbance term.
Table 1 summarizes estimates of the effect of migrant-native cultural distance on public primary
schools. With all controls, a one-standard-deviation of MNCD reduces the number of public
primary schools by approximately 0.18 school per 10,000 persons, which is equal to a fifth of
the mean or 40% of the standard deviation. We show an unconditional regression of MNCD on
primary schools during the decade of 1900-1910 in Column 1. The relationship is both negative
and significant. In Columns 2 and 3 we add population and urbanization rate sequentially. In
Alesina et al. (1999), a larger population means greater economy of scale to provide public goods,
but a higher transaction cost in raising taxes. Urbanization may enhance the economic return
of education, affecting the demand side of education. In Column 4, we include basic education
access in 1880. We interpret this as a measure of the stock of human capital in an area, and as a
summary statistic of those slow-moving components in local culture and institutions that shape
the decision to receive education in the long run. By isolating the influence of past educational
achievement under the private education system, we are one step closer to focusing on the impact
of MNCD on public goods provision. From Column 5 to Column 10, we introduce geographic
controls such as ruggedness, share of arable land and agricultural suitability, distance to the
Grand Canal, distance to the Yangzi River, distance to Shanghai, and distance to the provincial
capital.7 Share of arable land and agricultural suitability can proxy the opportunity cost of
receiving modern education. Distance to the Grand Canal and distance to the Yangzi River are
used to proxy the market potential and access to trade. Distance to the provincial capital is
included to account for the reach of provincial government or state capacity. Shanghai, which
became a treaty port as early as 1845, was exposed to rapid modernization and industrialization.
We control for distance to Shanghai to account for the spillover effects of Shanghai on the rest of
the region. The coefficient estimate of MNCD in Column 1 (-0.21) with no control is of similar
magnitude to that in Column 10 (-0.18) with all controls, when there is a sizable increase in R2
(from 0.174 to 0.399).
5
Data are from the 1907, 1908 and 1909 Census. Those are the only censuses in the decade of 1900 that
contain information on schools.
6
Dippel (2014) stresses the role of the history of shared governance in the discussion of ethnically and
linguistically fragmented jurisdictions having poorer economic performances.
7
Nanjing for Jiangsu, and Hangzhou for Zhejiang.
10
11
54
0.174
54
0.267
-0.272∗∗∗
(0.052)
-0.237∗∗∗
(0.086)
-0.219∗∗∗
(0.051)
54
0.253
54
0.238
-0.271∗∗∗
(0.058)
-0.245∗∗∗
(0.085)
0.000
(0.003)
0.035
(0.264)
54
0.225
-0.272∗∗∗
(0.059)
-0.252∗∗∗
(0.087)
0.000
(0.002)
0.023
(0.272)
-0.013
(0.040)
54
0.242
-0.255∗∗∗
(0.061)
-0.228∗∗
(0.093)
0.001
(0.002)
-0.056
(0.264)
0.010
(0.064)
0.588
(0.358)
0.014
(0.045)
54
0.231
-0.246∗∗∗
(0.067)
-0.203∗
(0.107)
0.001
(0.003)
-0.003
(0.310)
0.026
(0.077)
0.698∗
(0.407)
0.008
(0.047)
0.196
(0.328)
54
0.271
-0.192∗∗∗
(0.060)
-0.185
(0.113)
-0.000
(0.003)
-0.080
(0.296)
-0.002
(0.078)
0.424
(0.430)
-0.036
(0.041)
0.021
(0.361)
0.246∗
(0.130)
54
0.305
-0.161∗∗∗
(0.059)
-0.197∗
(0.107)
-0.002
(0.003)
0.018
(0.289)
0.023
(0.084)
0.132
(0.502)
-0.039
(0.044)
-0.067
(0.369)
0.205∗
(0.121)
-0.002∗
(0.001)
(4)
(5)
(6)
(7)
(8)
(9)
Dependent variable: #public primary schools per 10,000 persons
-0.272∗∗∗
(0.052)
-0.240∗∗
(0.091)
0.000
(0.003)
(3)
54
0.399
-0.180∗∗∗
(0.051)
-0.333∗∗∗
(0.108)
0.003
(0.003)
0.076
(0.263)
0.016
(0.068)
0.483
(0.428)
-0.006
(0.039)
-0.594
(0.400)
0.144
(0.119)
0.002
(0.001)
0.004∗∗∗
(0.001)
(10)
(11)
60
0.407
-0.185∗∗∗
(0.041)
-0.347∗∗∗
(0.092)
0.003
(0.003)
0.244
(0.260)
0.041
(0.041)
0.535
(0.387)
-0.004
(0.037)
-0.469
(0.347)
0.099
(0.099)
0.002
(0.001)
0.004∗∗∗
(0.001)
Notes: The table reports the impact of migrant-native cultural distance on the number of public primary schools in the decade of
1900-1910. The unit of analysis is a county in Republican China. All baseline controls are included in Column 4: the natural log
of population, urbanization rate and the natural log of primary school enrollment in 1880. Both baseline controls and geographic
controls are included in Column 10. Column 11 includes six additional observations in which MNCD is imputed. Robust standard
errors are included in all specifications. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Observations
Adjusted R2
Dist. to provincial capital
Dist. to Shanghai
Dist. to Yangtze
Dist. to Grand Canal
Agricultural suitability
%Arable land
Ruggedness
Log primary enrollment 1880
Urbanization
Log population
MNCD
(2)
(1)
Table 1: The Impact of of Migrant-Native Cultural Distance: Main Specification
This assures us that selection on unobservables are likely limited.8 In Column 11 we add six
counties with imputed values of MNCD to the sample. Our coefficients of interest remain stable
throughout the columns. The coefficient estimate in Column 11 is somewhat smaller. We
attribute this to the addition of six counties with imputed values of MNCD resulting in an
increase in measurement error. We use Column 4 as our baseline for the rest of the paper.
We next show how MNCD compares to measures of population heterogeneity. We use the
same surname data to construct the traditional fractionalization index and two polarization
indices (See Appendix G). In Table 2, we estimate the effects of MNCD, fractionalization and
polarization on public primary schools. Our main finding is that MNCD has a strong and
positive effect on public primary schools, and neither fractionalization nor polarization in the
entire population has a statistically significant effect.
The traditional fractionalization index is positively correlated with the number of public primary
schools per 10,000 persons (Col. 1-3), once MNCD is controlled for. This suggests that once
partialling out the effects of the cultural distance between migrants and natives, the fractionalization of the population does not necessarily reduce public goods provision. The coefficient
estimates of the polarization indices are negative across the columns (Col. 4-9). By comparing
MNCD to the polarization index adjusting for intergroup distances between surname groups
(Col. 7-9), we find that MNCD can reduce both the statistical significance and magnitude of
the coefficient estimate of the distance-based polarization index. From Column 7 to Column 8,
the coefficient estimate of polarization of the entire population moves from -0.165 (the p-value
is 0.138) to -0.118 (the p-value is 0.269).
In face of other measures of population heterogeneity, the effects of MNCD largely remain.
This first suggests that MNCD likely captures an important component distinct from what is
encapsulated in the fractionalization and polarization indices. Secondly, MNCD likely shares a
few traits in common with polarization. And they both have negative effects on public primary
schools. Third, I can also infer that the history of shared governance mitigates the effects of
cultural distance on local public goods provision. Likely by taking into account the history of
shared governance, MNCD outperforms polarization in explaining outcomes of public primary
schools. Our interpretation is that the cultural distance between previously detached groups is
a more important type of cultural distance. Unfortunately, given the nature of the betweensurname-group “distance” measure we use (see Appendix G), we can only draw a tentative
conclusion about the mitigating role of the history of shared governance on the negative effects
of the cultural distance between groups.
8
Oster (2014) formalizes this approach. The results of this analysis suggest that the ratio of selection on
unobservables relative to selection on observables has to be five times larger to explain away my results. Based on
the reasoning outlined by Altonji et al. (2005) that unobservables should not be more important than observables
in explaining the treatment, it is highly unlikely that unobservables are biasing my results.
12
Table 2: MNCD, Fractionalization and Polarization
(1)
MNCD
FRAC
0.013
(0.012)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Dependent variable: #public primary schools per 10,000 persons
-0.214∗∗∗
(0.056)
0.023∗∗
(0.011)
Native FRAC
-0.215∗∗∗
(0.059)
0.023∗
(0.012)
-0.001
(0.015)
POL
-1.621
(2.851)
-0.206∗∗∗
(0.056)
-0.173∗∗
(0.072)
-3.722
(2.664)
-2.196
(3.539)
-1.193
(1.716)
Native POL
POL DIST
-0.163∗∗∗
(0.052)
-0.175∗∗
(0.066)
-0.165
(0.109)
-0.118
(0.106)
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
-0.091
(0.155)
-0.114
(0.386)
Y
Y
54
0.318
54
0.435
54
0.421
54
0.306
54
0.412
54
0.405
54
0.341
54
0.405
54
0.392
Native POL DIST
Baseline controls
Geographic controls
Observations
Adjusted R2
(9)
Notes: The table reports OLS results of the impact of fractionalization and polarization on
public primary schools. The unit of observation is a county in Republican China. Robust
standard errors are used in all specifications. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
V. ROBUSTNESS CHECKS
A Subsamples
We first check if our results are robust to the omission of outliers. Table 3 provides evidence
that outliers do not have a substantial impact on our results. In Column 2 we exclude counties
with extreme values of MNCD. Column 3 excludes counties with extremely large human capital
shocks from the sample. Column 4 drops Shanghai. Coefficients in Columns 3 and 4 are fairly
comparable to those in the baseline. The coefficient of MNCD is slightly larger in Column 2.
This indicates possible attenuation bias due to measurement errors in MNCD. Our measurement
of MNCD may be particularly poor for those outliers, and MNCD may become more sensitive
to sampling errors in the range of large values. In such cases, removing those observations with
mis-measured values should improve our estimation.
B Initial Conditions
Public goods provision depends on the level of social capital. At the same time, higher social
capital within natives could imply more exclusivity, which could result in fewer migrants, and
especially, fewer migrants who disagree with values possessed by natives. To check if social
capital drives both MNCD and public primary schools, we explicitly include initial social capital
13
Table 3: Robustness: Subsamples
(1)
(2)
(3)
(4)
Dependent variable: #public primary schools per 10,000
-0.180∗∗∗
(0.051)
Y
-0.201∗∗
(0.074)
Y
-0.191∗∗∗
(0.056)
Y
-0.180∗∗∗
(0.051)
Y
Subsamples
Full sample
MNCD outliers
Human capital outliers
Shanghai
Observations
Adjusted R2
54
0.399
52
0.374
51
0.371
53
0.394
MNCD
Baseline controls
Notes: The table reports the impact of migrant-native cultural distance on public primary
schools during the decade of 1900-1910 on various subsamples. The unit of observation
is a county in Republican China. Column 1 provides benchmark results from Column 10
of Table 1. Column 2 excludes three counties with the highest MNCD (MNCD>5.25).
Column 3 excludes counties with the largest decline in human capital (Human capital-after
minus Human capital-before is less than -0.1). Column 4 excludes the important treaty port
city—Shanghai. Robust standard errors are used in all columns. ∗ p < 0.10, ∗∗ p < 0.05,
∗∗∗ p < 0.01
Table 4: Initial Conditions
Dependent variable: #public primary schools per 10,000 persons
(1)
(2)
(3)
(4)
(5)
MNCD
-0.180∗∗∗
(0.051)
#initial charities
-0.181∗∗∗
(0.052)
0.003
(0.006)
-0.201∗∗∗
(0.055)
Initial pop. density
Observations
Adjusted R2
0.119
(0.259)
-0.242∗∗∗
(0.063)
0.005
(0.006)
0.887∗∗
(0.329)
-0.351
(0.278)
0.608∗∗
(0.239)
Initial population
Baseline controls
Geographic controls
-0.170∗∗∗
(0.053)
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
54
0.399
54
0.385
54
0.461
54
0.388
54
0.458
Notes: The table reports reports OLS results of the impact of MNCD on public primary
schools, accounting for initial conditions. The unit of observation is a county in Republican
China. Initial # charities are number of charities before the Taiping Rebellion. Initial population and initial population density are population and population density in 1820. All
specifications include baseline controls and geographic controls. Baseline controls include
logged primary enrollment rate in 1880, urbanization rate and logged population. Geographic controls include ruggedness, share of arable land, agricultural suitability, distance
to the Grand Canal, distance to the Yangtze River, distance to the provincial capital and
distance to Shanghai. Robust standard errors are used in all specifications. ∗ p < 0.10, ∗∗
p < 0.05, ∗∗∗ p < 0.01
14
among the natives in our regression, proxied by number of charitable organizations in 1850.9 To
overcome endogeneity, we use the number of charitable organizations before the mass migration.
In addition, initial population size can contain crucial information about the socioeconomic
traits of the natives. So we also include initial population size or initial population density in
our regression, and find that initial population size is positively correlated with the number of
public primary schools per 10,000 in the 1900s.
C War-Related Conditions
The Taiping Rebellion was undoubtedly a major blow to the region. Other than bringing
in migrants with cultural dissimilarity with natives, the rebellion likely had an independent
effect on schools by affecting infrastructure, income and the demographics. Below we carefully
evaluate the overall impact of war on primary schools, as well as the specific effects of the
Taiping Rebellion on the human capital stock and the demographic profile of a county. We
show coefficient estimates of MNCD partialling out the other effects of war, as well as coefficient
estimates of those variables without MNCD. In the next section, we propose an instrumental
variable strategy to establish the causality going from cultural distance to primary schools.
Battle Exposure As mentioned in Section II.A, war could have lingering consequences several
decades after. The scale of damage can be a function of battle exposure. We measure battle
exposure by the number of months a county was exposed to war. Overall, we find no evidence
that battle exposure had an impact on the number of public primary schools fifty years later
(Col.2 & 3).
Demographic Shock War and migration could have affected schools by altering the demographic structure. Population tend to rebound rapidly after major disturbances (Davis and
Weinstein, 2002), resulting in a youthful population. Migrants tend to be younger, male. If a
larger number of migrants flock into a county, who are systematically younger, from the supply
side, education can become more affordable in the short run, due to lower dependency ratio, but
from the demand side, there could be less demand for education if opportunities for unskilled
work are abundant. We include both the %elderly and youth and %adult men (aged 20 to 40)
9
During the period we study, charitable organizations were mostly funded by local communities.
15
16
54
0.399
54
0.306
54
0.392
54
0.297
54
0.387
N
N
Y
Y
54
0.300
54
0.385
N
N
Y
Y
0.315
(1.209)
-0.182∗∗∗
(0.053)
54
0.365
N
N
Y
Y
-0.205∗∗∗
(0.067)
0.008
(0.012)
-0.046
(0.037)
0.007
(0.031)
0.261
(1.350)
54
0.516
Y
Y
Y
Y
-0.264∗∗∗
(0.077)
0.003
(0.011)
-0.039
(0.033)
0.012
(0.027)
1.046
(1.365)
(9)
Notes: The table reports reports OLS results of the impact of war on public primary schools. The unit of observation is a county in
Republican China. Battle exposure is measured by month. % elderly and youth refers to those under 20 or over 40. % adult men
refers to men between 20 and 40. Human capital shock is the difference in human capital between natives and the entire population
including natives and migrants. All specifications include baseline controls and geographic conrtrols. Robust standard errors are
used in all specifications. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Observations
Adjusted R2
N
N
Y
Y
-0.045
(0.039)
0.004
(0.028)
N
N
Y
Y
N
N
Y
Y
-0.018
(0.034)
-0.023
(0.026)
-0.200∗∗∗
(0.067)
Initial conditions
Inventions in education
Baseline controls
Geographic controls
N
N
Y
Y
0.007
(0.012)
-0.181∗∗∗
(0.052)
0.008
(0.011)
-0.006
(1.250)
N
N
Y
Y
-0.180∗∗∗
(0.051)
Dependent variable: #public primary schools per 10,000 persons
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Human capital shock
%adult men
%elderly and youth
Battle exposure
MNCD
(1)
Table 5: War-Related Conditions
in Columns 4 and 5, and we do not find a statistically significant effect of either variable.
Human Capital Shock Human capital could be another channel through which the Taiping
Rebellion affected the long-run prospects of a county. Even though migrants and natives had
little contact before the settlement of migrants, some selection could be on the amount of human
capital migrants had. In the event that migrants simultaneously posed a cultural shock and a
human capital shock to the native population, the human capital shock may be a confounder.
We define a human capital shock as the difference between migrant and native human capital
(See Appendix H for variable construction details). When the migrant human capital stock is
higher than native human capital, we conclude there was a positive human capital shock. In
Columns 6 and 7, we control for the difference between migrant and native human capital. We
find no effect of differences in human capital on schools. However, once MNCD is added to
the regression, human capital shock attains a positive coefficient which is consistent with our
prior. The advantage of having migrants with high human capital on school formation is likely
partially offset by the cultural distance between migrants and natives.
D Interventions in Education
Missionary Activities Christian missionaries came to China in the 19th century to spread
Christianity. In the process, they also built a large number of schools. Compared to secondary
education, however, primary education was not affected to the same degree by local missionaries
(Bai and Kung, 2014). Demand for public or semi-public modern primary education, may still
have been affected by long-standing missionary activities. For example, if church-sponsored
secondary schools already existed in the area before the campaign for modern education, it
might increase demand for modern primary education through a complementarity mechanism.
Due to the earliest missionary expansion and the Taiping Rebellion being concurrent events, we
cannot rule out the possibility that a hostile native-migrant relationship could deter missionaries,
or that migrants are more likely to settle in areas with missionaries, for aid and help. We include
number of communicants per 10,000 as a rough proxy for missionary activities. As predicted,
the density of communicants has a positive effect on public primary schools, but the effect is
not statistically significant (Col. 2). The coefficient estimate of MNCD remains very similar to
baseline estimates.
Temple Conversion In the first few decades of the 20th century, government closed tens
of thousands of Buddhist or Taoist temples and took over temple assets to support modern
education. Two million Buddhist and Taoist temples are estimated to have closed in late Qing.
At the time, they owned about 16 million houses, 13,000 square kilometers of land and millions
teal of silver altogether (Xu, 2010). In Column 3, temple conversion is positively associated
with the density of public primary schools. The coefficient estimate of MNCD slightly declines,
suggesting that temple conversion might be a channel for MNCD to impede the establishment
of public primary schools.
17
Table 6: Interventions: Missionary Activities and Temple Conversion
Dependent variable: #public primary schools per 10,000 persons
(1)
(2)
(3)
(4)
(5)
MNCD
-0.180∗∗∗
(0.051)
Missionary activities
-0.178∗∗∗
(0.047)
0.148
(0.095)
N
Y
Y
N
Y
Y
0.183∗∗
(0.085)
N
Y
Y
54
0.399
54
0.443
54
0.458
Temple conversion
Initial conditions
Baseline controls
Geographic controls
Observations
Adjusted R2
-0.168∗∗∗
(0.053)
-0.167∗∗∗
(0.053)
0.134
(0.085)
0.169∗
(0.085)
N
Y
Y
-0.220∗∗∗
(0.068)
0.107
(0.082)
0.173∗∗
(0.084)
Y
Y
Y
54
0.493
54
0.543
Notes: The table reports OLS results of the impact of MNCD on public primary schools,
accounting for interventions in education. The unit of observation is a county in Republican China. “Missionary activities” is measured by the log of number of communicants plus
one. “Temple conversation” is measured by the log of number of one plus temple-converted
schools. Baseline controls include logged primary enrollment rate in 1880, urbanization
rate and logged population. Geographic controls include ruggedness, share of arable land,
agricultural suitability, distance to the Grand Canal, distance to the Yangtze River, distance to the provincial capital and distance to Shanghai. Robust standard errors are used
in all specifications. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
VI. INSTRUMENTAL VARIABLE STRATEGY
A Construction of the Instrument
We have provided evidence that “selective migration” was as unlikely as was self-sorting after
the arrival of the migrants. “Selective migration” refers to migrants and natives make decisions
on entry and acceptance based on their cultural proximity. Nevertheless, the observed migrantnative cultural distance might still be correlated with unobserved characteristics of the native
or migrant population, or with unobserved characteristics of the county of destination or county
of origin. To sidestep omitted variable bias, we introduce an instrumental variable based on the
pre-existing cultural distance between the native population and the population just outside of
the Lower Yangzi (“nearby population”).10
We use the following equation to construct a surname-based measure of the pre-existing cultural
distance between the native population and the nearby population:
Pre-existing native-nearby culture distancei =
10
1
,
normalized isonomynative,nearby,i
(3)
The nearby population refers to residents in Zhejiang and Jiangsu Province excluding the Lower Yangzi.
18
19
(4)
IV
(5)
IV
(6)
IV
-0.271∗∗∗
(0.058)
0.238
-0.384∗∗
(0.183)
0.190
-0.354∗∗
(0.173)
0.305
-0.324∗∗
(0.155)
0.391
Second Stage
#public primary schools per 10,000 persons
(3)
OLS
-0.323∗∗
(0.156)
0.376
(7)
IV
N
N
N
N
N
N
N
N
N
N
N
N
N
N
54
Human capital shock
Battle exposure
Fractionalization
Initial social capital
Pop. 1820
Georgraphic controls
Baseline controls
Observations
54
N
N
N
N
N
N
Y
54
N
N
N
N
N
N
Y
4.12
0.968∗∗
( .477)
54
N
N
N
N
N
Y
Y
5.95
1.021∗∗
(0.419)
54
N
N
N
Y
Y
Y
Y
6.35
0.977∗∗
( 0.388)
54
N
N
Y
Y
Y
Y
Y
6.04
0.981 ∗∗
( 0.399)
54
Y
Y
Y
Y
Y
Y
Y
7.40
1.021∗∗∗
(0.375)
-0.312∗∗
(0.154)
0.370
(8)
IV
Notes: The table reports IV estimates. The instrument is native-nearby cultural distance. The unit of observation is a county
in Republican China. “Fractionalization” includes both within-native fractionalization and overall fractionalization. First-stage F
statistic is Kleibergen-Paap rk Wald F statistic. Robust standard errors are used in all specifications. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗
p < 0.01
54
7.13
1.123 ∗∗
(0.420)
First Stage
Dependent variable: migrant-native cultural distance
-0.250∗
(0.149)
0.170
-0.219∗∗∗
(0.051)
0.174
First-stage F Statistic
Native-nearby culture distance
Adjusted R2
MNCD
(2)
IV
(1)
OLS
Table 7: Instrumental Variable Analysis: Pre-existing Native-Nearby Culture Distance
where normalized isonomynative,nearby,i =
PS
P
P
qP k k,native,iPk,nearby
.S
S 2
S 2
P
k k,native,i
k Pk,nearby
is the number of the same
surnames in the two groups. Pk,native,i and Pk,nearby are the relative frequencies of surname
k within natives and within the nearby population. For the nearby population, we use the
proportion of each surname to the population total in the 2005 Population Census to proxy for
Pk,nearby . The denominator measures how likely any individual randomly drawn from within
natives bears the same surname as one drawn from within the nearby population. The variable,
native-nearby cultural distance (“NNCD”), is intended to capture the cultural dissimilarity
between the native population and the nearby population.
The intuition behind the instrument is that natives are more likely to be very different from the
incoming migrants, if ex ante, natives are more different from all prospective migrants. I treat
the nearby population as the pool of prospective migrants. This is a reasonable assumption as
the vast majority of migrants in the Lower Yangzi after the war are found to have originated
from that area (Cao, 1998; Cao and Li, 2000; Ge, 2002a; Liu, 2012; Zhe and Zhe, 1896). The
cultural distance between natives and the pool of prospective migrants is independent of the
Taiping Rebellion.
B IV Results
Table 7 compares six IV specifications. We run our regressions with no controls in Columns 1
and 2. The first stage is slightly above 7. The IV estimate is -0.25, only slight larger than the
OLS estimate in Column 1 (-0.219) in terms of magnitude. We add baseline controls in Columns
3 and 4. In Column 5, we add geographic controls, resulting in an IV estimate of 0.354. We then
add initial conditions, fractionalization and war-related conditions sequentially from Column 6
to Column 8. IV estimates in those specifications are highly comparable ranging from -0.312 to
-0.354. Our first-stage F statistics are not particularly high, but this is likely due to the small
sample size. The first-stage F statistics do improve when MNCD outliers are excluded, possibly
due to improved linearity. We also show reduced-form estimates in Table A-2. Our instrument
is negatively correlated with the density of public primary schools in specifications. In Columns
1, 5 and 6, the coefficient estimates drops slightly below the conventional cutoff of statistical
significance, with a p-value of 0.15.
Across the columns, IV estimates are slightly larger than the OLS estimates for the same specifications. We attribute the discrepancies partly to the measurement error in our migrant-native
cultural distance variable. As our surname sample is on the small side for some counties, we
may have a lot of sampling error in determining the relative frequency of a surname in the native
or overall population. In addition, when constructing our measure of MNCD, we assume people
with different surnames in the same location, had similar exposure to the war, and had similar
recovery growths in population after the war. While this assumption is unlikely systematically
violated, our measure can suffer measurement error and be noisy. In comparison, our instrument uses a much larger surname sample from the 1% population census. We have reason to
believe that our instrument has less measurement error with regards to the relative frequency of
20
a surname in a population; hence, our IV estimates do not suffer the same attenuation bias as
our OLS estimates do. In addition, MNCD is partially determined by population losses during
the rebellion, whereas our instrument is not affected by differential population losses during
the rebellion. We control for battle exposure, but it is far from being the perfect indicator of
population losses.11 IV estimates can differ from OLS estimates as they are unaffected by the
magnitude of population losses.12
VII. MECHANISMS: THEORY AND SUGGESTIVE EVIDENCE
Our results suggest cultural differences between migrants and natives have a negative impact on
public goods provision. In this section, we provide suggestive evidence on possible mechanisms.
The Lower Yangzi, in spite of being more developed than the rest of the country, remained
highly agrarian at the time. The average urbanization rate was merely 12%, which means the
vast majority of people lived in the countryside. In the sixty counties we study (fifty-four in
the main sample), every county comprises hundreds of villages. Many natives in those villages
typically have lived there for hundreds of years. Close kinship ties ensured that levels of trust
between villagers were relatively high, and this allowed them to collaborate in the provision of
public goods. In some villages all the inhabitants were from the same linage, in which case,
most of them would share the surname; other villages had a few lineages. Many of the public
goods, such as security and education, were provided by and within the clan. But some of
those goods were provided by the gentry to the whole village in the form of private donations.
The long tradition of village life, which went back centuries, if not millennium, constituted a
deep institutional memory of how to deal with problems such as the provision of local public
goods. Both natives and migrants (back in their own hometowns) had adapted to this stable
and slow-moving environment. So when migrants first showed up in the villages of the natives,
they had to sort out an arrangement in which their coexistence with natives would be possible;
a few decades after, migrants and natives were confronted with a new task: to build modern
schools for their next generation.
Based on theoretical and empirical research, there are several ways migrant-native cleavages
can affect public goods provision and prevent them from building modern schools together:
first, differences in preferences between migrants and natives would mean a lower chance for
migrants and natives to reach a consensus on whether to provide a public good or the best way
to provide that good. Second, mutual dissatisfaction, which is often a function of the cultural
11
Despite scholarly efforts to estimate population losses due to the rebellion, the actual damage remains
unknown (Cao and Li, 2000; Cao, 1998).
12
Population losses could have an independent effect on primary schools, and MNCD could be picking up the
effects of population losses in addition to the effects of the cultural distance between migrants and natives. Also,
although population losses were substantial for most counties in the Lower Yangzi, there were a few exceptions.
As places with little to none exposure to the war are “never-takers”, the local average treatment effect is estimated
on counties with actual in-flow migration after the rebellion.
21
distance between migrants and natives, can block any bilateral collaboration and negotiation.
For example, when a village leader originates from the group of migrants, natives would decide
to actively oppose whatever policies he initiates. Esteban et al. (2012a) shows empirically that
ethnic polarization, which accounts for distances between groups, will influence conflict if the
prize is “public”.13 And a closely related scenario is that a leader from the side of natives (or
migrants) indeed looks out for the best interest of natives (or migrants), and at the expense of
the other. When a leader from the side of natives (or migrants) takes advantage of the other
side, often in the name of public interest, it is to be expected for the other side to oppose
expropriatory policies proposed by the leader.
We find qualitative evidence in line with the mechanisms laid out above. First, villagers found it
difficult to raise taxes or use clan or temple assets to establish modern schools, when they were
from different dialectal and cultural background, due to mutual misunderstanding or outright
disagreement between migrants and natives. Second, migrants opposed the policies put forward
by natives. Tian and Chen (2008) document cases where migrants into rural areas refused to pay
county taxes designated to finance upper-primary schools on the grounds that “the schools only
serve the rich in the cities and towns”. In this example, we see both migrants being victims to
expropriatory polices, and a strong identity held by migrants reflected in their statement. Third,
native-migrant conflicts had a direct impact on public goods provision by intervening with the
day-to-day operation of a village. Community leaders were reluctant to build public modern
schools in the midst of frequent native-migrant conflicts. lowered the ability of a community
to perform effective decision-making in the provision of public education. Below we present
quantitative evidence to corroborate our analysis.
A Evidence from Types of Schools
We exploit a feature of the school system in early 20th century China: the financing of lowerprimary and dual-primary schools differed from that of upper-primary and secondary schools.
Overall, the financing of different types of educational institutions after 1905 was parallel with
that of the traditional system. The role of central and provincial governments was limited
to financing universities, students studying abroad, and secondary schools (including teacher
training schools) at provincial capitals.14 County governments financed primarily upper-primary
schools, which were mostly located in cities and towns, with a combination of county tax receipts,
business tax surcharges, the reallocation of endowments from traditional schools, and private
contributions by local elites (Chaudhary et al., 2012). Village communities worked to finance
13
Esteban et al. (2012b) develops a theory of conflict across groups allowing for “public” and “private” prizes.
A leadership position is an example of a “public” prize. In our context, possible sources of dissatisfaction, or
even conflict, include a lack of trust in migrants or a loss of pride due to losing the election as natives. The
lack of trust can arise from the perception of the village leader being susceptible to predating on the interest of
natives.
14
The decentralization of fiscal authority after the Taiping Rebellion left the majority of tax revenues to
county-level authorities. So the central state had very limited resources to finance primary schooling at the local
level.
22
dual-and-lower-primary schools in rural villages. A frequent practice was for community leaders
to transform clan schools and other properties into modern dual-and-lower-primary schools,
and to cover the day-to-day expenses with local taxes and private donations. Those schools
received very little financial support from the county government and its affiliates.15 Figure 3
lists distribution of financial sources for different types of schools in Hangzhou County (Zhang,
2008). According to Figure 3, lower-and dual-primary schools had a higher share of their budget
made up by local funds.
Another institutional feature we exploit is that Chinese villages are highly autonomous and selfgoverned. Constrained by state capacity, the reach of the central government tended to end at
the county level (Qu, 2003, pp.11-15, 179-201, 248-255).At the county level, again, each country
comprised of a large number of villages, making it difficult to project power into those villages.
The self governance of villages has a long tradition in China (Fei, 1939, 1992; Kung-Chuan,
1967; Shi et al., 1988).In an environment of self-governance, migrant-native cultural distance
might manifest itself most forcefully in the outcomes of collective decision-making processes, as
there will not be countervailing forces to their decisions.
With these two institutional features in mind, we can derive the following testable implication:
ceteris paribus, MNCD should have a greater impact on lower-primary and dual-primary schools
than on other types of schools. Fortunately, the educational census (1907-08) reports number
of schools and schools separately for five types of schools: secondary schools, upper-primary
schools, dual-primary schools and lower-primary schools and pre-schools. So we can empirically
test how MNCD affects the density of different types of schools.
We begin our analysis by running a simple regression of county government spending on density
of lower-primary and dual-primary schools. Panel A of Table 8 suggests that government funding
is strongly correlated with upper-primary and secondary schools rate (Col. 3, 4), but not with
lower-primary and dual-public primary schools (Col. 1, 2). These are highly consistent with
the information contained in Figure 3: the county government was not a major contributor to
the budget of the lower-primary and dual-primary schools, but provided funding to secondary
schools and upper-primary schools. It is not surprising that both upper-primary school and
secondary schools increase in county government spending on education, while there are no such
relationships for lower-primary or dual-primary schools.
When we replace county government spending with migrant-native cultural distance in Panel
B of Table 8, we find the type of schools with coefficients that are significant, flipped. We
find MNCD has a strong impact on schools at lower-primary and dual-primary schools, but
not on schools at upper-primary and secondary schools. This is consistent with our testable
implication that MNCD should matter the most for the setting of self-governance—given the
period we examine, both villages and townships were self-governed, and lower-primary and dual15
Schools also charged tuition to supplement other sources of funding. Tuition accounted for 10%-20% of
school budgets in 1917.
23
Commu Secondary Schools
nity or
industry
raised
13%
Tuition
and fees
2%
Upper-Primary Schools
Commu
nity or
industry
raised
6%
Tuition
and fees
8%
County
or
townshi
p raised
85%
Tuition
and fees
2%
County
or
township
raised
86%
Dual-Primary Schools
Lower-Primary Schools
Tuition
and fees
7%
Commu
nity or
industry
raised
56%
County
or
township
raised
42%
County
or
township
raised
7%
Commu
nity or
industry
raised
86%
Figure 3: The Makeup of the Budget of Different Types of Schools
primary schools were determined, established and financed locally. When migrants and natives
could not get along, they could face severe challenges in solving important and urgent issues
including establishing modern schools for their children.
B Evidence from Timing
Having showed the differential impact of MNCD on different types of schools in the 1900s, we
next explore the time-varying impact of MNCD as funding responsibilities over basic education
changed over time. This task is complicated by the fact that public primary schools did not
exist prior to 1905. So we instead look at access to basic education over time.
We exploit two structural breaks in funding responsibilities over basic education. The first
structural break came around in 1905. Prior to 1905, individual households or extended families
were responsible for basic education (as well as for advanced education). There is no evidence for
either village taxes, government aid, or publicly funded basic education. As basic education was
essentially a private good before 1905, exposure to basic education should not vary in MNCD.
The second structural break kicked in close to the end of the 1920s. From then on, the fiscal
authority over lower-primary and dual-primary schools migrated to upper-level governments. In
24
Table 8: Government Funding, Migrant-Native Cultural Distance and the Density of
Schools
Lower-primary
Dual-primary
Upper-primary
Secondary
Panel A. Government Funding: By School Type
Baseline controls
0.067
(0.979)
Yes
0.252
(0.338)
Yes
0.327∗∗∗
(0.073)
Yes
0.242∗∗∗
(0.053)
Yes
Adjusted R2
-0.078
0.069
0.171
0.328
Government spending on education
Panel B. Migrant-Native Cultural Distance: By School Type
MNCD
Baseline controls
Observations
Adjusted R2
-0.185∗∗∗
(0.042)
Yes
-0.052∗∗
(0.021)
Yes
-0.003
(0.005)
Yes
-0.003
(0.003)
Yes
54
0.142
54
0.155
54
0.122
54
0.178
Notes: Panel A reports the impact of government spending on then density of different
types of schools (#schools per 10,000 persons). Panel B reports the impact of migrantnative cultural distance by school type. The unit of observation is a county in Republican
China. Robust standard errors are included in all specifications. ∗ p < 0.10, ∗∗ p < 0.05,
∗∗∗ p < 0.01
the meanwhile, upper-level governments become much more autocratic over time. The combined
effect of schools being funded by upper-level governments and upper governments becoming more
autocratic, should render local decision making increasingly less relevant, which should yield a
less economically significant coefficient of MNCD. These two structural breaks give rise to the
prediction that the impact of MNCD should be the most relevant between 1905 and the end of
1920s.
We test this prediction in a fully flexible difference in differences specification. Our variables of
interest are interaction terms between MNCD and time period dummies:
%access to basic educationi,t =
X
βt M N CDi × T imeperiodt +
t
X
γt Xi × T imeperiodt
t
(4)
+Zi , tζ + ρt + ηi + εi , t ,
where the year 1960 is left out as the comparison group. The dependent variable is log share of
population receiving basic education in County i during time period t. Xi is a vector of other
time invariant controls and we allow these variables to have time-varying effects on access to
basic education. Zi,t ζ is a vector of time variant controls.ρt denotes a full set of time fixed effect
and ηi denotes a full set of county fixed effects. εi,t is a disturbance term. Our time periods are
defined as follows: ∈ 1820, 1850, 1880, 1900, 1910, 1920, 1930, 1940, 1950.
Table 9 summarizes our results. Across the specifications, the only M N CD × T imeperiod
25
Table 9: The Dynamic Impact of of Migrant-Native Cultural Distance: Baselines
(1)
MNCD×1820
MNCD×1850
MNCD×1880
MNCD×1900
MNCD×1910
MNCD×1920
MNCD×1930
MNCD×1940
MNCD×1950
(2)
(3)
(4)
%access to basic education
-0.054∗∗
(0.025)
-0.077∗∗∗
(0.022)
-0.078∗∗∗
(0.025)
-0.339∗∗∗
(0.062)
-0.168∗∗∗
(0.040)
-0.087∗∗
(0.043)
-0.031
(0.045)
-0.043∗
(0.023)
-0.039∗
(0.023)
-0.025
(0.031)
-0.048
(0.030)
-0.049
(0.035)
-0.309∗∗∗
(0.072)
-0.138∗∗∗
(0.045)
-0.058
(0.036)
-0.002
(0.039)
-0.014
(0.017)
-0.009
(0.011)
N
Y
540
0.904
Y
Y
-0.029
(0.030)
-0.050∗
(0.029)
-0.025
(0.037)
-0.294∗∗∗
(0.076)
-0.129∗∗∗
(0.046)
-0.051
(0.037)
0.001
(0.040)
-0.012
(0.019)
-0.008
(0.013)
0.145∗
(0.084)
Y
Y
-0.021
(0.025)
-0.041∗
(0.024)
-0.031
(0.031)
-0.266∗∗∗
(0.067)
-0.087∗
(0.045)
-0.044
(0.027)
0.010
(0.028)
-0.006
(0.014)
-0.002
(0.010)
0.103
(0.084)
Y
Y
540
0.936
540
0.937
600
0.937
Log population
County FE
Time period FE
Observations
Adjusted R2
Standard errors in parentheses
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Notes: The table reports the dynamic impact of migrant-native cultural distance before and
after the Taiping Rebellion.Column 1 only includes time period fixed effectds. Both time
period effeccts and county fixed are included in all other specifications. Column 3 controls
for log population. Column 4 includes the six counties with imputed MNCD. Column 3 is
our preferred specification. Standard errors are clusterered at the county level.
coefficients that are consistently significant are M N CD × 1900 and M N CD × 1910 . This
is highly consistent with our prediction that MNCD has the most impact on access to basic
education, during the period where basic education is a public good and fiscal authority is
decentralized. We begin our analysis by including only M N CD × T imeperiod, time period
FE. To make full use of the panel data, in Column 2, we add county FE to the regression.
We find M N CD × T imeperiod is negative but insignificant before 1900, and there is no pretrend. It is not the case that the counties became better or worse in providing basic education
over the time. Consistent with our prediction, MNCD did not matter for the period between
the end of the war (1864) and the abolition of the exam (1905), when basic education was
26
Figure 4: The Dynamic Impact of Migrant-Native Cultural Distance
largely financed privately within clans. For 1900-1920, MNCD has a negative effect on schools
of great economic magnitude, and coefficient estimates of interactions between MNCD and time
periods drop substantially after 1920, and are no longer significant.16 In Column 3, we add
a possibly endogenous variable, log population, to the specification, and we get fairly similar
estimates to Column 2.17 For the decade of 1900-1910, a one-standard-deviation increase in
MNCD is associated with 32.3% decrease (0.294*1.1) in access to basic education, relative to
the effect of MNCD on schools in 1960, our omitted category. This estimate is comparable to
our baseline OLS estimate in Column 4 of Table 1. In Column 4, we run the same regression
as in Column 3, but on a sample including counties with imputed MNCD. The estimates are
somewhat smaller, likely due to the attenuation bias posed by less well measured MNCD of the
newly added counties.
The differential impact of MNCD in Table 9 before and after two structure breaks, are most
similar to the approach featured in Dippel (2014). Dippel (2014) finds the impact of forced
coexistence on economic prosperity only emerged after 1990, when Indian reservations were
given substantially more autonomy. We find the impact of MNCD only began to surface after
the modern educational reform, when the responsibilities of individual households and clans
in providing basic education was partly taken over by villages, and only died out after 1920,
when village and small town autonomy began to be eroded by upper-governments. Evidence
from the timing of MNCD kicking in and dying out suggests that the primary mechanism runs
16
accesst ob asice ducation1910 is the average of public primary schools recorded in the 1914, 1915 and 1916
Census.
17
We would like to include time-varying urbanization rates as well, just to be comparable to the specifications
in Table 1. Unfortunately, we only have urbanization for one period.
27
from migrant-native cultural distance through quality of collective decision-making to public
primary schools. In addition, the general consistency between our panel and OLS estimates
further enhances our confidence in the findings.
VIII. CONCLUSION
This paper uses quasi-exogenous migration to identify the impact of cultural proximity between
migrants and natives on public goods provision outcomes. Our key finding is that cultural
distance has an independent effect on public goods provision outcomes, conditional on the
history of shared governance.
In addition to our key finding, we also find partial evidence that a history of shared governance
mitigates the negative effects of the cultural distance. We find that the culture distance between
previously detached groups has a statistically significant effect on public goods provision, whereas
the polarization measure taking into account differences between surname groups (“surname
group distance”) without making any adjustment for their shared history, does not have the same
effect. A most likely explanation for this is that the culture distance between previously detached
groups (migrants and natives) does more damage to public goods provision than does the cultural
distance between groups with a history of shared governance. A history of shared governance
might have nurtured trust and increased the level of collaboration. This is consistent with
Dippel’s (2014) finding that a history of shared governance facilitates public goods provision.
Therefore, we believe our findings are consistent with both sides of the debate as to whether
ethnic cleavages or artificial jurisdictions has caused poor economic performance.
Our analysis sheds light on how migrant-native cultural distance, as a key aspect of population
heterogeneity, is related to the failure of educational reforms in the early 20th century. Our
finding that a greater cultural distance between migrants and natives is associated with a lower
density of public primary schools reveals the unique challenge faced by communities with many
migrants from afar face in their attempt to modernize their education. For those communities,
educational reforms proved to be incredibly challenging and frustrating, due to intense social
antagonism between two groups. The under provision of modern education could prevent have
prevented the accumulation of modern human capital and increased barriers to industrialization
for those communities.
A more general point is that cultural distance is associated with a higher cost in modernizing the
governance structure in China. Following China’s constitutional reforms, village communities
aimed to provide a wider range of public goods, as well as to formalize its governance structure
(Kuhn, 2002, Chapter 4). The failure to provide a key public good—public primary schools—
actually reveals a deeper issue with the process of modernizing the village governance structure.
The idea that low social capital can prevent effective self-governance in Chinese villages is
28
explored in Qian, Xu, Yao, et al. (2015).18 Formalization and democratization of village decisionmaking following the constitutional reforms was hampered by strong social antagonism between
migrants and natives, the two groups with the most mistrust for each other. Social antagonism
between different groups in the village, created substantial barriers for post-1905 villages to
modern its political structure and to provide a wider range public goods than before. We argue
that cultural distance can affect the quality of village governance through social capital and
delay the emergence of full-fledged modern village governance.
Lacking the support of well-functioning local villages, Qing China made little meaningful progress
in its political reforms before its collapse in 1911. While both constitutional reforms and ”modern school movement” continued into Republican China, they still faced very similar constraints
that Qing China had during its last ten years. By 1930s, modern schools eventually attained
rapid growth but only at the expense of village self-governance and political decentralization
(Kuhn, 2002). The stagnant growth of public primary schools in a time of decentralization
parallels the experience nineteenth-century Prussia had with a decentralized education system
in a linguistically polarized society (Cinnirella and Schueler, 2016). The transition to autocratic
rule after 1930 seems to be consistent with Galor, Klemp, et al. (2015) which establishes the
point that population heterogeneity can provide a breeding ground for autocratic rule.
For those interested in the historical role of the gentry class, our paper also sheds some light
on the role of gentry men in modernizing village governance in a discussion of the mechanisms
linking MNCD to poor public goods provision outcomes. Gentry men were elite members of their
village. Many of them were enthusiastic about making contributions to their village. Schools
before 1905 relied heavily on their private donations, and remained so to a certain degree even
after 1905. However, our findings most likely indicate that gentry men were not best suited to
fulfill leadership roles in a modern context—gentry men, because they were viewed as closely
associated with their clan, encountered criticism about their impartiality; these tensions helped
fuel political conflict as leadership positions were seen as “public” prizes by a heterogeneous
population.
18
Qian, Xu, Yao, et al. (2015) find the presence of village temples, their proxy for social capital, enhances
public goods provision through the channel of elections.
29
References
Ager, P. and M. Brückner (2013). Cultural diversity and economic growth: Evidence from the us
during the age of mass migration. European Economic Review 64, 76–97.
Aghion, P., A. Alesina, and F. Trebbi (2005). Choosing electoral rules: theory and evidence from us
cities. Technical report, National Bureau of Economic Research.
Alesina, A., R. Baqir, and W. Easterly (1999). Public goods and ethnic divisions. The Quarterly
Journal of Economics 114 (4), 1243–1284.
Alesina, A. and E. L. Ferrara (2005). Ethnic diversity and economic performance. Journal of Economic
Literature 43 (3), 762–800.
Altonji, J. G., T. E. Elder, and C. R. Taber (2005). Selection on observed and unobserved variables:
Assessing the effectiveness of catholic schools. Journal of Political Economy 113 (1), 151–184.
Ashraf, Q. and O. Galor (2013). The’out of Africa’hypothesis, human genetic diversity, and comparative
economic development. The American Economic Review 103 (1).
Bai, Y. and J. K.-s. Kung (2011). Genetic distance and income difference: Evidence from changes in
china’s cross-strait relations. Economics Letters 110 (3), 255–258.
Bai, Y. and J. K.-s. Kung (2014). Diffusing knowledge while spreading god’s message: Protestantism
and economic prosperity in china, 1840–1920. Journal of the European Economic Association.
Bai, Y. and J. K.-s. Kung (2015). Does genetic distance have a barrier effect on technology diffusion?
evidence from historical china.
Bazzi, S., M. Gudgeon, et al. (2015). Local government proliferation, diversity, and conflict. Technical
report, Boston University-Department of Economics.
Bossert, W., C. D’AMBROSIO, and E. La Ferrara (2011). A generalized index of fractionalization.
Economica 78 (312), 723–750.
Bossert, W., P. K. Pattanaik, and Y. Xu (2003). Similarity of options and the measurement of diversity.
Journal of Theoretical Politics 15 (4), 405–421.
Buck, J. L. and C. Press (1937). Land Utilization in China: a study of 16,786 farms in 168 localities,
and 38,256 farm families in twenty-two provinces in China, 1929-1933, Volume 2. University of
Chicago Press.
Cao, S. (1998). The impact of the taiping rebellion on population in south jiangsu. History Research 2,
64. .
Cao, S. and Y. Chen (2002). Malthusian theory and chinese population since qing: comments on recent
studies by american scholars. History Study (lishi yanjiu) 1, 41–54.
Cao, S. and Y. Li (2000). The impact of the taiping rebellion on population in zhejiang. Journal of
Fudan University: Social Sciences 5, 005. .
Cao, z. (2006). On the rise and fall of linguistic enclaves: the case of wu-hui mandarin dialect. Linguistic
study (Yuyan yanjiu).
Caselli, F. and W. J. Coleman (2013). On the theory of ethnic conflict. Journal of the European
Economic Association 11 (s1), 161–192.
Chaudhary, L., A. Musacchio, S. Nafziger, and S. Yan (2012). Big brics, weak foundations: The
beginning of public elementary education in brazil, russia, india, and china. Explorations in Economic
History 49 (2), 221–240.
CHGIS, H. Y. I. (2007). CHGIS.
Cinnirella, F. and R. M. Schueler (2016). Dp11274 the cost of decentralization: Linguistic polarization
and the provision of education.
Crayen, D. and J. Baten (2010). Global trends in numeracy 1820–1949 and its implications for longterm growth. Explorations in Economic History 47 (1), 82–99.
Davis, D. R. and D. E. Weinstein (2002). Bones, bombs, and break points: the geography of economic
activity. The American Economic Review 92 (5), 1269–1289.
Desmet, K., M. Le Breton, I. Ortuño-Ortı́n, and S. Weber (2011). The stability and breakup of nations:
a quantitative analysis. Journal of Economic Growth 16 (3), 183–213.
30
Dippel, C. (2014). Forced coexistence and economic development: evidence from native american
reservations. Econométrica: Journal of the Econometric Society 82 (6), 2131–2165.
Du, R., C. Xiao, and L. Cavalli-Sforza (1997). Genetic distances between chinese populations calculated
on gene frequencies of 38 loci. Science in China Series C: Life Sciences 40 (6), 613–621.
Duclos, J.-Y., J. Esteban, and D. Ray (2004). Polarization: concepts, measurement, estimation. Econometrica, 1737–1772.
Easterly, W. and R. Levine (1997, November). Africa’s growth tragedy: Policies and ethnic divisions.
The Quarterly Journal of Economics 112 (4), 1203–50.
Esteban, J., L. Mayoral, and D. Ray (2012a). Ethnicity and conflict: An empirical study. The American
Economic Review , 1310–1342.
Esteban, J., L. Mayoral, and D. Ray (2012b). Ethnicity and conflict: Theory and facts. science 336 (6083), 858–865.
Esteban, J. and D. Ray (2011). Linking conflict to inequality and polarization. The American Economic
Review 101 (4), 1345–1374.
Esteban, J.-M. and D. Ray (1994). On the measurement of polarization. Econometrica: Journal of the
Econometric Society, 819–851.
Fei, H.-T. (1939). Peasant life in China. 19, 39.
Fei, X. (1992). From the Soil, the Foundations of Chinese Society: A Translation of Fei Xiaotong’s
Xiangtu Zhongguo, with an Introduction and Epilogue. University of California Press.
Galor, O., M. Klemp, et al. (2015). Roots of autocracy. Technical report.
Ge, J. and S. Cao (2001). Zhongguo renkoushi qingdaijuan (History of China’s Population: Qing
Dynasty). Fudan University Press.
Ge, q. (2002a). jindai suzhewan jiaojie diqu renkou qianxi yanjiu (Migrants from Hunan and Hubei in
the intersection of Jiangsu, Zhejiang and Anhui). Shanghai: Shanghai Academy of Social Science
Press.
Ge, Q. (2002b). Minguo shiqi zhejiang huzheng yu renkou diaocha (Household and Population Census
in Zhejiang During the Republican Period ). Chinese Academy of Social Sciences press. .
Gennaioli, N. and I. Rainer (2007). The modern impact of precolonial centralization in africa. Journal
of Economic Growth 12 (3), 185–234.
Guo, y. (1989). Historical geography of the taiping rebellion (taiping tianguo lishi dituji.
Huang, X. (2004). An study on the mandarin dialect enclaves of Anji County, Zhejiang province
(Zhejiang anjixian fangyandao yanjiu). Dissertation, Beijing Language University.
Junya, M. and T. Wright (2010). Industrialisation and handicraft cloth: The jiangsu peasant economy
in the late nineteenth and early twentieth centuries. Modern Asian Studies 44 (06), 1337–1372.
Kuhn, P. A. (2002). Origins of the modern Chinese state. Stanford University Press.
Kung-Chuan, H. (1967). Rural China: Imperial control in the nineteenth century. University of Washington Press.
Leung, A. K. C. (1994). Elementary education in the lower yangtze region in the seventeenth and
eighteenth centuries. Education and Society in Late Imperial China, 1600-1900 , 381–416.
Li, N. (2011). Essays on cultural diffusion, migration, and human capital: investigation from china’s
historical experience.
Li, W. (1981). The change in land ownership in the area occupied by the taiping rebels in jiangsu,
zhejiang and anhui in late qing. History Study (lishi yanjiu) 6, 81–96.
Li, W. (2002). Restoration of temples and religious networks in rural lower yangzi after the taiping
rebellion. Zhejiang Chronicles (Zhejiang Fangzhi) 1, 62–69.
Liang, q. (2001). Charity and Edification: Charitable Organizations in Ming and Qing (Cishan yu
jiaohua: mingqing de cishan zuzhi). Hebei Education Press.
Lin, C. and N. Li (2014). The impacts of taiping rebellion on long-term economic development. Journal
of Translation from Foreign Literature of Economics (jinji ziliao yicong) 2.
Liu, Y. (2012). Conflict and assimulation in Jiangnan during Late Qing and Early Republic China—
Jurong in Jiangsu. History (Shi lin) 2, 016.
Mann, S. (1994). The education of daughters in the mid-ch’ing period. Education and Society in Late
31
Imperial China, 1600–1900.
Michalopoulos, S. and E. Papaioannou (2013). Pre-colonial ethnic institutions and contemporary
african development. Econometrica 81 (1), 113–152.
Mickey, G. A. and T. Shiroyama (2009). China during the great depression: Market, state, and the
world economy, 1929-1937.
Oster, E. (2014). Unobservable selection and coefficient stability: Theory and evidence. University of
Chicago Booth School of Business Working Paper .
Ouyang, N. and W. Zhang (2010). Spatial-temporal analysis of temple-converted schools in Jiangnan
in Late Qing and Republican China. Historical Geography (Lishi dili), 011.
Qian, N., Y. Xu, Y. Yao, et al. (2015). Making democracy work: Culture, social capital and elections
in China. Technical report, National Bureau of Economic Research.
Qu, t. (2003). Local Government in Qing (Qingdai de difang zhengfu). Ph. D. thesis, Beijing: Law
Press.
Rawski, E. S. (1979). Education and popular literacy in Ch’ing China. University of Michigan Press
Ann Arbor, MI.
Reynal-Querol, M. (2002). Ethnicity, political systems, and civil wars. Journal of Conflict Resolution 46 (1), 29–54.
Reynal-Querol, M. and J. G. Montalvo (2005). Ethnic polarization, potential conflict and civil war.
American Economic Review 95 (3), 796–816.
Shi, J., H. Wu, and X. Fei (1988). Emperor and gentry (huangquan he shenquan).
Simmons, R. V., R. Shi, and Q. Gu (2006). An historical geographic study on the division line
of Jianghuai-mandarin dialect and Wu-mandarin dialect. (Wu-huai guanhua yu wuyu bianjie de
fangyan dilixue yanjiu). Shanghai Education Press.
Spolaore, E. and R. Wacziarg (2009). The diffusion of development. Quarterly Journal of Economics 124 (2).
Tian, Z. and S. Chen (2008). Taxation for education and conflicts arising from provision of village
education. Journal of Zhejiang University: Humanities 38 (3), 142–149. .
Xu, X. (2010). Temple property in the Late Qing Dynasty and the Republic of China (1895-1916).
Technical report.
Xu, Y. and Y. Yao (2015). Informal institutions, collective action, and public investment in rural china.
American Political Science Review 109 (02), 371–391.
Yin, M. and Q. Tian (2009). Minguo renkou huji shiliao huibian (Compilation of Historical Population
Statistics During the Republican Period). National Library Press. .
Yuchtman, N. (2015). Teaching to the tests: An economic analysis of traditional and modern education
in late imperial and republican china. Unpublished manuscript 27.
Zhang, x. (2008). Qingdai jiangnan gonggong jiaoyu ziyuan choucuo peizhi de lishi dili xue fenxi (A
Historical Geographical Analysis of Public Education Resource Allocation in Jiangnan from 1644 to
1911). Ph. D. thesis, Fudan University. (1644-1911 ).
Zhe, c. and c. Zhe (1896). Draft Gazetteer of Yuhang County (Guangxu yuhangxian zhigao(fu buyi)).
32
Appendix
A Additional Tables
Table A-1: Summary Statistics
Variable
Mean
Std. Dev.
Observations
Source
#public primary schools per 10,000 persons
0.937
0.518
60
b
Measures of population heterogeneity:
Migrant-native cultural distance
Overall diversity
Within-native diversity
Overall polarization
Within-native polarization
2.12
35.357
30.166
1.297
0.85
1.024
5.984
6.311
0.729
0.293
54
54
54
54
54
a
a
a
a
a
Instrumental variables:
Native-nearby cultural distance
1.578
0.421
54
a
Baseline controls:
Log(%access to basic education), 1880
Urbanization rate, 1917
Log population
3.136
12.614
12.456
0.293
16.093
0.869
60
60
60
b
d
b
Geographic controls:
Ruggedness
% arable land
Agricultural suitability
Dist. to Grand Canal
Dist. to provincial capital
Dist. to Shanghai
Dist. to Yangzi
1.857
0.535
2.851
0.27
137.617
171.15
1.158
2.532
0.242
2.054
0.26
95.719
87.681
0.694
60
60
60
60
60
60
60
g
e
g
g
g
g
g
Initial conditions:
#initial charities
Log population, 1820
Log population density, 1820
2.583
12.978
6.095
4.928
0.739
0.655
60
60
60
h
b
b
Impact of war:
Battle exposure (months)
% adult men
% elderly and youth
Human capital shock
31.167
20.108
16.527
-0.022
16.955
2.090
1.661
0.053
60
60
60
54
i
c
c
b
Interventions in education:
Log (#communicants per 10,000+1)
Log (#temple-converted schools+1)
2.921
1.158
1.152
0.930
60
60
f
j
33
Variable
Mean
Std. Dev.
Observations
Source
Mechanisms:
Government spending on education
#lower-primary schools per 10,000
#dual-primary schools per 10,000
#upper-primary schools per 10,000
#secondary schools per 10,000
0.014
0.589
0.179
0.063
0.01
0.036
0.4
0.169
0.055
0.02
60
60
60
60
60
b
b
b
b
b
For panel estimates:
Log(%access to basic education)
Log population
2.971
12.707
1.105
0.859
600
600
b
b
a. Surname data sources (see Appendix B)
b. Primary schools data sources (see Appendix C)
b. Ge and Cao (2001)
c. Yin and Tian (2009, Volumn 4), Compilation of Historical Population Statistics During
the Republican Period; Ge (2002b), Household and Population Census in Zhejiang During
the Republican Period.
d. Yin and Tian (2009, Volumn 1), Compilation of Historical Population Statistics During
the Republican Period.
e. Buck and Press (1937), Land Utilization in China: a study of 16,786 farms in 168
localities, and 38,256 farm families in twenty-two provinces in China, 1929-1933.
f. Local gazetteers on religious facilities.
g. CHGIS (2007)
h. Liang (2001)
i. Guo (1989)
j. Ouyang and Zhang (2010)
Table A-2: Instrumental Variable Analysis: Reduced-Form Estimates
(1)
(2)
(3)
(4)
(5)
(6)
Dependent variable: #public primary schools per 10,000 persons
Native-nearby culture distance
Human capital shock
Battle exposure
Fractionalization
Initial social capital
Pop. 1820
Georgraphic controls
Baseline controls
Observations
Adjusted R2
-0.280
(0.194)
N
N
N
N
N
N
N
-0.371∗
(0.191)
N
N
N
N
N
N
Y
-0.362∗
(0.182)
N
N
N
N
N
Y
Y
-0.367∗
(0.184)
N
N
N
Y
Y
Y
Y
-0.317
(0.216)
N
N
Y
Y
Y
Y
Y
-0.319
(0.212)
Y
Y
Y
Y
Y
Y
Y
54
0.034
54
0.047
54
0.381
54
0.367
54
0.348
54
0.331
Notes: The table reports reduced-form estimates of native-nearby culture distance. The
unit of observation is a county in Republican China. The specifications correspond to those
in Table 7. Robust standard errors are used in all specifications. ∗ p < 0.10, ∗∗ p < 0.05,
∗∗∗ p < 0.01
34
B Data Sources for Surnames
An Overview of Surname Data Sources
Pre-Taiping period
Deaths during the
Exam degree holders
Taiping Rebelliona
1645-1850b
31126
Post-Taiping period
Diseased soldiers
College students
1927-52c
1906-1949d
13340
39419
15584
Sources:
a: Traditional county gazetteers (pre-1949): chapter for deaths of high moral worth. 旧志-咸同忠烈
姓名录
b: Traditional county gazetteers (pre-1949): chapter for exam degree holders. 旧志-选举志（by year
of degree）
c: Modern county gazetteers(post-1949): chapter for deceased soldiers. 新志-抗战英烈和革命英烈
姓名录 (by year of death)
d: College yearbooks during the Republican era (by year of admission). See Hao and Clark (2014) for
more details.
35
36
Figure A-1: Data Sources for Surnames
C Data Sources for Basic Education
Year
Region
Sources
1820-1900
The Lower Yangzi Same as surname data sources. See “Some notes on surname data”.
The “Statistical Chart of Education” (Jiaoyu Tongji Tubiao) by Ministry of Education (教育部)
1907
Nation
光绪三十三年第一次教育统计图表
1908
Nation
光绪三十四年第二次教育统计图表
1909
Nation
宣统元年第一次教育统计图表
1915
Nation
中华民国第三次教育统计图表
1916
Nation
中华民国第四次教育统计图表
1917
Nation
中华民国第五次教育统计图表
1923
Jiangsu
江苏政治年鉴
The Yearbook of Politics: Jiangsu Province (1923)
The “Statistical Chart of Education” (Jiaoyu Tongji Tubiao) by Department of Education (教育
厅), Zhejiang Province (1925-1929)
1925
Zhejiang
中华民国十四年度浙江省教育统计图表
1927
Zhejiang
中华民国十六年度浙江省教育统计图表
1929
Zhejiang
中华民国十八年度浙江省教育统计图表
1929
Jiangsu
江苏教育概览 The Overview of education: Jiangsu province (1929)
The “Statistical Chart of Education” (Jiaoyu Tongji Tubiao) by Department of Education (教育
厅), Jiangsu Province (1932-1937)
1932
Jiangsu
民国二十一年度江苏省教育经费统计图表
1933
Jiangsu
民国二十二年度江苏省教育经费统计图表
1935-37
Jiangsu
江苏省教育统计图表
1935-37
Zhejiang
浙江省三年来教育概况
Education for the Last Three Years: Zhejiang province
1947
Jiangsu
江苏省 36 年教育统计
浙江经济年鉴
1947
Zhejiang
The Yearbook of Economy: Jiangsu Province (1923)
1950-60
Jiangsu
江苏五十年(1949-1999)
The fifty-year statistics of Jiangsu (1949-1999)
1950-60
Zhejiang
新浙江五十年统计资料汇编(1949-1999)
The fifty-year statistics of Zhejiang (1949-1999)
Figure A-2: Data Sources for public primary schools (1820-1960)
37
D The Traditional Education System
Prior to 1905, the primary education system was based upon Confucian classics and aimed at
success in the imperial examinations. At national, provincial and county levels, highly competitive exams selected a few degree holders, and brought the individual, his clans, and communities
“prestige, power, and wealth through government service (Rawski, 1979, p.21). Within a community or a clan, an individual’s literacy helped determine his social status, occupation and
wealth. The return to passing exams generated considerable demand for privately or publicly
provided traditional schooling throughout the country. On the other hand, due to commercialization in this area since the 16th century there was a growing demand for education, especially
basic education. The educational system consisted of two phases: mass primary education
teaching basic literacy and select students of talent, and the more advanced education drilling
candidates to prepare for exams (Leung, 1994). Those who reached the threshold level in literacy
to attend the exam (4,000 characters) accounted for only 1%-2% of the male population(Rawski,
1979, p.96). A much higher percentage of male population, roughly 30% to 40%, acquired basic literacy (numeracy and about 1000 characters) Crayen and Baten (2010). Female literacy
(2%-5%) was confined to those from elite families (Mann, 1994; Rawski, 1979).
Both basic and advanced education were mainly provided privately. Children in elite households were instructed by family members to begin with (age 3-6), and by hired private tutors
between age 6 and 15 (typically exam degree holders who had not been to a government position). Households of modest backgrounds pooled their resources and paid local teachers for
instructions. There is no evidence that a village tax was collected, nor any aid from government
was received to finance those educational instructions. Parents paid the teacher in either money
or commodities. Each tutor taught 1-30 students in the tutor’s own house or the village temple.
Hours and schedules were flexible, adjusted to weather and season. The cost varied with the
quality of instructions, and degree holders could demand a higher salary as tutors. At the age
of nine or ten, decisions were made based on the observed talent of children. Families would
support the talented ones to pursue exam careers, while they prepare others for various lesser
occupations.
Boys from families too poor to pay for schooling were not necessarily barred from the classroom.
Clan schools (zu xue) were often established primarily to aid such students (Rawski, 1979, pp.
30-32). These schools were financed by contributions, a clan tax, and clan land (xue tian) with
the rent going to special funds for education and exam preparation. Clan temples (zong miao)
often served as classrooms. For the talented students, their future study and exam taking were
fully covered by land funds. Most clan schools limited admission to kin members but exceptions
were made for talented non-kin members within the community . Overall, in the lower Yangzi
about 5-8% of males were supported by their kinships for their education.
The central and local government in late imperial China took a hands-off approach to financing
education. Local magistrates advocated setting up primary schools for the poor but there
38
was no record of direct funding from the government (Rawski, 1979, pp. 38-40). The education
establishments with direct government funding, called ”government-based county and prefecture
schools”, were for no more than a few hundred lower degree holders to accomplish higher degrees
(an allowance was provided to cover basic living costs and travel costs). There were a few
exceptions to the rule: in frontier regions populated by non-Han minorities, the government
funded the establishment and maintenance of schools; in areas with recent exposure to war and
famine, the government often helped to establish schools as a means to restoring the Confucian
order (Rawski, 1979, p.89). For instance, after the Taiping rebellion, the magistrate of the
county of Wujiang set up sixteen charity schools with the county budget, but those schools
admitted no more than 500 students county-wide, which accounted for only 1% of school-aged
children. Even so, these schools were closed after a decade because of ”a tight budget” and
the restoration of ”private and clan schools to a pre-war level.In most cases, such governmentinitiated charity schools only admitted 0.5%-1% of the school-aged children, or 2%-4% of the
entire educated population.
E The Modern Education System
Content
Pedagogy
Early childhood
(4-7)
Traditional
Modern
Basic Chinese charac- Memorizing
ters
Tutored at home by
family members
Primary (7-12)
Poetry and Confucius Memorizing; Tutored
classics
at home or at sishu by
private tutors
Juvenile (13-17) Preparing for exams Memorizing
and
and youth (18- (writing eight-legged writing; Studied at
25)
essays based on Con- home or in private
fucius classics)
academies
Early childhood
(4-7)
Primary (7-12)
Juvenile (13-18)
Youth (19-25)
Basic Chinese characters
Chinese, art, mathematics
Chinese, art, mathematics, western science
Art, western sciences,
foreign languages, law
and engineering
Tutored at pre-schools
Studied in primary
schools
Studied in high primary schools and secondary schools
Studied in universities
Table A-3: Traditional versus Modern Education. Sources: Yuchtman (2015)
39
Example county: County of Wu （吴县）
Village
Small
town
Lower-primary
schools
and dual-primary
schools
Upper-Primary Schools
Bureau of Education
县劝学所，教育局
City of Suzhou (苏州）
Secondary Schools
District-level
agency 区劝
学员
Figure A-3: Types of Schools and The Fiscal Authority in Charge
40
F The Taiping Rebellion
Figure A-4: Battle Exposure
G Measures of Population Heterogeneity
Below we construct measures of population heterogeneity as alternative explanatory variables,
mainly, the fractionalization index and the polarization index. Those indices do not contain
information on migrant-native status. We compare MNCD to measures of population heterogeneity to put the explanatory power of the cultural distance between migrants and natives in
perspective.
G.1
Fractionalization
Akin to the fractionalization measure in (Alesina et al., 1999), we use the inverse of isonomy
to define fractionalization. Rather than ethnic groups, our measure relies on surnames (k). We
measure the probability of randomly drawn two individuals sharing the same surname for a
population at a given time:
S
X
2
I=
Pk,i
,
(5)
k
41
where S is the number of surnames and Pk is the relative frequency of surname k, which is the
proportion of the population with surname k to the population. This is a Herfindahl-Hirschman
index of surname distribution. And,
1
(6)
F RAC = .
I
A higher FRAC corresponds with a lower likelihood that any two individuals randomly drawn
from a population bear the same surname (which indicate closeness in dialect, culture and even
genetics). This fractionalization index treats all groups symmetrically.
MNCD can change without increasing fractionalization. An example is that when people carrying a certain surname are wiped out by the rebellion, and replaced by migrants carrying a
different surname, the overall fractionalization of a population remains unchanged, but MNCD
can still change.
G.2
Polarization
Our first polarization index is taken from Reynal-Querol (2002) and Reynal-Querol and Montalvo
(2005):
POL =
S
X
1
(2
i=1
− Pk
1
2
)2 Pk
(7)
where S is the number of surnames and Pk is the relative frequency of surname k, which is
the proportion of the population with surname k to the population. This measure employs a
weighted sum of population shares and assumes the any two groups are either completely similar
or completely dissimilar.
We then rely on Duclos, Esteban, and Ray (2004); Esteban and Ray (2011); Esteban, Mayoral,
and Ray (2012a) to build a second measure of polarization:
POL DIST =
S X
S
X
Pi2 Pj dij .
(8)
i=1 j=1
The group i here is defined as a surname group. The population share of each surname group is
Pi , the intergroup distance di j. Our measure of intergroup distance is the difference in human
capital between groups. This is not the perfect measure, but this is the best measure we can find
given our data. A similar measure to ours is the difference in incomes between groups (Esteban
and Ray, 1994) and (Aghion et al., 2005). For that reason, our polarization measure should be
interpreted with caution.
We report correlations between MNCD and fractionalization and MNCD and polarization (Table
A-4). We find MNCD is positively correlated with overall fractionalization and polarization, but
42
Table A-4: Correlations Between MNCD, Fractionalization and Polarization
MNCD
Overall FRAC
Overall POL
Overall POL DIST
MNCD
Overall FRAC
Overall POL
Overall POL DIST
1
0.203
-0.229
0.151
1
-0.786∗∗∗
-0.506∗∗∗
1
0.434∗∗
1
∗
∗∗∗
p < 0.1,
∗∗
p < 0.05,
p < 0.01
not with within-native fractionalization and polarization. This is consistent with the nature of
MNCD: it captures the relationship between migrants and natives, not the relationship between
natives. The high correlations between MNCD and overall fractionalization and polarization
suggest that MNCD captures crucial information in overall population heterogeneity.
H Human Capital Shock
Migrants do not always have the same level of human capital as natives. When their levels of
human capital differ, this measure captures the resulting “shock” to the stock of local human
capital:
X
(Pk,M N − Pk,N )RRk
(9)
Human capital shock =
k
where RRk is a surname-specific variable that denotes the representation of surname k among
the educated elites relative to its population, i.e. RRk is the ratio of the share of k in education
elites to the share of k in the total population. Supposing that natives have two surnames:
Wang and Li. Each surname accounts for 50% of the population. Li’s are over represented
among education elites by a factor of 2, in other words, Li’s are two times more likely to show
up among national elites than in an average population, whereas Wang’s are just as likely to be
educated as is a member of the average population. For illustrative purposes, take the extreme
case where all Li’s die in the war. A family of Zhangs came to the village, with a RR of 0.5.
The human capital shock is (50%-50%)*1+(0%-0.5%)*2+(50%-0%)*0.5=-0.5.
We acknowledge that our measure of human capital shock can have measurement errors due to
migrant selectivity—migrants do not have to have the same capital level as the overall population
with their surnames. When skilled migrants migrate to more industrialized areas, our measure
can underestimate the level of their human capital. However, we do not see how this can bias our
results aside from introducing an attenuation bias. Also, this sort of measurement error is going
to be minor for most counties in our sample, with the exception of industrial Shanghai. Figure
A-5 shows geographic pattern of human capital shocks. The human capital shock appears to be
the most positive in the east, consistent with the fact that Shanghai often attracted migrants
with higher human capital relative to its native population.
43
Figure A-5: Human Capital Shock
I Evidence from Timing
I.1 Pre-1900 Access to Basic Education
We construct our time-varying access to basic education variable using two methods: for 19001960, we obtain primary school enrollment rates from national and provincial censuses on modern
education. For 1820-1900, there were no primary schools in the modern sense, but a percentage
of people did acquire basic education as explained in Appendix D, but data on this are generally
not available. Below we explain our method in estimating pre-1900 rates of access to basic
education and how it affects the validity of our panel analysis.
We estimate pre-1900 %access to basic education from time-varying pre-1900 rates of collegeequivalent education (juren), college schools rate in 1947, and literary rate in 1947. While the
content of provincial-exams was vastly different from modern university education, successful
candidates of provincial exams were similarly at the top of the distribution of the educated.
This yields a method to construct a rough estimate of the share of population with a basic
education prior to 1900. We employ the following parallel projection to estimate time-varying
rates of receiving basic education:
%basic education accessi,1947 = a + b ∗ college student per 10,000i, 1947 + µit .
(10)
Second, using the obtained a and b (1.72 and 13.56 respectively), we project %access to basic
44
education in the exam era from juren per 10,000,
%basic education accessi,t = a + b ∗ ft ∗ juren per 10000i,t ,
(11)
where ft is a time-varying ratio that transforms juren per 10,000 into college students in 1947.
For example, in 1820 and 1850, the proportion of juren is 0.03% of male population in Lower
Yangzi and in 1880 the proportion of juren is 0.06% (juren quota changed little but population
halved after 1860), whereas college students account for 0.3% of total population in 1947. So
the ft =10 for 1820 and 1850, and ft =5 for 1880.
We should point out that this method does assume a fixed ratio between advanced and basic
education in 1947 and in the 19th century. We want to point out that in our panel, our results
are unlikely affected by differential advanced-to-basic education ratios in 1947 and in the 19th
century. The difference between two ratios will be captured by our county fixed effects.
It should be noted that official 1900-1950 schools rates only comprise schools at modern schools
that provided western-style education and was open to the public. Comparing the primary
schools rate in the decade of 1900, and the rate of basic education in the period just before that,
we can estimate the differential response to modern basic education in relation to migrant-native
cultural distance, holding constant a county’s past coverage of basic education, and factors
common to all counties in modernizing their education. Due to data limitations, we cannot
track the evolution of traditional schools during the same period.19 Therefore, we do not know
whether the gap left by the differential response to modern education was fully compensated by
additional provision of traditional education.
19
In 1935, and in 1949, both Zhejiang and Jiangsu collected data on and facilitated the registration of traditional schools. Unfortunately, data quality was low and varied greatly from county to county. By 1953, all
traditional schools had been converted to formal public schools. Our basic education measure captures the
amount of county- and community-level public and semi-public education. Schools initiated by individuals and
clans, but did provide education to the wider population, are not excluded from the measure.
45