URBAN AND CLUSTER AGGLOMERATION ECONOMIES'S EFFECTS ON
RURAL HOUSEHOLDS IN ASIA
By
Chaoran Hu

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Agricultural, Food, and Resource Economics - Doctor of Philosophy
2017

ABSTRACT
URBAN AND CLUSTER AGGLOMERATION ECONOMIES'S EFFECTS ON
RURAL HOUSEHOLDS IN ASIA
By
Chaoran Hu
Agglomeration effects play important roles for rural households in participating in
farm and nonfarm activities. With the rapid growth of cities of different sizes and the
development of food value chain, how these agglomerations of urban effects, networks,
and food value chain clusters will affect rural householdsâ participation in nonfarm
employment and farm behavior (technology adoption) are not yet well known. The
dissertation consists of three chapters that aim to assess the impacts of these urban and
cluster agglomeration economiesâ on rural households in Asia.
Chapter 1 links urban proximity, agricultural performance, and nonfarm
employment of rural households in China to study how the distance to cities of different
sizes will impact rural households income diversifications. I calculate unique composite
distance measures and find that proximity to cities of different sizes impacts rural
householdsâ nonfarm employment differently.
Chapter 2 is an extension of essay 1 and studies how networks will impact ruralurban migrantsâ destination choice in Guangdong, China. I calculate the origindestination paired networks using two methods and find that network effects play positive
roles in choice of cities for rural-urban migrants in China. The results also emphasize the
importance to include migration costs and to separate physical migration costs and
network effects in rural-urban domestic migration studies.

Besides the agglomeration of urban effects on rural households nonfarm
employment, the agglomeration of actors within food value chain, particularly in the rural
areas, will also affect rural households farming decisions. Chapter 3 explores value-chain
clusters and aquaculture innovation in Bangladesh. I test whether the clustering, and thus
economies of agglomeration with implied lower transaction costs, encourage and
facilitate farmers to innovate. I use a unique data set from our own primary survey of the
aquaculture value chain in Bangladesh and calculate an index to include both horizontal
agglomeration and vertical interconnections among actors in the value chain. I find that
being in an area with a high clustering index is associated with a higher probability of
farmers using more modern inputs and growing non-traditional commodity fish species,
controlling for farmersâ other characteristics as well as proximity to cities.

Copyright by
CHAORAN HU
2017

ACKNOWLEDGMENTS

I would like to express my sincere gratitude and thanks to my major professor,
Thomas Reardon, for his guidance and support in my graduate studies. His commitment
and enthusiasm to research has inspired me a lot and his creative ideas have expanded my
horizons. It is a great pleasure and honor to work with him. I am grateful to my
dissertation committee members, Songqing Jin, Jeffrey Wooldridge, and Joseph Herriges,
for their constructive advice, comments and support.
Also, I would like to thank a group of people providing valuable advice and
support for my work. Thank Ricardo Hernandez, Ben Belton, Xiaobo Zhang, and
Shahidur Rashid for their insights and advice on fish value chain project. Thank Kevin Z.
Chen, Duncan Boughton, and David Tschirley for their support in my graduate studies.
I am also thankful to the graduate students, faculties, and staff in the Department
of Agricultural, Food, and Resource Economics at MSU. I wouldnât have made it without
them.
Last, thank my dear family in China and USA for always being considerate and
caring!

v

TABLE OF CONTENTS

LIST OF TABLES ........................................................................................................... viii
LIST OF FIGURES ........................................................................................................... xi
INTRODUCTION .............................................................................................................. 1
CHAPTER 1: LINKING PROXIMITY TO SECONDARY CITIES VS. MEGA CITIES,
AGRICULTURAL PERFORMANCE, AND NONFARM EMPLOYMENT OF RURAL
HOUSEHOLDS IN CHINA ............................................................................................... 4
1. Introduction ..................................................................................................................... 4
2. Survey, sample, and variables ......................................................................................... 7
2.1. Farm household survey and sample ......................................................................... 7
2.2. City classification and sample ................................................................................. 8
2.3. Definition of urban proximity ................................................................................ 10
3. Descriptive statistics ..................................................................................................... 13
3.1. Distances of villages from cities of different sizes ................................................ 13
3.2. Sample characteristics ............................................................................................ 15
4. Model and estimation methods ..................................................................................... 18
4.1. Model ..................................................................................................................... 18
4.2. Estimation issues and methods .............................................................................. 20
5. Empirical results ........................................................................................................... 22
5.1. Impacts of urban proximity .................................................................................... 22
5.2. Interacted effects from urban proximities and agricultural performance .............. 27
5.3. Robustness check ................................................................................................... 29
5.3.1. Different definitions of nonfarm employment ................................................ 29
5.3.2. Estimations with additional controls............................................................... 31
6. Conclusions ................................................................................................................... 32
APPENDIX ....................................................................................................................... 34
REFERENCES ................................................................................................................. 57
CHAPTER 2: NETWORK EFFECTS AND RURAL-URBAN MIGRATION
DESTINATION CHOICES: EVIDENCE FROM CHINA.............................................. 62
1. Introduction ................................................................................................................... 62
2. The model ..................................................................................................................... 65
3. Data ............................................................................................................................... 66
3.1. Data and sampling.................................................................................................. 66
3.2. Characteristics of the sample ................................................................................. 67
4. Estimations.................................................................................................................... 68
4.1. Networks ................................................................................................................ 68
4.2. Calculation of counterfactual incomes................................................................... 70
4.3. Migration cost ........................................................................................................ 72
4.4. Amenities of cities ................................................................................................. 73

vi

4.5. Weighting used to correct for on-site sampling ..................................................... 73
5. Estimation results .......................................................................................................... 74
5.1. Impacts of network effects ..................................................................................... 74
5.2. Impacts of the migration cost ................................................................................. 76
5.3. Results with individual characteristics interactions ............................................... 77
5.4. Results with city amenities interactions ................................................................. 80
6. Conclusions ................................................................................................................... 81
APPENDIX ....................................................................................................................... 83
REFERENCES ................................................................................................................. 99
CHAPTER 3: VALUE-CHAIN CLUSTERS AND AQUACULTURE INNOVATION
IN BANGLADESH ........................................................................................................ 102
1. Introduction ................................................................................................................. 102
2. Background on Aquaculture in Bangladesh ............................................................... 107
3. Survey and sample ...................................................................................................... 109
4. Clustering index and its patterns ................................................................................. 112
4.1. Calculation of clustering index ............................................................................ 112
4.2. Clustering patterns ............................................................................................... 115
5. Regression estimations................................................................................................ 117
5.1. Regression specification ...................................................................................... 117
5.2. Econometric issues and their resolution .............................................................. 120
5.3. Descriptive results ................................................................................................ 122
6. Results of Regressions ................................................................................................ 125
6.1. Adoption of modern inputs .................................................................................. 125
6.2. Adoption of nontraditional commodity and newly commercialized aquaculture
niche species ............................................................................................................... 127
6.3. Specialization versus diversification of fish production ...................................... 128
6.4. Robustness check ................................................................................................. 128
7. Conclusions ................................................................................................................. 129
APPENDIX ..................................................................................................................... 133
REFERENCES ............................................................................................................... 151

vii

LIST OF TABLES

Table 1.1. Distribution of cities and population statistics by size of cities ....................... 36
Table 1.2. Distance measures and village level descriptive statistics ............................... 37
Table 1.3. Distribution of sample villages by measure of distance .................................. 38
Table 1.4. Distribution of incomes of sample households, by years and sources............. 39
Table 1.5. Descriptive Statistics of Family Members (Individual Level)......................... 40
Table 1.6. Distribution of nonfarm income of sample households, by year and location 41
Table 1.7. Descriptive Statistics of other variables........................................................... 42
Table 1.8. Impacts of urban proximities on levels of income, by sources and by six
methods of measurements of distance .............................................................................. 43
Table 1.9. Impacts of urban proximities on income shares, by sources and by six methods
of measurements of distance ............................................................................................. 44
Table 1.10. Impacts of local agricultural performance (AP) and urban proximities on
migration income levels and shares, by three methods of measurements of distance ...... 45
Table 1.11. Robustness check using different definitions of local nonfarm income levels
........................................................................................................................................... 46
Table 1.12. Robustness check using different definitions of Local-Migration income
levels ................................................................................................................................. 47
Table 1.13. Robustness check using different definitions of Migration income levels and
shares................................................................................................................................. 48
Table 1.14. Robustness check, including additional controls in the estimations .............. 49
Table 1.A1. Full results of impacts of urban proximities on local income values, by
sources and by six methods of measurements of distance ................................................ 51
Table 1.A2. Full results of impacts of urban proximities on local-migration income
values, by sources and by six methods of measurements of distance ............................... 53
Table 1.A3. Full results of impacts of urban proximities on migration income values, by
sources and by six methods of measurements of distance ................................................ 55
Table 2.1. Summary statistics of demographic variables, Individual level ...................... 84

viii

Table 2.2. Average incomes and rents per month ............................................................. 85
Table 2.3. Origin-destination paired networks (%), by cities and measures .................... 86
Table 2.4. Distribution of sample migrants by types ........................................................ 87
Table 2.5. Probability of individual from region r to city j, by education levels .............. 88
Table 2.6. Imputed incomes for all observations by cities ............................................... 89
Table 2.7. Travel distance and time, by destination cities ................................................ 90
Table 2.8. Summary statistics of city attributes, city level ............................................... 91
Table 2.9. Distribution of sample migrants in nine destination cities .............................. 92
Table 2.10. Estimation results of Network I on destination choice, without travel time.. 93
Table 2.11. Estimation results of Network II on destination choice, without travel time 94
Table 2.12. Estimation results of networks on destination choice, with travel time ........ 95
Table 2.13. Estimation results of networks and education selections............................... 96
Table 2.14. Estimation results of networks and micro-level nonfarm experience............ 97
Table 2.15. Estimation results with networks and city amenities interactions ................. 98
Table 3.1. Farmed fish output of sample households, by species ................................... 134
Table 3.2. Distribution of size categories of different types of actors in the study areas 135
Table 3.3. Clustering in the 20 districts in four cardinal-points zones ........................... 136
Table 3.4. Descriptive statistics of household and district characteristics ...................... 137
Table 3.5. Patterns in fish farmersâ adoption of technology innovations ....................... 138
Table 3.6. Regression of clustering degree on adoption of modern inputs..................... 140
Table 3.7. Regression of clustering degree on per acre expenditures on modern inputs 141
Table 3.8. Regression of clustering degree on output share, by species ......................... 142
Table 3.9. Regression of clustering degree and urban proximity on modern inputs
adoption, by product-type farmers .................................................................................. 143
Table 3.10. Regression of clustering degree on specialization ....................................... 144
Table 3.11. Regression of clustering degree and urban proximity, by different samples145
ix

Table 3.A1. Fishes that mapped into each specie category ............................................ 149
Table 3.A2. IV regression of clustering degree on specialization, by IV methods ........ 150

x

LIST OF FIGURES

Figure 1.1. Trend of income structure of rural households in China, 1990-2011............. 50
Figure 3.1. 20 sample districts in six major aquaculture areas of Bangladesh ............... 146	 Â 
Figure 3.2. Distribution of actors per 1,000 rural people in 2013 by district ................. 147	 Â 
Figure 3.3. Clustering degrees of Sample Districts, 2013 .............................................. 148	 Â 

xi

INTRODUCTION

Agglomeration effects play important roles for rural households in participating in
farm and nonfarm activities. On the one hand, the geography and migration literature has
summarized that the urban proximity could provide push effects to increase the rural
labor supply from farm to nonfarm employment. On the other hand, the industrial
organization literature has shown that network as well as clustering effects could help the
actors within the clusters to benefit from agglomeration externalities. However, the gaps
remained in the literature motivated the goal of this dissertation.
The first gap remained in the literature is that little is unknown whether distance to
cities of different sizes would impact rural householdsâ nonfarm employments
differently. In the first chapter, I link urban proximity, agricultural performance, and
nonfarm employment of rural households in China to study how the distance to cities of
different sizes will impact rural households income diversifications. To capture the
potential impacts from all cities I calculate unique composite distance measures. I then
analyze the distance measuresâ impacts on rural employment. Several main findings stand
out. First, proximity to cities of different sizes impacts rural householdsâ nonfarm
employment (including agricultural wage) versus agricultural employment (own farming
only) differently. Specifically, proximity to smaller cities increases both income levels
and shares in total income from agriculture and local nonfarm activities, while proximity
to mega cities increases nonfarm income from migratory activities. Second, rural
households seeking nonfarm employment undertaken at shorter distances (within their
province) or by commuting tend to be attracted by service sectors in cities of medium

1

size. By contrast, the agglomeration effects and manufacturing sectors in smaller cities
are important to householdsâ undertaking rural local nonfarm employment.
Manufacturing in mega cities induces rural households to undertake out- of-own province
migration. Third, in contrast to the common view, local agriculture performance spurs
migration by rural households, in particularly for people living far from mega cities.
The second chapter is an extension of the first chapter. This paper studies how
networks in the destination city will impact rural-urban migrantsâ location choice.
Existing studies calculate the network effects from a relatively narrow aspect, the impacts
of which may be limited and even decrease over time. In this paper, I calculate the origindestination paired networks using two methods, and study their impacts on rural-urban
migrantsâ destination choices in Guangdong, China. Filling in the gaps by allowing
migration cost being varied by location, and calculating counterfactual expected incomes
to account for selection concerns, I provide another evidence on education-related
selection to the migration literature with mixed results. The results show that network
effects play positive roles in choice of cities for rural-urban migrants in China, and I find
positive selection of networks in terms of education. The results emphasize the
importance to include migration costs and to separate physical migration costs and
network effects in rural-urban domestic migration studies.
Besides the agglomeration of urban effects on rural households nonfarm employment,
the agglomeration of actors within food value chain, particularly in the rural areas, will
also affect rural households farming decisions. Chapter three explores value-chain
clusters and aquaculture innovation in Bangladesh. Farmers adopting and implementing
innovations, such as new technologies and new products, often require âcollaborative

2

inter-segment innovationâ by other actors in other segments of the value chain, such as
wholesalers implementing new product innovations such as supply of commercial fish
feed and chemicals and buying and marketing non-traditional fish species. I test whether
the clustering, and thus economies of agglomeration with implied lower transaction costs,
encourage and facilitate farmers to innovate. That potential determinant of farmer choices
has not been studied in agriculture or aquaculture or indeed the food sector. I use a
unique data set from our own primary survey of the aquaculture value chain in
Bangladesh, including micro data for 1500 fish farm households and 20 districts for meso
level data. I calculate an index to include both horizontal agglomeration and vertical
interconnections among actors in the value chain. I find that being in an area with a high
clustering index is associated with a higher probability of farmers using more modern
inputs and growing non-traditional commodity fish species, controlling for farmersâ other
characteristics as well as proximity to cities.

3

CHAPTER 1: LINKING PROXIMITY TO SECONDARY CITIES VS. MEGA
CITIES, AGRICULTURAL PERFORMANCE, AND NONFARM
EMPLOYMENT OF RURAL HOUSEHOLDS IN CHINA

1. Introduction
Nonfarm income from manufactures or services, in rural or urban areas, is important
to rural households in developing countries. The share of nonfarm income over total rural
household income averages 34%, 51%, and 47% in developing Africa, Asia, and Latin
America, respectively (Haggblade et al., 2007). The National Bureau of Statistics of
China notes that the share of nonfarm income in total income of rural households has
increased from 34% to 52% in one decade, from 2000 to 2010 (see Figure 1.1).
Three strands of literature have explored the determinants of rural householdsâ
participation in nonfarm activities.
A first strand starting in the 1950s focused on decisions of rural households (who
were considered to be autarchic farmers before migrating) to migrate to urban areas and
internationally. The strand conceived of the drivers of migration as the relative
rural/urban wage (Lewis, 1954), the expected urban wage and employment (Todaro,
1969); Harris and Todaro, 1970), and perceptions of risks and capacities to migrate (in
general, Stark and Lucas, 1988; for China, Rozelle et al., 1999; Mu and Giles, 2014; and
Du et al., 2015)
A second strand of literature focused on activity portfolio selection by rural
households of local rural nonfarm employment, farm work, and migration (e.g., Hymer
and Resnick, 1969, in general, and Barrett et al., 2001, Yunez-Naude and Taylor, 2001,
for Africa and Mexico; and for China, examples are Zhao, 1999a, and Zhao, 1999b).

4

This literature modeled the determinants of this multisectoral, multi-spatial activity
choice as a function of labor-supply origin-zone characteristics (such as rural
infrastructure, and local agriculture performance which can create demand for local
nonfarm employment via production and consumption linkages (Mellor and Lele, 1973).
A third strand of literature introduced spatial determinants into the modeling of local
and migratory employment. This strand started with von ThĂźnen's (1826) observation
that different degrees of proximity to an urban center affect the spatial pattern of
agricultural activities as a function of transaction costs and factor costs. Recently this was
extended to model the spatial distribution of sectoral activity (e.g., Desmet and
Fafchamps, 2005, for the US) and rural householdsâ participation in nonfarm
employment as a function of distance to the nearest urban center (e.g., Lucas, 2001,
Fafchamps & Shilpi, 2003, for Nepal; De Janvry et al., 2005, for China; and Fafchamps
et al., 2016 for Ghana). Some work has crossed distance to the nearest city with
agricultural performance of the rural householdâs zone (Deichmann et al., 2009, for
Bangladesh).
However, I perceive a gap in the third strand of literature. Gains in understanding of
the impact of city distance on rural household labor choices could be made by
differentiating distance to different sizes of cities, rather than just a city per se. This has
not been done in the literature and is addressed by the present paper. The justification for
the importance of the gap is made in a general way in the work of Christiaensen and
Todo (2014) and Christiaensen et al. (2017) in Africa and BerdeguĂŠ et al. (2015) in Latin
America. They show that the distinction between mega cities and secondary cities is
important in terms of how the city influences the rural hinterlandâs poverty rates and

5

territorial development, respectively. Both works argue that secondary cities may have
more direct dependence on and interactions with rural areas, and secondary versus mega
cities have different draw potentials for rural migrants. For the US, Partridge et al. (2008)
found stronger effects of proximity to larger urban centers compared with small urban
centers on population growth of U.S. hinterland counties.
But one can go further, and I do in this paper, than just distinguishing distance from
different sizes of cities. I also bring in the basic concept from Deichmann et al. (2009) by
crossing the distance/city size effect with agricultural performance of the labor-origin
zone. This cross also has not been done before in the literature. Agricultural performance
conditions the degree of pull effect of a city on a rural household by either being so risky
or poor that it encourages through a push effect the household to send migrants or
commuters to work in the city (given the distance), or it is so good that it creates a pull to
stay in the farm area for direct farm employment or employment in local nonfarm
activities in consumption linkages or production linkages to local agriculture. The idea
of including and differentiating rural agricultural performance by interacting it with
distance to cities is the same as that done in Henderson and Wang (2005). In their model,
they allow for technological improvement in agriculture. Thus, agricultural performance
of the rural areas may not only enter as a push effect for labor to migrate from the rural
area but also a pull effect for rural-urban migration, and these effects may be impacted by
surrounding urban sectors.
Besides modeling the impact on rural household labor choices of the distance of cities
of different sizes crossed with local rural agricultural performance, I aim to make a
methodological contribution. Most studies that include a measure of the distance to urban

6

areas tend to measure urban proximity as the shortest distance to a nearby city, or as the
distance to a largest city. Such studies do not, however, take into account the potential
effects of proximity to other cities beside the nearest city. I thus measure urban proximity
as a composite distance by summing the distances to all cities of different size categories
with two different weights (population and sectoral GDPs) that are surmised to have
impacts on household decisions. To my knowledge no other study uses this composite
measure.
We examine the above research questions using data from five provinces in a
balanced two âyear panel, 2005 and 2006, from a nation-wide survey collected annually
by Chinaâs Research Center for the Rural Economy (RCRE) of the Ministry of
Agriculture.
The paper is organized as follows. Section 2 discusses the survey, sampling, and
definitions of key variables. Section 3 provides descriptive statistics on nonfarm activities
of rural households in the sample. It also shows distances of survey villages from cities of
different sizes, depicting the different measures used in this paper. Section 4 presents the
conceptual and empirical approach. Section 5 shows the regression results. Section 6
concludes.

2. Survey, sample, and variables
2.1. Farm household survey and sample
The data sets come from a nation-wide survey collected annually by Chinaâs
Research Center for the Rural Economy (RCRE) of the Ministry of Agriculture. The
survey sample method was as follows. In each province, RCRE stratified counties into

7

three groups by income level. In each selected county, villages were randomly selected;
the number of villages varied by the size of the province. Per village, 40-120 households
were randomly selected; the number varied by the size of the village (Benjamin et al.,
2005).
We use data from five provinces (Heilongjiang, Shandong, Sichuan, Hunan, and
Jiangxi) from the 31 RCRE annually surveyed provinces. I chose these five because they
represent a diversity of rural settings: intensive horticulture, plains rice, mountainous
mixed-product farming. I use two years of data, 2005 and 2006. These were selected to
predate the 2008 earthquake, which affected Sichuan deeply, and the 2007-08 financial
crisis, which affected all the provinces. Also, sample villages in the RCRE survey were
changed after 2006, so using 2005 and 2006 allows me to have a two-year balanced panel
of households. For each year, the sample consists of 2,560 rural households from 59
villages.

2.2. City classification and sample
In China, a province is divided into several âprefectural cities,â the second level
administrative division, also called the capital city of the prefecture. The country is
divided into 344 prefectural âcitiesâ each of which has an urban center as well as
surrounding areas that are further divided into several counties. Each county has a
county center and the surrounding rural areas that are divided into townships. Typically,
the urban center of a âprefectural cityâ is much bigger than a county center inside the
same prefectural city. In this study, I use âprefectural cityâ as a proxy for âcity,â since the
data from county-level cities are very limited (Au and Henderson, 2006). To include the

8

potential impacts from all urban areas, I take all prefecture level cities into consideration.
The size of the city is measured as the total urban population of the âprefectural cityâ.
Only the registered urban population is included; the registration is urban âhukou,â a kind
of citizenship certificate for urban residents.. So in a sense, the city in this study is a
collection of urban areas (urban center and the county centers) within a prefectural city,
not a city in a strict sense.
While this is not an ideal measure of a city, I believe it is a reasonable proxy. First,
migrants typically answer with the name of a âprefectural cityâ when they are asked
about their migration destination (e.g., Dongguan, Guangzhou, Shenzhen and Wenzhou,
prefectural cities, are among the most common migration destinations).1 Second, it is
reasonable to argue that the core urban center of a prefecture city is a more populated
area and plays a more important role in attracting migrants. Evidence shown in Fan and
Scott (2003) suggests a positive correlation between a large city region, industrial
agglomeration, and economic performance, and thus attraction for rural migrants. The
issue for the measure however is that the county level cities are left out of the analysis.
Therefore, to control for the impacts from lower (e.g. county) level cities, I include the
distance from the sample village to the nearest railway and long-distance bus station,
which tend to be located in county and township level cities.
Chinese cities are officially classified into five types by resident population; e small
(<0.2 million); medium (0.2-0.5 million); large (0.5-1 million); extra-large (1-5 million);

1

Using the urban population of prefectural level cities in China to reflect the city

proportion is typically used in the literature. See for example Au and Henderson (2006).

9

and extra-extra-large (>5 million) with the last category being added in 2014.2 This
resident population includes rural migrants staying in the city for more than six months.
The latter is endogenous. Thus, to get an exogenous definition of city size, I use the
population of permanent urban residents (hukou registered urban population). My
definition is more exogenous as it is not temporary, since âhukouâ is passed from
mothers to children (Li and Gibson, 2013). The distribution of these cities by both
official classifications (the five noted above) and my own classification of city sizes are
shown in Table 1.1.
To obtain comparable sizes of sets, I made some combinations. I combined the
officially defined set of small and medium cities into one category comprising 152 cities,
what I call the âsmaller-cityâ group. The latter range from 0.03 million to 0.99 million,
but only 20% of this group have fewer than 0.3 million (and thus would be considered
âtownsâ). I also group the officially defined extra-large and extra-extra-large cities into
what I call a âmega-cityâ group since there are few cities in this group. According to my
definition, overall, 44% of the cities in China are in the âsmaller citiesâ group, but their
total population is only 15% of the overall urban population.

2.3. Definition of urban proximity
We measure urban proximity as the spherical distance (shortest distance between two
points on the earth) from the center of the village to the center of a city. This is calculated
using QGIS software. I do not use travel distance, since it will be endogenous in my case,
and it is hard to find a valid instrumental variable. For example, the construction of roads

2

http://www.gov.cn/zhengce/content/2014-11/20/content_9225.htm

10

between villages and cities is due partly to economic or political factors, such as to attract
the settlement of firms in the suburban and rural areas, which will also affect rural
households' participation in and earnings from nonfarm activities. (Also, as mentioned in
Volpe Martincus et al. (2017), even using historical roads would be hard to meet the
exclusion restriction required by a valid instrument variable.)
We examine the effects of six categories of distance variables. The first three types of
distance are âunweighted physical distancesâ. The first is the one most commonly used
in the literature (such as in Escobal, 2001, the distance to the nearest city/market,
regardless of the size of city). Apart from the nearest city, researchers also use the
distance to the countyâs capital city (de Janvry et al., 2005) or provincial capital city
(Gibson, 2000) or largest city of the country (Deichmann et al., 2009). The latter citytype specific measures assume that the pull effects from these cities are higher than other
cities near to the household.
The second category is the distance to the nearest city for each of three sets of cities
(âsmaller,â âlarger,â and âmegaâ cities). This measure has been used for example by
Jonasson and Helfand (2010).
The third category is called the incremental distance method, used in Partridge et al.
(2008). It is based on the assumption that cities in the higher tier (for example mega
cities) offer goods and services that are not available in the lower tier (for example
smaller cities). The distance to the nearest city of any size is the first distance controlled
for. Then, the additional distance to the nearest city in the next highesttier is controlled
for, and finally, the distance to the nearest city in the highest tier is controlled for.
Incremental distances to larger and mega cities may be equal to zero.

11

However, results from Fafchamps and Wahba (2006) and de Janvry and Sadoulet
(2001) imply that the above nearest distance methods may ignore the potential impacts
from other cities (beyond the nearest but within a certain perimeter). Therefore, I
introduce three more measures to capture the potential effects from all cities in the
country on any rural village (and household). But unlike Fafchamps and Wahba, and de
Janvry and Sadoulet, I do not limit the maximum travel time, for three reasons. First, in
China, several cities, in particular larger and mega cities, even when they are far from a
village, may also have some impact on nonfarm employment of rural households in that
village. For example, several cities in Guangdong province (such as Shenzhen and
Dongguan), instead of the nearest or capital city (Guangzhou), are attracting lots of rural
migrants. Second, limiting the travel time or distance by any size is too subjective, as it
might exclude cities potentially important to a certain village. Third, massive investments
in trains and highways in the past several decades have made it practicable and even
relatively cheap for a villager to reach any major city in China.
My measures also contain distance variables to each of the three city categories
(smaller /larger/mega cities). However, instead of including a single distance from the
sample village to the nearest city in a given category, I calculate the composite distances
using all (near and far from the village) the cities in a given category.
The composite distances are calculated using different weights. In the fourth measure
the distances are weighted by the populations of cities (with the weight being the share of
a cityâs population in total urban population). In the fifth measure the distances are
weighted by GDP from the secondary sector (manufacturing and construction, which for
simplicity I will just call manufacturing hereafter). The sixth measure weights distances

12

by GDP from the tertiary sector (which I term the service sector hereafter). The GDP
weights are the GDP of a city in the aggregate GDP of cities for that sector. I include the
fifth and sixth measures by GDP to give greater weight to cities with higher GDP values
from manufacturing and service sectors.
To show how I have done the weighting, take for example the measurement of
population-weighted distance.
The composite distance from village j to nine mega cities = Â 
where đ¤đđđâđĄ! =

đđđđ˘đđđĄđđđ!

!
!!! đˇ!"

â đ¤đđđâđĄ!  Â 

!
!!! đđđđ˘đđđĄđđđ!

The measures of distance to other type of cities and distance weighted by GDP by
sectors are similar. All methods used to calculated distance are at village level. Table 1.2
summarizes the statistics of the six measures.

3. Descriptive statistics
3.1. Distances of villages from cities of different sizes
The spatial distribution of villages (in terms of distances from various types of cities)
is shown in Table 1.3. The first two columns in Section I (the horizontal section) show
the distribution of villages by the shortest distance to a nearby city without regard to the
size of the city (M1). The remaining columns of that table section show the distribution
of villages by the shortest distance to a nearby city by size (smaller, larger, and mega
city) (M2). The first part of section II of Table 1.3 shows the distribution of villages by
distance to all cities by size, with the distance being weighted by the population of the
city (M4). The following two parts show the distribution of villages by distance to all
cities by size. The distance is weighted by GDP from the manufacturing sector (M5) and

13

from the service sector (M6). The distribution of villages from the third measure (M3) is
explained in the notes to the Table.
Several points stand out in Table 1.2 and Table 1.3. First, contrary to conventional
wisdom that rural people live far from cities, 80% of the sample villages are only 80
kilometers away from their nearest city of any size. All sample villages are within 160
kilometers, a few hours, from some city. As one moves over the distance categories from
0-40 to 120-160 it is interesting to see that there is no clear pattern of city size at the
various distances, and all four are in a rather tight range averaging from 1 to 2 million.
Second, compared to the average distance to the nearest smaller cities (119km), more
sample villages are closer to a larger city (81km). Yet 86% of the sample villages do not
have a ânearestâ mega city within 160 kilometers.
Third, once I relax the nearest-city condition and include all cities, the populationweighted distance shows that almost all sample villages can access a larger city or mega
city within 2500 kilometers, yet over 10% of sample villages cannot reach a small city
within 2500 kilometers. The reason for this is that as I include all cities and group them
into three categories by size (recall Table 1.1), a larger number of cities with larger sizes
but still within the small city category are far from villages compared with smaller cities
within the small city category.
Fourth, the weighted composite distances (M4, M5, and M6) show that a high share
of villages are within 1000 kilometers of larger cities (27% in M4, 39% in M5, and 36%
in M6) and mega cities (47% in M4, 49% in M5, and 41% in M6).

14

Fifth, results of the incremental distance measure show that 36% of the sample
villages have a ânearest cityâ belonging to the category of smaller cities, and 63% have a
ânearest cityâ belonging to the category of larger cities.

3.2. Sample characteristics
Table 1.4 shows household income composition. Note that average household income
increased by 9% from 2005 to 2006 (compared with overall GDP/capita growth, of 11%).
The income from own-farming decreased slightly (2%), while nonfarm incomes
(including agricultural wage earnings) increased by 21%. The average rural household
has income very diversified beyond agriculture: in the two years analyzed, agriculture is
only 44% and 39% of household income, respectively.
Table 1.5 shows for each of the two years the decomposition of nonfarm employment
by location using individual (not household) level data with observations about the
location of employment and how many days each family member worked outside the
home and stayed at home. Several interesting points emerge.
First, the results are consistent with the main body of literature that finds a high
participation rate of rural labor in nonfarm employment in developing countries. The data
show that 74% in 2005 (and 76% in 2006) of rural labor participated in some type of
nonfarm employment, compared with slightly lower for participation in own-farming
(71% in 2005, 68% in 2006).
Second, while much of the literature on nonfarm employment in China focuses on
out-of-province migration, my results show that only 27% of people participating in some
type of nonfarm activity is working in another province. Among 83% of people doing

15

nonfarm activities within their own province, 57% of them even undertake the nonfarm
activities within own village, and 43% of them work outside their own village but within
their own province.
Even for people doing nonfarm activities outside of their own villages, a substantial
share of labor works locally. For instance, 39% of rural people work within their own
county (22% in another village but within their own township, 17% in another township
but within their own county), 21% of them work out of their own township but within
their own province, and only 40% of them work in another province.
Based on these results, I classify nonfarm income into three categories. The first is
âlocalâ. This is income earned from nonfarm activities within oneâs township. The
second category is âlocal migrationâ, being income earned from nonfarm activities
outside of oneâs own township but within the province. The third category is âmigrationâ,
being income from activities outside of oneâs own province.
Table 1.6 shows nonfarm income composition based on location. Several points are
salient.
First, local nonfarm income is substantial: the share of local nonfarm income (36%)
is similar to the share of migration income (38%) in the householdsâ total nonfarm
income. If I add income earned in another township in the householdâs own county as
local nonfarm income, then local nonfarm income is 49% of total nonfarm income - 13%
higher than migration income.
Second, Table 1.5 shows that 61% of nonfarm laborers are employed within their
own township, while Table 1.6 shows that only 36% of total nonfarm income is from
within-township employment. The results suggest that participation is widespread

16

perhaps because entry barriers and capital requirements for this local nonfarm
employment are low â but the earnings are also relatively low.
Third, local-migration income is also substantial. While only 17% of nonfarm rural
laborers have jobs outside their township and within their province, income from these
jobs forms fully 26% of total nonfarm income.
Table 1.7 shows the characteristics of sample households and villages. Household
heads are late middle aged, and almost all are male. The highest education is 9 years
(after middle school). The farms are small. The average size of owned land under
cultivation per household is 6.2 mu (0.41 hectare). Households are generally close to
transport facilities: the mean distance between the sample villages and a paved road is 1.5
km, to the nearest railway station, 38 kilometers, and to the nearest long-distance bus
station, 9.4 kilometers.
A villageâs agricultural performance is posited to be a critical factor affecting the
nonfarm economy, both locally and non-locally. This occurs either by pull effects (such
as intersectoral linkages) or push effects (such as poor performance driving farm
households to diversify incomes into nonfarm activities to manage risk or cope with farm
sector shocks). I define a villageâs agricultural performance as the village output in value
terms per mu of all crops. On average, the latter is around 750 RMB per mu ($1125/ha at
the 2006 exchange rate). But there is substantial variation over villages. Poorer
performance is a push factor in local rural households seeking nonfarm employment. One
could associate this case with Heilongjiang, with the crop composition focus on lower
value products (rice). By contrast, high performance with its pull factors for nonfarm

17

employment can be associated with Shandong, with its high share of high value cropping
(horticulture in particular).
Some other village level variables that are hypothesized to affect rural householdsâ
labor choices are included, such as the degree of development of the local land rental
market. The existing evidence shows that households are more likely to obtain their main
income from migration if the land rental market is developed so they can migrate and rent
their farmland out while they are away (Deininger and Jin, 2006, Jin and Deininger,
2009). To avoid the potential reverse causality of nonfarm incomes on the householdâs
land renting decision, I use a proxy for the village-level land rental market in the controls.

4. Model and estimation methods
4.1. Model
The analysis spatializes a labor supply function of a rural household with n types of
labor characterized by geography and sector. The explained and determinant variables are
as follows.
The explained variables are quantities of labor supply to various locations and
sectors. Following Barrett, Reardon, & Webb (2001), the relevant n types are a cross
between spatial distribution of the work performed by the rural household (local versus
migratory employment) and the sectoral distribution (agricultural versus non-agricultural
employment). Non-agriculture includes services and manufactures, which I lump in the
analysis of employment. The latter could be further disaggregated into the functional
distribution (self-employment versus wage employment) but I lump those.
The determinants in a labor supply function are generally incentive variables (such as
relative urban/rural wages) and capacity variables (such as education). I apply that
18

general framework with specific incentives capturing urban employment âpullâ by
distance to various size categories of cities controlling for access to earnings from ownfarming and intersectoral linkage employment by including village âagricultural
performanceâ measured as crop income per acre. The capacity variables are farm assets
(such as land). .
The basic econometric model used is an application of the above:

Yijt =Îą + đ˝1Smallj + đ˝2Largej + đ˝3Megaj +đ˝4 APjt + Î´Zijt + Îłyear + Ďprovince +vj + cij + Îľijt ,

(1)

The specific variables are shown in the Table 1.A1, and summarized here. The Yijt are
either the level of nonfarm income or the share of nonfarm income in total household
income of household i in village j in year t. Nonfarm income is further divided into local
nonfarm income (earnings from nonfarm activities within the householdâs own
township), local-migration income (outside of their own township but within their own
province) and migration income (from outside their own province).
The determinants include a set of distance variables, Smallj, Largej, and Megaj,
measuring distances from the center of village i to a smaller city, a larger city, and a
mega city. As noted above, the regressions also explore variations on the simple distances
with several weighted distance measures. Other determinants are as follows. APit is
village-level agricultural performance, measured as the value of total output per acre of
cultivated land in village i. Zjit is a vector of individual, household, and village level
controls, the descriptive statistics for which are shown in Table 1.7. Îłyear are time-fixed

19

effects and Ďprovince are province-fixed effects. vj and cij are village- and householdspecific unobserved characteristics. Îľijt is a stochastic error term.

4.2. Estimation issues and methods
While one can safely assume the exogeneity of the physical distances between the
village center of any given rural household and cities of different sizes3, it cannot be ruled
out that some unobserved household traits (cij), such as ability, risk aversion, and
experience, and village unobserved factors (vj), such as history, shocks, and local
policies, are correlated with household labor supply choices and village-level agricultural
performance. The existence of vj and cij causes the OLS estimates to be biased and
inconsistent. Ideally I would use a panel fixed effects model (FE) to remove vj and cij.
Unfortunately, neither a household FE nor a village FE method can be used to estimate
equation (1) because the distance variables are time-constant.
Hence, I use the correlated random effect (CRE) estimation method, following
Chamberlain (1979) and Mundlak (1978). The CRE method relaxes the assumption of
strong independence of unobserved stochastic errors from the random effects model, and
provides fixed effect model estimators of time-variant variables and random effect
estimators for time-constant variables. Specifically it is assumed,

đ!" =  Â Îťđ Â !"  Â  +  Â  Îą!" , and đŁ! =  Â Îˇđ¨đˇ Â !  Â  +  Â Ďđ Â !" + ÂľÎź!

3

An average rural household typically does not move from one village to another village. Temporary
migration has been the dominant type of migration in China so a large majority of migrants still belong to
the original villages.

20

The CRE model is obtained by substituting cij and vi in equation (1) by mean values
of time-varying household and village level variables. Specifically, the CRE model
estimates the following equation,

đ!"# = đź + đ˝! đđđđđ! + đ˝! đżđđđđ! + đ˝! đđđđ! + đ˝! đ´đ!" + đżđđđđ
+đ¸đđđđ + đ!"#$%&'( + Îťđ Â !" + Îˇđ¨đˇ Â ! +  Â  đ!"# ,

(2)

Equation (2) can now be estimated consistently by OLS. Since the key variables of
interest are measured at the village level and there are multiple household observations
for a given village, I follow Wooldridge (2003) and correct the standard errors by
adjusting for the clustering effect at the village level. The statistical significance of
estimates of Îť and Îˇ would signal the existence of unobservables. The key coefficients of
interest in equation (2) are the Î˛âs.
In a different set of estimations, the first equation is augmented by including the
interaction terms of APjt and three distance variables (smallit, largeit and megait), to test
the potential multiplier effects of village agricultural performance. As in the case of
equation (1), I adopt the CRE model to obtain consistent estimates of the augmented
model. The equivalent CRE model for the augmented model can be written as,

đ!"# = đź + đ˝! đđđđđ! + đ˝! đżđđđđ! + đ˝! đđđđ! + đ˝! đ´đ!" + đ! (đđđđđ! Ăđ´đ! ) +
đ! (đżđđđđ! Ăđ´đ! )+đ! (đđđđ! Ăđ´đ! )+đżđđđđ + đ¸đđđđ + đ!"#$%&'( + Îťđ Â !" +
Îˇđ¨đˇ Â ! + đ! (đđđđđ! Ăđ´đ! ) + đ! (đżđđđđ! Ăđ´đ! )+đ! (đđđđ! Ăđ´đ! ) +  Â  đ!"# , (2â)

21

Like equation (2), equation (2â) can be estimated consistently using OLS. The key
coefficients of interest in equation (2â) are the Î˛âs and Î¸âs.

5. Empirical results
5.1. Impacts of urban proximity
Because of space limitations, the full results for income levels are reported in the
tables in the Appendix. All the estimations include year- and provincial-level fixed
effects, with errors clustered by villages. The results of demographic variables and village
level controls are consistent and robust to different measurements of distances.
Table 1.A1, Table 1.A2, and Table 1.A3 show that the highest education level of
family member, the size of the household, access of the village to a railway and a longdistance bus, and the village-level land rental market are all important to household
income levels. Several results stand out.
First, as expected, more education increases migration income; a bigger household
(and thus more labor available) increases not only migration income but also localmigration income.
Second, the effect of distance to the nearest long-distance bus station on local
nonfarm income is significantly negative, which is robust to different measures/models
(see Table 1.A1).
Third, households in villages with higher shares of rented lands have higher migration
incomes, consistent with land rental market studies (Deininger and Jin, 2005; Chernina et
al., 2014). Rural households in villages with more active land rental markets could rent

22

their lands out easily and thus are less restricted in their labor movements. Also, land
rental may generate cash for households to undertake migration.
The key variables of interest are impacts of urban proximity on income levels and
shares, which are shown in Table 1.8 and Table 1.9, respectively. The results are
generally robust to both income levels and shares using six different measurements of
distances. In Table 1.8, the results using the physical distance measurement are shown in
the first three columns (Models M1, M2 and M3), and the results using the composite
weighted distance measurements are shown in Models M4, M5 and M6.
The distance variables in M1 and the first distance variable in M3 are measured as the
distance to the nearest city of any size. Their impacts are similar. But there are no
significant impacts of incremental distances on local-migration incomes. Recalling the
descriptive patterns in Table 1.3, section M2, I surmise that the reason for insignificance
of the nearest city on local nonfarm income and agriculture income in the regressions
may be because large share of the sample villages have the nearest city as a larger city
instead of smaller city. If the nearest cities of different size categories are considered,
only the nearest one of any size will affect non-local nonfarm income levels and shares.
Rural households close to a city of any size have higher local-migration nonfarm incomes
while having lower migration income in both levels and shares. Thus these results
emphasize the importance of nearest cities on nonfarm employments, which are
consistent to existing findings using nearest cities. Moreover, the results also suggest that
being close to a larger city, when other factors are controlled for, will increase localmigration nonfarm employment. However, from these measures, including nearest cities

23

only, I do not find any other linkages (e.g. rural nonfarm -small cities, and migration mega cities).
The results using the nearest distances to cities of the three size categories are shown
in model M2. The variance inflation factors of the three distance variables are 4, 2 and 7,
so we do not worry about multicollinearity if they are together present in the same
regression. The results provide some supplemental points to results in M1 and M3.
Specifically, households closer to the nearest larger cities have higher local-migration
income levels and shares. This is line with the findings above.
Moreover, Table 1.9 shows that once we control for distance to the nearest cities of
the three different sizes, a household being closer to a smaller city will not only increase
the share of local-migration income but also the share of local nonfarm income. This
implies linkages between small cities and local nonfarm employment. As noted above,
even if sample villages are close to large cities, being also close to small cities increases
income from local nonfarm employment. We surmise that local nonfarm employment
requires lower threshold investments and transaction costs, as well as lower skill
requirements compared with participation in migratory nonfarm employment. In
addition, small cities tend to rely more closely on their rural hinterlands than do big
cities, both for food and concomitant nonfarm activities such as transport and wholesale
and processing (BerdeguĂŠ and Proctor, 2014). The small city might itself be home to
agroindustries and agricultural services closely linked to the surrounding rural areas and
employing rural workers.
The results using population- and sectoral GDP-weighted distances, including all
cities, rather than the nearest ones only, are reported in the right three columns (M4, M5

24

and M6) in both tables. Unlike the results in M1-M3, where only partial linkages (small
city vs. local nonfarm income or larger city vs. local-migration income) could be
observed, several significant results with more detailed information emerge regarding
M4-M6.
First, the results using weighted distances show that the impacts of distance to cities
of different sizes are significantly different. Table 1.8 shows that rural households close
to smaller cities have higher agricultural income. In addition, being close to smaller cities
increases income from local nonfarm employment, while being close to larger cities
increases local-migration nonfarm income. Being closer to mega cities increases
migration income. We surmise that the observed small city - agricultural employment
linkage is similar to that discussed above for small city-local nonfarm rural employment
linkages: the small city depends on its rural hinterland and that market spurs agricultural
production with attendant effects on farm employment and nonfarm-farm linkage
employment. Moreover, it is probable that the link observed between proximity to mega
cities and migration income is that mega city proximity reduces the transaction cost to
migrate, and the mega cityâs employment opportunities and relative amenities are easily
observable by nearby rural households.
Second, effects of distances are differed by type of weight. Results using populationweighted (M4) and manufacturing GDP-weighted distance (M5) are identical for all
types of incomes, while the effects of distances on local and migration nonfarm incomes
are not significant using service GDP-weighted distance (M6). These suggest the pull
effects from larger cities on local-migration employment might operate through the effect
of the service sector, rather than manufacture or construction sectors of large cities.

25

By contrast, the pull effects from smaller (mega) cities on local (migration) incomes
are especially from the manufacturing sector of small (mega) cities. Compared with
findings for international migration, my mega city result is opposite to the argument that
the growth of the global cities would increase low-paid service jobs (Sassen, 1991),
supplied for example by the arrival of immigrants from poor countries into London
(Gordon and Kaplanis, 2014). But my findings are consistent with studies that show
mega cities serving as the base for clusters of advanced services firms, which demand
skilled workers (Taylor et al., 2014). Moreover, in the domestic migration literature my
findings concerning the effects of mega cities are consistent with those of Heath and
Mushfiq Mobarak (2015), who find that female migrants originating from close to Dhaka
are more likely to be involved in manufacturing occupations in Dhaka than those coming
from far from the city.
To further resolve the question of why, controlling for distance, the service sector
appears to be the major inducement for rural households to migrate to larger cities, but
the manufacture sector is the main factor (inferred from the effect of its weight) for small
as well as mega cities, I would need to have detailed information about those sectors in
the receiving cities and to analyze the capacity determinants (such as skills) of the rural
households migrating to different city sizes and choosing manufactures versus service
jobs in those cities. I lack both these kinds of data to pursue that analysis. I can only note
a hypothesis for future research that it may be that low skill-requiring agroindustrial jobs
are being sought by rural households in small cities, and higher paying jobs in urban
factories in mega cities. The latter jobs probably require more skill to enter. By contrast,
larger cities may have more developed service sectors that at the same time are low skill-

26

requiring; this may be for example where large cities are ports or transport hubs and the
migrants go there to seek jobs as drivers and loaders.
As the models using weighted distances (M4, M5 and M6) have better goodness-offit, and the former models may suffer from ignoring cities of other sizes or potential
effects problems, to explore how the distance and agricultural performance will jointly
impact the incomes I use below the latter three methods (M4, M5 and M6)

5.2. Interacted effects from urban proximities and agricultural performance
The results on both income levels and shares are reported in Table 1.10. Only the
results for migration incomes are discussed here, as they alone among the results are
significant for interacted effects.
Table 1.10 shows that using composite weighted distances, the signs and significance
of distance estimators in M4 and M5 alone are similar to the results without interactions.
The exception is that the effects are diminished with respect to local agricultural
performance. To wit, households have more migration income (levels and shares) if they
are closer to a mega city, but if they are living in the villages with better agricultural
performance, the effects of proximity to mega cities are smaller. This re-emphasizes
(along with the previous modelâs results) the importance of proximity for rural workersâ
migration to mega cities.
Moreover, while I expected from the traditional development literature to find that
there would be a strong effect of agricultural performance on local nonfarm employment
due to intersectoral linkage effects, instead I find that local agriculture performance has a
positive effect on migration income (both level and share) in M4 and M5. That is,

27

households in rural areas with better local agriculture performance have higher migration
income levels and shares, and these effects are magnified by being far from mega cities.
This is an interesting result for several reasons. On the one hand, much of the literature
(such as Taylor et al. 2003 for China) shows the effects of migration on farming (such as
investments from remittances into farm capital that can be especially important where
rural credit markets are constrained). On the other hand, some literature tests for the
effect of farming conditions (land tenure and land holdings, and farm productivity and
other measures of performance such as the net income from crops) on the decision to
migrate. The traditional hypothesis is that poor farming conditions drives migration as a
way to generate alternative income (Massey et al. 1993). Other work (such as de Janvry
and Sadoulet, 2001, for Mexico) finds that farm performance actually promotes migration
because better-off farmers have the initial wealth to make the investment in migration
(transport, set-up at destination, and so on); having oneâs own initial wealth source has
been found to be important in the situation where rural credit markets are constrained.
Also, as most migrants are from Sichuan, Hunan, and Jiangxi, where some remote areas
are very poor, better-off agricultural performance areas there might have an especially
important effect on enabling migration investment. To my knowledge, this finding is the
first time in the literature to test for a relation between distance from cities and the effect
of agricultural performance on migration.
Interestingly, distances and agricultural performance are jointly significant in M6
using distance weighted by GDP from the service sector. However, the results in M6 are
contrary to those in M4 and M5, which may be because migration employment is mainly

28

via the effects of the manufacturing sector in the mega cities, instead of the service
sector, as noted above.

5.3. Robustness check
In this section, I conduct two types of robustness checks. One is to rerun the above
estimations using different definitions of nonfarm incomes, and another is to consider the
potential concerns of missing variables in the above estimations.

5.3.1. Different definitions of nonfarm employment
On the one hand, it could be argued that the results might be sensitive to the
definitions of nonfarm employment by three sources (local vs. local-migration vs.
migration). I further explore the results in three other ways, used in some literature. First,
instead of defining local nonfarm employment as within the township, I include as
âlocalâ the nonfarm employment in another township but within the county.
Second, for local-migration, instead of defining all people doing nonfarm
employment outside their township and within their province as local-migration, I define
income earned by commuters as local-migration income. Similar to the local-migration
definition I use above, commuters are defined as rural people working outside of their
own township and within their own province, but different from the definition I use
above, I class commuters as being only those rural workers returning home every evening
(zaochu wangui, or âmorning leave evening returnâ).
Third, instead of defining all people working outside their own province as migrants,
I separate migrants into two subsamples based on the number of months per year the rural

29

worker stays at home. The first and second types of migrants are those working outside
their own province but staying outside for more than three and six months, respectively.
Table 1.11 compares the results for local nonfarm income with different definitions of
local-migration employment. The upper (in the table) results are based on the withintownship definition and the bottom results are based on the within-county definition.
Despite my varying the definitions, the signs and significance of results using M4 and
M5 discussed above remain unchanged.
Table 1.12 compares the results for local-migration income levels. The upper (in the
table) results repeat the previous results where I had defined local-migration as working
outside oneâs own township but within oneâs own province (local-migrants). The middle
(of the table) and bottom parts are the new results. The middle section shows results
including commuters only and the bottom results only include non-commuters (local
migrants excluding commuters).
The results show that the service sector GDP effects (M6) from larger cities are only
significant for commuters. Aggregation effects (the model using composite distance
weighted by population) (M4) and manufacturing sector weight effects (M5) from larger
cities are only significant for non-commuters. For non-commuters, their situations in
larger cities may be similar to rural migrants to mega cities, and thus they are more likely
to undertake jobs in the manufacturing sector, which may be relatively undemanding in
skills. However, for commuters, they may be likely to undertake jobs that involve
transiting between rural and urban areas as part of the service itself, such as in jobs in
transport and commerce. Yet rural residents might commute rather than take jobs in
cities that require long hours and inflexible schedules, such as in factories or construction

30

sites, because they are bound in some way to the rural areas. The latter might be due to
having to care for their farm land, or care for children and old parents.
Table 1.13 compares the results of population (M4) and manufacturing GDPweighted (M5) distances on migration income (levels and shares). The M4 results are in
the upper part of the table and the M5 in the bottom. The first two columns show the
previous results including all migrants working outside their own province. The middle
two columns include migrants working outside their own province for more than three
months and the left two columns include migrants working outside their own province for
more than six months.
The significances, signs, and even values are robust for all types of migrants using
these different methods. The distance to mega cities has negatively significant effects on
migration income (levels and shares) for all types of migrants.

5.3.2. Estimations with additional controls
A potential concern is that even though I controlled for the city effects by size and
potential urban effects from all cities using a composite distance, I may underestimate the
effects of the nearby cities. To avoid missing effects from nearby cities, I add
specifications with additional dummy variables in the regressions: whether there is a
nearby small city within 100 km; whether there is a nearby large city within 100 km; and
whether there is a nearby mega city within 300 km.
Table 1.14 compares the results of population (M4), manufacturing GDP-weighted
(M5), and service GDP-weighted (M6) distances on income levels by four sources. Both
results with and without the three additional dummies are reported.

31

The results from the robustness check columns of agricultural, local nonfarm, and
local-migration nonfarm incomes show that controlling for the composite distances,
having a city nearby decreases non-migratory income levels. This implies that for nonmigratory employment, it is not the nearby city alone, but total urban effects around the
rural areas that matter. Specifically, most of the results on composite distances are
consistent with my previous findings. Being close to the smaller cities increases incomes
from agricultural and local nonfarm employment, due to population and manufactures
aggregation effects. Being close to large cities increases income from local-migration
nonfarm employment, due to population and services aggregation effects.
The results on migration nonfarm incomes are not robust in the two models. Models
with the additional three dummy variables suggest that once we have controlled for the
composite distances, having a small or a large city within 100 km increases rural
householdsâ migration incomes. However, once we add the interactions between local
agricultural potential and distances, the results from the bottom part of Table 1.14 remain
consistent with the previous results. This shows that controlling the nearest cities by
sizes, local agricultural performance still has positive multiplier effects on migration
incomes. Specifically, agricultural performance increases migration incomes, particularly
for rural people far away from mega cities. The reasons are discussed above.

6. Conclusions
This paper seeks to contribute to the literature by examining the labor supply choices
of rural households in China as a function of distance from mega versus secondary cities,
controlling for the agricultural performance of the zones and household characteristics. It
also seeks to contribute by moving beyond physical distances to distances associated with
32

economic influence on household decisions of a city, including population and sectoral
GDPs. Three results stand out.
First, access to cities, measured in various ways related to distances, affect rural
household labor supply decisions to agriculture and non-agriculture. However, the
impacts differed by income sources and city sizes. In summary, I find three significant
linkages. The first set of linkages is between smaller cities and agriculture as well as local
nonfarm employment, which may be due to production and expenditure linkages between
rural and smaller cities. The second linkage is between larger cities and local-migration
employment. The third linkage is between mega cities and migration employment, which
probably is due to expected wage differences.
Second, pull effects from cities of different sizes on nonfarm income levels may come
from distinct sectors. Specifically, pull effects from manufacture and construction sectors
are important to all types of nonfarm incomes. However, it seems that rural commuting
nonfarm labor is more likely to be attracted by the pull effects from the services sector in
larger cities. This might be due to the types of jobs undertaken by commuters, who
probably prefer jobs linking rural-urban areas, like traders, or prefer jobs with flexible
time schedules due to labor constraints to their time in the rural areas.
Third, I find strong multiplier effects of local agricultural performance on migration
employment. This emphasizes the importance of not only urban proximity but also local
agriculture, which will help the poor to overcome credit constraints to diversify income.
My results shed some light on how important it is to take potential cities that may
influence householdâs income diversification into consideration, rather than including the
nearest city only. The work points possible directions for future work. Due to the

33

limitation of the data, I do not separate the nonfarm employments further, as by industry
or by low-high returns. With more detailed information on jobs chosen by rural labor
would make it feasible to clarify the linkages between city effects by sizes and sectors
and income diversification.

34

APPENDIX

35

APPENDIX

Table 1.1. Distribution of cities and population statistics by size of cities
City
Total Population
City
Population
Type
(million)
Number
Share
Population
Share
Governmentâs Classification
Small
18.24
0-0.5
64
18.6
3.3
Medium
66.82
0.5-1
88
25.6
12.1
Large
380.56
1-5
183
53.2
68.6
XL
40.75
5-10
6
1.7
7.3
XXL
48.2
>10
3
0.9
8.7
Total
344
100
554.57
100.00
Our Classification
Smaller
85.06
0-1
152
44.2
15.3
Larger
380.56
1-5
183
53.2
68.6
Mega
88.95
>5
9
2.6
16.1
Total
344
100
554.57
100.00
Note: population is number of people registered as urban citizen (chengshi hukou)

36

Table 1.2. Distance measures and village level descriptive statistics
Method
(M)
Mean

Std.
Dev.

Min

Max

I. Physical Distance (including nearest cities only)
M1
to nearest city of any size

54

34

10

142

M2
M2
M2

to nearest Smaller city
to nearest Larger city
to nearest Mega city

119
81
493

78
49
424

10
12
42

324
313
1,896

M3
M3
M3

to nearest city of any size (same as M1)
incremental distance to Larger city
incremental distance to Mega city

54
20
418

34
37
414

10
0
0

142
219
1,824

II. Composite Distance (including all cities, with different weights)
M4
to all Smaller cities (Population weighted)
1,587
486
M4
to all Larger cities (Population weighted)
1,320
411
M4
to all Mega cities (Population weighted)
1,243
471

1,222
881
864

3,067
2,534
2,653

M5
M5
M5

to all Smaller cities (Manufacturing GDP wgt)
to all Larger cities (Manufacturing GDP wgt)
to all Mega cities (Manufacturing GDP wgt)

1,504
1,275
1,246

448
426
473

1,172
788
845

2,905
2,418
2,629

M6
M6
M6

to all Smaller cities (Service GDP weighted)
to all Larger cities (Service GDP weighted)
to all Mega cities (Service GDP weighted)

1,465
1,309
1,265

533
418
430

1,071
851
824

3,012
2,437
2,518

(N=59)

Measures of Distance

37

Table 1.3. Distribution of sample villages by measure of distance
City observations

(Village #=59)

All
Cities
N=344
Village
%

Ave.
Pop.

Smaller
Cities
N=152
Village Cum.
%
%

Larger
Cities
N=183
Village
%

Cum.
%

Mega
Cities
N=9
Village
%

Cum.
%

I. Physical Distance (Including Nearest Cities Only)
M1. Distance to the nearest city
M2. Distance to the nearest city, by size
34
1.4
14
14
20
20
0
0-40
44
1.9
25
39
38
58
3
40-80
15
1.1
17
56
18
78
5
80-120
7
1.9
10
66
15
92
5
120-160
0
/
34
100
9
100
86
>160

0
3
8
14
100

II. Composite Distance (Including All Cities)
M4. Distance weighted by population
/
/
0
<1000
/
/
69
1000-1500
/
/
8
1500-2000
/
/
7
2000-2500
/
/
14
2500-3000
/
/
2
>3000

0
69
78
85
98
100

27
42
15
14
2
0

27
69
85
98
100
100

47
29
3
19
2
0

47
76
80
98
100
100

M5. Distance weighted by GDP from manufacture sector
<1000
/
/
0
0
39
1000-1500
/
/
75
75
22
1500-2000
/
/
3
78
27
2000-2500
/
/
15
93
12
2500-3000
/
/
7
100
0
>3000
/
/
0
100
0

39
61
88
100
100
100

49
27
3
19
2
0

49
76
80
98
100
100

M6. Distance weighted by GDP from service sector
<1000
/
/
0
0
36
36
41
41
1000-1500
/
/
76
76
24
59
34
75
1500-2000
/
/
2
78
29
88
12
86
2000-2500
/
/
8
86
12
100
12
98
2500-3000
/
/
12
98
0
100
2
100
>3000
/
/
2
100
0
100
0
100
Notes: Results from incremental distance measurement (M3) show that 21 (36%) sample villages
have a nearest city belonging to a smaller city, 37 (63%) have a nearest city belong to a larger
city, and one village can reach a mega city as its nearest city.

38

Table 1.4. Distribution of incomes of sample households, by years and sources
Year=2005 (N=2,560)
HH total income (RMB)
...Farm Income %
...Nonfarm Income %
...Transfer income %
...Other income %

Mean
23,121
10,046
10,953
782
1,340

Std. Dev.
20,025
14,888
10,831
1,828
9,882

Min
1,929
0
0
0
0

Max
317,250
241,500
225,000
36,050
303,000

Share of total
100.0
43.5
47.4
3.4
5.8

Year=2006 (N=2,560)
Mean
Std. Dev.
Min
Max
Share of total
HH total income (RMB)
25,348
20,612
1,675
351,512
100.0
...Farm Income %
9,854
14,733
0
244,000
38.9
...Nonfarm Income %
13,245
12,419
0
241,000
52.3
...Transfer income %
955
2,289
0
40,100
3.8
...Other income %
1,294
8,549
0
306,320
5.1
Notes:
Nonfarm income includes income earned from all types of nonfarm activities in any place.
Transfer income includes earnings from remittances and cash gifts.
Other income includes earnings from interest, insurance, and pensions.

39

Table 1.5. Descriptive Statistics of Family Members (Individual Level)
Group of people
(Col. A)

Days
at
Home

Days
doing
farm
prod.

Days
doing
NFA
invil.

Days
doing
NFA
outvil.

N.

Share of people (in col. A)
doing:

Share
(%)

Obs.

Age

Total individual

10,388

Any types of work (labor)

73.1

7,592

40

270

94

40

112

...Total labor

7,592

70.7

5,364

43

332

133

44

52

...Total labor

7,592

Any farm productions
Any nonfarm activities
(NFA)

74.1

5,623

37

240

64

54

152

......NFA labor

5,623

Any within village NFA

46.6

2,622

42

339

106

116

28

......NFA labor

5,623

62.7

3,528

.........Out-village NFA

3,528

Any out-village NFA
Any NFA in another
village

22.3

787

38

321

76

22

173

.........Out-village NFA

3,528

17.6

620

37

263

57

13

215

.........Out-village NFA

3,528

1.8

64

40

250

98

20

173

.........Out-village NFA

3,528

7.7

271

33

135

27

4

252

.........Out-village NFA

3,528

7.5

263

32

167

35

17

234

.........Out-village NFA

3,528

4.1

143

34

73

11

1

291

.........Out-village NFA

3,528

38.9

1,374

29

59

10

2

292

.........Out-village NFA

3,528

Any NFA in another town
Any NFA in another
county-rural
Any NFA in another
county-urban
Any NFA in provincial
capital city
Any NFA in another
province-rural
Any NFA in another
province-urban
Any NFA in another
country

0.2

6

28

79

2

2

293

Total individual

10,324

Any types of work (labor)

73.4

7,577

40

268

91

41

116

...Total labor

7,577

67.9

5,141

44

329

128

44

56

...Total labor

7,577

Any farm productions
Any nonfarm activities
(NFA)

76.4

5,786

38

238

61

55

157

......NFA labor

5,786

Any within village NFA

46.7

2,702

43

335

103

118

35

......NFA labor

5,786

63.2

3,657

.........Out-village NFA

3,657

Any out-village NFA
Any NFA in another
village

21.9

800

39

324

66

19

188

.........Out-village NFA

3,657

16.6

608

38

268

52

10

224

.........Out-village NFA

3,657

1.8

65

42

252

99

23

176

.........Out-village NFA

3,657

8.4

306

34

139

37

7

247

.........Out-village NFA

3,657

6.9

253

32

148

31

19

244

.........Out-village NFA

3,657

5.0

182

34

63

8

3

292

.........Out-village NFA

3,657

39.2

1,434

30

66

9

9

290

.........Out-village NFA

3,657

Any NFA in another town
Any NFA in another
county-rural
Any NFA in another
county-urban
Any NFA in provincial
capital city
Any NFA in another
province-rural
Any NFA in another
province-urban
Any NFA in another
country

0.2

9

28

74

4

3

284

Year=2005

Year=2006

40

Table 1.6. Distribution of nonfarm income of sample households, by year and
location
Year
(N=2,560 for each year)
HH Nonfarm income

2005
Mean
Share
10,953
100

2006
Mean
Share
13,245
100

1. Local
... within village
... outside village within township
âŚsub total

2,622
1,297
3,919

23.9
11.8
36

3,111
1,655
4,766

23.5
12.5
36

2. Local-Migration
... outside township within county
... outside county (rural) within province
... outside county (urban) within province
... provincial capital city
... sub total

1,445
221
626
604
2,896

13.2
2.0
5.7
5.5
26

1,815
159
762
686
3,422

13.7
1.2
5.8
5.2
26

3. Migration
... outside province (rural)
... outside province (urban)
... outside country
... sub total

451
3,663
24
4,138

4.1
33.4
0.2
38

482
4,514
61
5,056

3.6
34.1
0.5
38

41

Table 1.7. Descriptive Statistics of other variables
N=2,560 for each year
Variable
Individual/HH level
Age of HH head
Share of household head male
Highest education level of HH member
HH size
Children and elderly share within household
Cultivated land areas owned by household
Village level
Distance to nearest railway station
Distance to nearest long-distance bus station
Distance to main road
Share of cultivated land rented in/out
Share of labor working out of village
Share of HH with electricity
Number of phones per person
Agricultural performance (AP)
Note: 1 mu=0.0667 hectare.

2005

2006

Unit

Mean

Std.
Dev.

years
%
years
people
%
mu

48.8
96.9
8.9
4.1
18.1
6.2

10.5
17.4
2.3
1.4
20.2
9.9

49.7
96.8
9.0
4.0
17.8
6.2

10.4
17.5
2.4
1.4
20.2
9.7

km
km
km
%
%
%
number
RMB/m
u

38.0
9.4
1.4
7.7
33.2
99.4
0.27

53.2
6.7
1.9
11.8
19.8
3.2
0.11

38.0
9.4
1.5
8.1
33.6
99.3
0.31

53.2
6.7
2.4
11.8
20.3
3.4
0.12

710

511

735

541

42

Mean

Std.
Dev.

Table 1.8. Impacts of urban proximities on levels of income, by sources and by six
methods of measurements of distance
Composite Distance (weighted (Wgt.))
Physical Distance (M1-M3)
M2
Nearest
city by
size

M3

M4

Increment
Distance

Wgt. by
Population

5,120
2,560

5,120
2,560

Agricultural income (ln)
Dis_nearest city -0.743
Dis_smaller city /
Dis_larger city
/
Dis_mega city
/

/
0.247
-0.073
-0.582

Local nonfarm income (ln)
Dis_nearest city 0.466
Dis_smaller city /
Dis_larger city
/
Dis_mega city
/

/
0.128
2.138**
-0.809

Distance
measurements

Observations
Number of HH

M1
Nearest
city
5,120
2,560

Local-Migration nonfarm income (ln)
Dis_nearest city -2.882*** /
Dis_smaller city /
-1.322
Dis_larger city
/
-2.209**
Dis_mega city
/
0.656

M5
Wgt. by
GDP from
Manuf.

M6
Wgt. by
GDP from
Service

5,120
2,560

5,120
2,560

5,120
2,560

-0.759
/
-0.056
-0.197

/
-36.747***
-39.293
54.303

/
-62.638***
-25.840
62.498**

/
-21.786**
-1.553
8.720

0.415
/
0.008
-0.487

/
-64.009***
-6.868
53.748

/
-107.118***
-43.350
125.844***

/
-17.892
35.697
-13.728

-2.840***
/
-0.442
0.064

/
-8.112
-128.254**
129.353**

/
-43.783
-73.111**
104.217**

/
-9.369
-89.180***
91.384**

Migration nonfarm income
(ln)
Dis_nearest city 1.753*
/
1.789*
/
/
Dis_smaller city /
1.970*
/
26.567*
75.668***
Dis_larger city
/
0.494
0.241
95.073*
75.743**
Dis_mega city
/
0.115
0.530
-115.785** -136.679***
Note: In column M3, "Dis_larger city" and "Dis_mega city" are incremental distances
*** p<0.01, ** p<0.05, * p<0.1

43

/
-10.792
-4.325
2.094

Table 1.9. Impacts of urban proximities on income shares, by sources and by six
methods of measurements of distance
Composite Distance (weighted
(Wgt.))
M5
Wgt. by
GDP from
Manuf.

M6
Wgt. by
GDP from
Service

5,120
2,560

5,120
2,560

/
-3.609***
-1.992**
4.535***

/
-0.655
0.407
0.082

Local nonfarm income as share of total households income (%)
Dis_nearest city -0.017
/
-0.018
/
Dis_smaller city /
-0.045** /
-0.478
Dis_larger city
/
0.007
-0.002
0.742
Dis_mega city
/
-0.018
-0.012
-0.335

/
-0.377
0.247
0.121

/
0.154
1.259
-1.172

Local-Migration nonfarm income as share of total
income (%)
Dis_nearest city -0.075*** /
-0.073***
Dis_smaller city /
-0.065** /
Dis_larger city
/
-0.062** -0.002
Dis_mega city
/
0.054
0.017

/
0.376
-0.972
0.608

/
0.206
-1.870**
1.545

Distance
measurements

Observations
Number of HH

Physical Distance (M1-M3)
M1
M2
M3
Nearest
Nearest
Increment
city by
city
Distance
size
5,120
2,560

5,120
2,560

5,120
2,560

M4
Wgt. by
Population
5,120
2,560

Agricultural income as share of total households income (%)
Dis_nearest city
Dis_smaller city
Dis_larger city
Dis_mega city

-0.009
/
/
/

/
0.028
0.013
-0.017

-0.010
/
0.000
-0.010

/
-2.069***
-3.551**
4.731***

households
/
0.559
-2.013
1.490

Migration nonfarm income as share of total households
income (%)
Dis_nearest city 0.078***
/
0.079***
/
/
Dis_smaller city /
0.072**
/
1.755***
3.101***
Dis_larger city
/
0.030
0.009
2.425
1.898*
Dis_mega city
/
-0.004
0.011
-3.642**
-4.293**
Note: In column M3, "Dis_larger city" and "Dis_mega city" are incremental distances
*** p<0.01, ** p<0.05, * p<0.1

44

/
0.408
-0.192
-0.340

Table 1.10. Impacts of local agricultural performance (AP) and urban proximities on
migration income levels and shares, by three methods of measurements of distance
Composite Distance
Distance
measurements

M4

(Observations=5,120)
(Number of HH=2,560)

Wgt. by
Population

Migration nonfarm income (ln)
Dis1_smaller city (ln)
Dis2_larger city (ln)
Dis3_mega city (ln)
AP*Dis1
AP*Dis2
AP*Dis3
AP ( Ag. Performance, ln(RMB/Mu))

233.594**
419.469***
-549.558***
-22.403*
-60.252**
65.048**
135.471**

(weighted (Wgt.))
M5
M6
Wgt. by GDP
Wgt. by GDP
from
from Service
Manufacturing

373.592*
281.883
-558.585*
-43.743
-41.305
67.247
136.997

-66.479
-252.082
300.760*
3.184
-1.937
-3.984
19.713

Migration nonfarm income as share of total households income (%)
Non-Linear
Dis1_smaller city (ln)
9.451**
11.149
-1.080
Dis2_larger city (ln)
14.167**
6.572
-8.501
Dis3_mega city (ln)
-19.271**
-14.646
9.501*
AP*Dis1
-0.947**
-1.202
0.111
AP*Dis2
-2.073**
-0.963
0.466
AP*Dis3
2.353**
1.665
-0.640
AP ( Ag. Performance, ln(RMB/Mu))
5.125**
3.829
0.420
Note: AP and Distance variables are jointly significant at 1% and 5% in M4 and M5,
respectively; in migration income level (upper) regressions. AP and Distance variables are jointly
significant at 1% in M4, M5 and M6 in migration income share (bottom) regressions.
*** p<0.01, ** p<0.05, * p<0.1

45

Table 1.11. Robustness check using different definitions of local nonfarm income
levels
Distance
measurements

Composite Distance (weighted (Wgt.))
M4
M5
M6

(Observations=5,120)
(Number of HH=2,560)

Wgt. by
Population

Wgt. by GDP
from Manuf.

Wgt. by GDP
from Service

Local: within own township
Dis_nearest city (ln)
Dis_smaller city (ln)
Dis_larger city (ln)
Dis_mega city (ln)

/
-64.009***
-6.868
53.748

/
-107.118***
-43.350
125.844***

/
-17.892
35.697
-13.728

Local: within own county (Including within
within county)
Dis_nearest city (ln)
/
Dis_smaller city (ln)
-56.013***
Dis_larger city (ln)
-20.619
Dis_mega city (ln)
59.897
*** p<0.01, ** p<0.05, * p<0.1

46

township & outside township but
/
-88.340***
-38.344
105.667**

/
-21.628
5.918
14.729

Table 1.12. Robustness check using different definitions of Local-Migration income
levels
Distance
measurements

Composite Distance (weighted (Wgt.))
M4
M5
M6

(Observations=5,120)
(Number of HH=2,560)

Wgt. By
Popul.

Wgt. By GDP
from Manuf.

Base: all people doing local-migration nonfarm activities (NFA)
Dis_nearest city (ln)
/
/
Dis_smaller city (ln)
-8.112
-43.783
Dis_larger city (ln)
-128.254**
-73.111**
Dis_mega city (ln)
129.353**
104.217**

Wgt. By GDP
from Service
/
-9.369
-89.180***
91.384**

Alternative 1: commuters only (doing local-migration NFA and staying at home
everyday)
Dis_nearest city (ln)
/
/
/
Dis_smaller city (ln)
9.833
-10.702
-10.849
Dis_larger city (ln)
-41.926
-35.198
-72.394**
Dis_mega city (ln)
29.580
36.963
69.641**
Alternative 2: Non-commuters only (doing local-migration NFA but do not stay at home
everyday)
Dis_nearest city (ln)
/
/
/
Dis_smaller city (ln)
-18.570
-35.827*
0.843
Dis_larger city (ln)
-93.372***
-41.484**
-20.028
Dis_mega city (ln)
105.520***
71.240**
23.775
*** p<0.01, ** p<0.05, * p<0.1

47

Table 1.13. Robustness check using different definitions of Migration income levels
and shares
All Migrants
(Obs.
=5,120)
(HH
=2,560)

M4

M5

Wgt. by
Popul.

Wgt. by
GDP from
Manuf.

Migration nonfarm income levels (ln)
Dis1
26.567*
75.668***
Dis2
95.073*
75.743**
Dis3
-115.785** -136.679***
AP
0.123
0.123

Migrants:
Out home > 3 Months
M4
M5

Migrants:
Out home > 6 Months
M4
M5

Wgt. by
Popul.

Wgt. by
GDP from
Manuf.

Wgt. By
Popul.

Wgt. by
GDP from
Manuf.

23.898
101.764**
-118.922**
-0.037

73.385***
78.127**
-135.729***
-0.037

29.030*
102.389*
-123.510**
0.009

79.745***
79.297**
-142.127***
0.009

Migration nonfarm income as share of total households income (%)
Dis1
1.755***
3.101***
1.693***
3.016***
1.742***
Dis2
2.425
1.898*
2.447
1.890*
2.347
Dis3
-3.642**
-4.293**
-3.590**
-4.190**
-3.536**
AP
-0.007
-0.007
-0.015
-0.015
-0.013
*** p<0.01, ** p<0.05, * p<0.1

48

3.033***
1.835*
-4.160**
-0.013

Table 1.14. Robustness check, including additional controls in the estimations
(Obs=5,120)
(HH=2,560)

M4.
Wgt. by Population

M5.
Wgt. by GDP from
Manuf.
Main
Robust
Model
Check

Main
Robust
Model
Check
Agricultural income levels (ln)
Dis1
-36.747***
-22.080**
-62.638***
Dis2
-39.293
-24.437
-25.840
Dis3
54.303
25.554
62.498**
Has a small city /
-0.235
/
Has a large city
/
-1.648**
/
Has a mega city /
-1.341**
/
Local nonfarm income levels (ln)
Dis1
-64.009***
-52.219*** -107.118***
Dis2
-6.868
5.624
-43.350
Dis3
53.748
23.950
125.844***
Has a small city /
1.601
/
Has a large city
/
-2.797**
/
Has a mega city /
-1.500
/
Local-Migration nonfarm income levels (ln)
Dis1
-8.112
41.322*
-43.783
Dis2
-128.254**
-91.530*
-73.111**
Dis3
129.353**
68.914
104.217**
Has a small city /
-5.289***
/
Has a large city
/
-3.329**
/
Has a mega city /
0.861
/
Migration nonfarm income levels (ln)
Dis1
26.567*
-12.208
75.668***
Dis2
95.073*
69.225
75.743**
Dis3
-115.785**
-69.075
-136.679***
Has a small city /
2.860**
/
Has a large city
/
4.084***
/
Has a mega city /
-1.582
/
Migration nonfarm income levels (ln), with interactions
Dis1
233.594**
193.912**
373.592*
Dis2
419.469***
345.823**
281.882
Dis3
-549.559*** -481.640** -558.583*
AP*Dis1
-22.403*
-20.414*
-43.743
AP*Dis2
-60.252**
-56.002**
-41.305
AP*Dis3
65.048**
59.976**
67.247
AP
135.471**
126.395**
136.997
Has a small city /
3.039***
/
Has a large city
/
4.629***
/
Has a mega city /
-1.782*
/
*** p<0.01, ** p<0.05, * p<0.1

49

M6.
Wgt. by GDP from
Service
Main
Robustnes
Model
s Check

-24.287
-0.403
5.664
-0.462
-1.692*
-1.280**

-21.786**
-1.553
8.720
/
/
/

-0.221
37.562*
-48.032*
-0.808
-2.072***
-1.845**

-67.784*
-11.838
57.799
1.354
-2.680*
-1.280

-17.892
35.697
-13.728
/
/
/

9.836
91.550***
-96.935**
0.383
-3.958***
-2.749**

42.353
-41.548
19.383
-5.641***
-3.592***
0.955

-9.369
-89.180***
91.384**
/
/
/

15.075
-99.652***
93.887**
-5.194***
-3.135**
1.881

10.797
51.756
-71.500
3.039**
3.947***
-1.849*

-10.792
-4.325
2.094
/
/
/

-38.262**
-9.260
23.280
3.494***
4.726***
-0.931

258.713
186.329
-405.906
-33.344
-33.144
52.127
110.253
3.397***
4.499***
-1.890*

-66.479
-252.081
300.760*
3.183
-1.937
-3.984
19.713
/
/
/

-83.901**
-349.961**
372.560***
3.183
-1.938
-3.983
19.714
3.510***
5.631***
-1.364

Figure 1.1. Trend of income structure of rural households in China, 1990-2011
Note1: SE Farm shows share of income from self-employment farm activities; Wage
Income shows share of income from wage-employment; SE Non-Farm shows the share
of income from nonfarm self-employment nonfarm; Other income shows the share of
income from other sources apart from the above three activities.
Note2: Data is from National Bureau of Statistics of China, available on-line:
http://www.stats.gov.cn

50

Table 1.A1. Full results of impacts of urban proximities on local income values, by sources
and by six methods of measurements of distance
Distance
measurements

Physical Distance (M1-M3)
M1
M2
M3
Nearest
Nearest
Increment
city by
city
Distance
size

Composite Distance (weighted (Wgt.))
M4
M5
M6
Wgt. By
Wgt. By
Wgt. By
GDP from
GDP from
Population
Manuf.
Service

Observations
Number of HH

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

Dis_nearest city (ln)

0.466
(0.864)
/
/
/
/
/
/

/
/
0.128
(0.876)
2.138**
(0.907)
-0.809
(1.651)

0.415
(0.855)
/
/
0.008
(0.285)
-0.487
(0.696)

/
/
-64.009***
(21.623)
-6.868
(45.455)
53.748
(40.324)

/
/
-107.118***
(25.995)
-43.350
(31.856)
125.844***
(42.880)

/
/
-17.892
(18.309)
35.697
(28.601)
-13.728
(28.721)

0.987**
(0.473)

0.987**
(0.473)

0.987**
(0.473)

0.987**
(0.473)

0.987**
(0.473)

0.987**
(0.473)

0.010
(0.034)
0.029
(0.019)
-0.316***
(0.118)
0.046
(0.204)
-0.003
(0.012)
-0.004
(0.044)

0.010
(0.034)
0.029
(0.019)
-0.316***
(0.118)
0.046
(0.204)
-0.003
(0.012)
-0.004
(0.044)

0.010
(0.034)
0.029
(0.019)
-0.316***
(0.118)
0.046
(0.204)
-0.003
(0.012)
-0.004
(0.044)

0.010
(0.034)
0.029
(0.019)
-0.316***
(0.118)
0.046
(0.204)
-0.003
(0.012)
-0.004
(0.044)

0.010
(0.034)
0.029
(0.019)
-0.316***
(0.118)
0.046
(0.204)
-0.003
(0.012)
-0.004
(0.044)

0.010
(0.034)
0.029
(0.019)
-0.316***
(0.118)
0.046
(0.204)
-0.003
(0.012)
-0.004
(0.044)

0.014**
(0.006)
-0.173**
(0.078)
-0.077
(0.078)
-0.121
(0.115)
0.030
(0.028)
0.147**

0.012**
(0.006)
-0.193***
(0.068)
-0.077
(0.078)
-0.121
(0.115)
0.030
(0.028)
0.147**

0.015**
(0.007)
-0.175**
(0.074)
-0.077
(0.078)
-0.121
(0.115)
0.030
(0.028)
0.147**

0.013**
(0.006)
-0.185***
(0.068)
-0.077
(0.078)
-0.121
(0.115)
0.030
(0.028)
0.147**

0.009
(0.006)
-0.192***
(0.066)
-0.077
(0.078)
-0.121
(0.115)
0.030
(0.028)
0.147**

0.017***
(0.006)
-0.173**
(0.072)
-0.077
(0.078)
-0.121
(0.115)
0.030
(0.028)
0.147**

Dis_smaller city (ln)
Dis_larger city (ln)
Dis_mega city (ln)

Ag. Performance

HH/Individual Var.
Age of HH head
% of HH head is male
Highest edu. level
HH size
% of kids and elder
Cultivated land areas
owned by household
Village level Var.
Distance to nearest
railway station
Distance to nearest longdistance bus station
Distance to main road
% of cultivated land are
rented in/out
% of labor working out
of village
% of HH with electricity

51

Table 1.A1. (contâd)
(0.071)
2.605
(3.236)

(0.071)
2.605
(3.237)

(0.071)
2.605
(3.237)

(0.071)
2.605
(3.237)

(0.071)
2.605
(3.237)

(0.071)
2.605
(3.237)

0.222
(0.210)
-1.143
(1.987)
0.920
(1.648)
-0.963
(2.136)
-0.692
(2.040)

0.222
(0.210)
-1.823
(2.795)
1.091
(2.754)
-0.482
(3.367)
-2.784
(4.228)

0.222
(0.210)
-2.071
(2.285)
0.261
(1.936)
-1.795
(2.491)
-2.276
(3.160)

0.222
(0.210)
-2.169
(6.497)
4.825
(5.186)
-6.773
(8.277)
-17.852
(12.434)

0.222
(0.210)
2.000
(5.492)
7.403
(5.116)
-0.371
(6.847)
-7.596
(8.337)

0.222
(0.210)
-1.685
(8.912)
5.318
(7.217)
-4.849
(9.891)
-10.297
(9.213)

Constant

-6.119
(12.391)

-14.482
(15.954)

-3.216
(12.672)

129.539
(88.507)

182.754**
(73.173)

-36.307
(96.662)

R2 (overall)
P
Chi2

0.195
0.000
873.5

0.215
0.000
1025

0.196
0.000
980.5

0.222
0.000
1364

0.223
0.000
1498

0.208
0.000
729.6

Number of phones per
person
Time and province Var.
Year=2006
Province=Jiangxi
Province=Shandong
Province=Hunan
Province=Sichuan

Note: In column M3, "Dis_larger city" and "Dis_mega city" are incremental distances.

52

Table 1.A2. Full results of impacts of urban proximities on local-migration income
values, by sources and by six methods of measurements of distance
Distance
measurements

Physical Distance (M1-M3)
M1
M2
M3
Nearest
Nearest
Increment
city by
city
Distance
size

Composite Distance (weighted (Wgt.))
M4
M5
M6
Wgt. By
Wgt. By
Wgt. By
GDP from GDP from
Population
Manuf.
Service

Observations
Number of HH

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

Dis_nearest city (ln)

-2.882***
(0.921)
/
/
/
/
/
/

/
/
-1.322
(1.259)
-2.209**
(0.963)
0.656
(1.300)

-2.840***
(0.985)
/
/
-0.442
(0.304)
0.064
(0.584)

/
/
-8.112
(22.482)
-128.254**
(55.414)
129.353**
(53.246)

/
/
-43.783
(32.882)
-73.111**
(35.591)
104.217**
(52.205)

/
/
-9.369
(16.709)
-89.180***
(34.516)
91.384**
(36.684)

-1.940***
(0.438)

-1.940***
(0.438)

-1.940***
(0.438)

-1.940***
(0.438)

-1.940***
(0.438)

-1.940***
(0.438)

-0.023
(0.039)
0.011
(0.015)
0.158
(0.144)
0.358**
(0.153)
-0.011
(0.009)
0.044
(0.057)

-0.023
(0.039)
0.011
(0.015)
0.158
(0.144)
0.358**
(0.154)
-0.011
(0.009)
0.044
(0.057)

-0.023
(0.039)
0.011
(0.015)
0.158
(0.144)
0.358**
(0.154)
-0.011
(0.009)
0.044
(0.057)

-0.023
(0.039)
0.011
(0.015)
0.158
(0.144)
0.358**
(0.154)
-0.011
(0.009)
0.044
(0.057)

-0.023
(0.039)
0.011
(0.015)
0.158
(0.144)
0.358**
(0.154)
-0.011
(0.009)
0.044
(0.057)

-0.023
(0.039)
0.011
(0.015)
0.158
(0.144)
0.358**
(0.154)
-0.011
(0.009)
0.044
(0.057)

-0.001
(0.007)
-0.007
(0.070)
0.003
(0.075)
-0.014
(0.106)
0.019
(0.040)
-0.565***

-0.006
(0.008)
-0.035
(0.080)
0.003
(0.075)
-0.014
(0.106)
0.019
(0.040)
-0.565***

0.001
(0.007)
-0.024
(0.076)
0.003
(0.075)
-0.014
(0.106)
0.019
(0.040)
-0.565***

-0.023***
(0.008)
-0.044
(0.088)
0.003
(0.075)
-0.014
(0.106)
0.019
(0.040)
-0.565***

-0.021***
(0.007)
-0.068
(0.091)
0.003
(0.075)
-0.014
(0.106)
0.019
(0.040)
-0.565***

-0.017**
(0.007)
-0.031
(0.085)
0.003
(0.075)
-0.014
(0.106)
0.019
(0.040)
-0.565***

Dis_smaller city (ln)
Dis_larger city (ln)
Dis_mega city (ln)

Ag. Performance

HH/Individual Var.
Age of HH head
% of HH head is male
Highest edu. level
HH size
% of kids and elder
Cultivated land areas
owned by household
Village level Var.
Distance to nearest
railway station
Distance to nearest longdistance bus station
Distance to main road
% of cultivated land are
rented in/out
% of labor working out
of village
% of HH with electricity

53

Table 1.A2. (contâd)
(0.076)
-8.837**
(3.561)

(0.076)
-8.837**
(3.562)

(0.076)
-8.837**
(3.562)

(0.076)
-8.837**
(3.562)

(0.076)
-8.837**
(3.562)

(0.076)
-8.837**
(3.562)

0.440**
(0.220)
-5.372**
(2.230)
-6.885***
(2.042)
-4.845**
(2.310)
-4.807**
(2.406)

0.440**
(0.220)
-3.935
(2.719)
-6.414**
(2.570)
-3.109
(2.825)
-2.358
(3.595)

0.440**
(0.220)
-4.434
(2.719)
-6.541***
(2.216)
-4.788*
(2.591)
-3.841
(3.492)

0.440**
(0.220)
11.674
(8.552)
1.218
(6.515)
14.841
(10.259)
20.425
(14.088)

0.440**
(0.220)
3.882
(6.680)
-3.630
(6.407)
7.718
(8.197)
9.695
(8.795)

0.440**
(0.220)
-3.808
(6.847)
-1.637
(6.489)
-0.222
(7.963)
5.943
(8.056)

Constant

9.974
(12.279)

12.822
(18.706)

10.057
(13.723)

40.931
(90.868)

89.278
(103.187)

54.671
(82.277)

R2 (overall)
P
Chi2

0.115
0.000
1546

0.102
0.000
1553

0.120
0.000
1840

0.102
0.000
4090

0.0984
0.000
4304

0.0999
0.000
2325

Number of phones per
person
Time and province Var.
Year=2006
Province=Jiangxi
Province=Shandong
Province=Hunan
Province=Sichuan

Note: In column M3, "Dis_larger city" and "Dis_mega city" are incremental distances.

54

Table 1.A3. Full results of impacts of urban proximities on migration income values, by
sources and by six methods of measurements of distance
Distance
measurements

Physical Distance (M1-M3)
M1
M2
M3
Nearest
Nearest
Increment
city by
city
Distance
size

Composite Distance (weighted (Wgt.))
M4
M5
M6
Wgt. By
Wgt. By
Wgt. By
GDP from
GDP from
Population
Manuf.
Service

Observations
Number of HH

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

5,120
2,560

Dis_nearest city (ln)

1.753*
(0.912)
/
/
/
/
/
/

/
/
1.970*
(1.084)
0.494
(0.711)
0.115
(1.511)

1.789*
(0.963)
1.789*
-0.963
0.241
(0.320)
0.530
(0.608)

/
/
26.567*
(15.379)
95.073*
(51.805)
-115.785**
(53.228)

/
/
75.668***
(26.703)
75.743**
(30.302)
-136.679***
(49.771)

/
/
-10.792
(13.108)
-4.325
(25.174)
2.094
(29.831)

0.123
(0.351)

0.123
(0.352)

0.123
(0.352)

0.123
(0.352)

0.123
(0.352)

0.123
(0.352)

0.059**
(0.026)
-0.016
(0.014)
0.312**
(0.147)
0.420**
(0.201)
-0.011
(0.009)
-0.027
(0.027)

0.059**
(0.026)
-0.016
(0.014)
0.312**
(0.147)
0.420**
(0.201)
-0.011
(0.009)
-0.027
(0.027)

0.059**
(0.026)
-0.016
(0.014)
0.312**
(0.147)
0.420**
(0.201)
-0.011
(0.009)
-0.027
(0.027)

0.059**
(0.026)
-0.016
(0.014)
0.312**
(0.147)
0.420**
(0.201)
-0.011
(0.009)
-0.027
(0.027)

0.059**
(0.026)
-0.016
(0.014)
0.312**
(0.147)
0.420**
(0.201)
-0.011
(0.009)
-0.027
(0.027)

0.059**
(0.026)
-0.016
(0.014)
0.312**
(0.147)
0.420**
(0.201)
-0.011
(0.009)
-0.027
(0.027)

-0.015
(0.012)
0.070
(0.072)
-0.062
(0.064)
0.182**
(0.084)
0.015
(0.018)
-0.071

-0.015
(0.013)
0.098
(0.072)
-0.062
(0.064)
0.182**
(0.084)
0.015
(0.018)
-0.071

-0.017
(0.012)
0.081
(0.073)
-0.062
(0.064)
0.182**
(0.084)
0.015
(0.018)
-0.071

0.001
(0.011)
0.096
(0.071)
-0.062
(0.064)
0.182**
(0.084)
0.015
(0.018)
-0.071

0.002
(0.011)
0.119*
(0.071)
-0.062
(0.064)
0.182**
(0.084)
0.015
(0.018)
-0.071

-0.006
(0.011)
0.106
(0.074)
-0.062
(0.064)
0.182**
(0.084)
0.015
(0.018)
-0.071

Dis_smaller city (ln)
Dis_larger city (ln)
Dis_mega city (ln)

Ag. Performance

HH/Individual Var.
Age of HH head
% of HH head is male
Highest edu. level
HH size
% of kids and elder
Cultivated land areas
owned by household
Village level Var.
Distance to nearest
railway station
Distance to nearest longdistance bus station
Distance to main road
% of cultivated land are
rented in/out
% of labor working out
of village
% of HH with electricity

55

Table 1.A3. (contâd)
(0.046)
-2.343
(2.014)

(0.046)
-2.343
(2.014)

(0.046)
-2.343
(2.014)

(0.046)
-2.343
(2.014)

(0.046)
-2.343
(2.014)

(0.046)
-2.343
(2.014)

0.320**
(0.147)
5.442***
(1.868)
3.040*
(1.811)
4.992**
(2.404)
4.876**
(2.458)

0.320**
(0.147)
6.488**
(2.650)
3.622
(2.508)
4.195
(2.995)
6.692
(4.207)

0.320**
(0.147)
5.991**
(2.372)
3.611*
(2.095)
5.927**
(2.679)
6.170*
(3.668)

0.320**
(0.147)
-10.140
(7.644)
-8.016
(6.186)
-10.902
(8.717)
-10.111
(10.580)

0.320**
(0.147)
-6.712
(6.106)
-5.800
(6.198)
-8.700
(6.793)
-6.095
(6.028)

0.320**
(0.147)
-6.225
(7.367)
-6.510
(7.473)
-7.114
(7.636)
-4.262
(6.079)

Constant

-32.423***
(12.324)

-36.622**
(15.334)

-35.843***
(13.725)

-60.808
(58.407)

-130.663**
(59.731)

82.245
(84.166)

R2 (overall)
P
Chi2

0.334
0.000
725.4

0.335
0.000
657.6

0.336
0.000
772.9

0.337
0.000
1218

0.339
0.000
1112

0.328
0.000
800.5

Number of phones per
person
Time and province Var.
Year=2006
Province=Jiangxi
Province=Shandong
Province=Hunan
Province=Sichuan

Note: In column M3, "Dis_larger city" and "Dis_mega city" are incremental distances.

56

REFERENCES

57

REFERENCES

Au, C.C. and Henderson, J.V., 2006. Are Chinese cities too small?. The Review of
Economic Studies, 73(3): 549-576.
Au C-C and Henderson JV (2006) How migration restrictions limit agglomeration and
productivity in China. Journal of Development Economics 80(2): 350â388.
Barrett CB, Reardon T and Webb P (2001) Nonfarm income diversification and
household livelihood strategies in rural Africa: concepts, dynamics, and policy
implications. Food Policy 26(4): 315â331.
Benjamin D, Brandt L and Giles J (2005) The evolution of income inequality in rural
China. Economic Development and Cultural Change 53(4): 769â824.
BerdeguĂŠ JA, Carriazo F, Jara B, et al. (2015) Cities, Territories, and Inclusive Growth:
Unraveling UrbanâRural Linkages in Chile, Colombia, and Mexico. World
Development.
BerdeguĂŠ, J.A., F. Proctor. 2014. Inclusive Rural-Urban Linkages. Doc no 123, Working
Group: Development with Territorial Cohesion. Santiago, Chile: Rimisp.
Chernina E, CastaĂąeda Dower P and Markevich A (2014) Property rights, land liquidity,
and internal migration. Journal of Development Economics, 110: 191â215.
Christiaensen L and Todo Y (2014) Poverty reduction during the ruralâurban
transformationâthe role of the missing middle. World Development 63: 43â58.
Christiaensen L, De Weerdt, J, Ingelaere, B and Kanbur, R (2017) Why Secondary
Towns Can Be Important for Poverty Reduction-A Migrant's Perspective. No.
12193. CEPR Discussion Papers.
De Janvry A and Sadoulet E (2001) Income strategies among rural households in
Mexico: The role of off-farm activities. World development 29(3): 467â480.
De Janvry A, Sadoulet E and Zhu N (2005) The role of non-farm incomes in reducing
rural poverty and inequality in China. Department of Agricultural & Resource
Economics, UCB.
Deichmann U, Shilpi F and Vakis R (2009) Urban proximity, agricultural potential and
rural non-farm employment: Evidence from Bangladesh. World Development
37(3): 645â660.
Deininger K and Jin S (2005) The potential of land rental markets in the process of
economic development: Evidence from China. Journal of Development
Economics 78(1): 241â270.
58

Deininger K and Jin S (2006) Dynamics of temporary migration in China: exploring the
changing role of networks. World Band Policy Research Working Paper.
Desmet K and Fafchamps M (2005) Changes in the spatial concentration of employment
across US counties: a sectoral analysis 1972â2000. Journal of economic
geography 5(3): 261â284.
Du X, Ifft J, Lu L, et al. (2015) Marketing Contracts and Crop Insurance. American
Journal of Agricultural Economics.
Escobal J (2001) The benefits of roads in rural Peru: a transaction costs approach. Grupo
de AnĂĄlisis para el Desarrollo (GRADE), Lima.
Fafchamps M and Shilpi F (2003) The spatial division of labour in Nepal. Journal of
Development Studies 39(6): 23â66.
Fafchamps M and Wahba J (2006) Child labor, urban proximity, and household
composition. Journal of Development Economics 79(2): 374â397.
Fafchamps M, Koelle M and Shilpi F (2016) Gold mining and proto-urbanization: recent
evidence from Ghana. Journal of Economic Geography.
Fan CC and Scott AJ (2003) Industrial Agglomeration and Development: A Survey of
Spatial Economic Issues in East Asia and a Statistical Analysis of Chinese
Regions. Economic Geography 79(3): 295â319.
Gibson J (2000) A Poverty Profile of Cambodia, 1999. A Report to the World Bank and
the Cambodian Ministry of Planning, Phnom Penh.
Gordon IR and Kaplanis I (2014) Accounting for Big-City Growth in Low-Paid
Occupations: Immigration and/or Service-Class Consumption. Economic
Geography 90(1): 67â90.
Haggblade S, Hazell PB and Reardon T (2007) Transforming the rural nonfarm
economy: Opportunities and threats in the developing world. Intl Food Policy Res
Inst.
Harris JR and Todaro MP (1970) Migration, unemployment and development: a twosector analysis. The American Economic Review: 126â142.
Heath R and Mushfiq Mobarak A (2015) Manufacturing growth and the lives of
Bangladeshi women. Journal of Development Economics 115: 1â15.
Henderson JV and Wang HG (2005) Aspects of the rural-urban transformation of
countries. Journal of Economic Geography 5(1): 23â42.
Hymer S and Resnick S (1969) A model of an agrarian economy with nonagricultural
activities. The American Economic Review: 493â506.

59

Jin S and Deininger K (2009) Land rental markets in the process of rural structural
transformation: Productivity and equity impacts from China. Journal of
Comparative Economics 37(4): 629â646.
Jonasson E and Helfand SM (2010) How important are locational characteristics for rural
non-agricultural employment? Lessons from Brazil. World Development 38(5):
727â741.
Lewis WA (1954) Unlimited Supplies of Labour. Manchester school.
Li C and Gibson J (2013) Rising regional inequality in China: Fact or artifact? World
Development 47: 16â29.
Lucas RE (2001) The effects of proximity and transportation on developing country
population migrations. Journal of Economic Geography 1(3): 323â339.
Massey, DS, Arango, J, Hugo, G, Kouaouci, A, Pellegrino, A, Taylor, JE. 1993. Theories
of International Migration: A Review and Appraisal. Population and Development
Review, 19(3), September: 431-466.
Mellor JW and Lele UJ (1973) Growth linkages of the new foodgrain technologies.
Indian Journal of Agricultural Economics 28(1): 35.
Mu R and Giles J (2014) Village political economy, land tenure insecurity, and the rural
to urban migration decision: evidence from China. World Bank Policy Research
Working Paper (7080).
Partridge MD, Rickman DS, Ali K, et al. (2008) Lost in space: population growth in the
American hinterlands and small cities. Journal of Economic Geography: lbn038.
Rozelle S, Guo L, Shen M, et al. (1999) Leaving Chinaâs farms: survey results of new
paths and remaining hurdles to rural migration. The China Quarterly 158: 367â
393.
Sassen S (1991) The Global City: New York, London, Tokyo, Princeton. NJ: Princeton.
Stark O and Lucas RE (1988) Migration, remittances, and the family. Economic
Development and Cultural Change: 465â481.
Taylor, JE, Rozelle S, de Brauw, A. 2003. Migration and Incomes in Source
Communities: A New Economics of Migration Perspective from China.
Economic Development and Cultural Change, 51(1), October: 75-101.
Taylor PJ, Derudder B, Faulconbridge J, et al. (2014) Advanced Producer Service Firms
as Strategic Networks, Global Cities as Strategic Places. Economic Geography
90(3): 267â291.

60

Todaro MP (1969) A model of labor migration and urban unemployment in less
developed countries. The American Economic Review: 138â148.
Volpe Martincus C, Carballo J and Cusolito A (2017) Roads, exports and employment:
Evidence from a developing country. Journal of Development Economics 125:
21â39.
von ThĂźnen JH (1826) Der isolierte staat in Beziehung auf Landwirtschaft und
NationalĂśkonomie. Jena: Gustav Fischer.
Wooldridge JM (2003) Cluster-sample methods in applied econometrics.
Yunez-Naude A and Taylor JE (2001) The determinants of nonfarm activities and
incomes of rural households in Mexico, with emphasis on education. World
Development 29(3): 561â572.
Zhao Y (1999a) Labor migration and earnings differences: the case of rural China.
Economic Development and Cultural Change 47(4): 767â782.
Zhao Y (1999b) Leaving the countryside: rural-to-urban migration decisions in China.
The American Economic Review 89(2): 281â286.

61

CHAPTER 2: NETWORK EFFECTS AND RURAL-URBAN MIGRATION
DESTINATION CHOICES: EVIDENCE FROM CHINA

1. Introduction
Networks or social connections are important in social capital theory to explain
whether (migration decision) and where to migrate (destination choice) (Massey et al.,
1994, Stark and Jakubek, 2013). Networks can help migrants to reduce the costs of
moving, credit, settlement and establishment, and job search, and thus will affect the
migrantâs utility of whether and where to migrate (Bauer et al., 2009).
In this paper, I study the impacts of network effects alone, and as a substitute or a
complement to other migrant attributes like education and city amenities like city size, on
rural-urban migrantâs destination choice.
The gaps in the literature discussed below motivate the goal of this paper.
In the migration literature, there are incentives and capacity factors to determine the
supply of labor by rural people to migrate and to where to go. The incentives can be push
effects like poor agricultural performance, and can be pull effects like better amenities in
the migrating cities. The incentives can be pull factors like good education of the migrant,
so he wants to move to a place that values better his education than the rural areas, and
can be pull factors like better network capital, so he can move with little search cost.
The literature calls the migrant labor supply choice âpositive selectionâ if an attribute
of the migrant (such as education or network capital) positively affects the personâs
capacity to migrate as well as where to migrate. An example could be education or
network or both help migrant to go to a city where he/she gets a higher return to labor
(Chiquiar and Hanson 2005 and Grogger and Hanson 2011). While an example of

62

ânegative selectionâ could be that lower-educated people are more likely to migrate to
cities in search of low-entry barrier (skill requirement) jobs (Borjas 1987, Durand et al.
1996, and Moraga 2011).
The âcapitalâ above is really a vector. For example it can be education as well as
network capital for destination. Logically they can be complements or substitutes. Their
effects on migration decisions have been tested in some recent literature (Pedersen et al.
2008, Bertoli 2010, McKenzie and Rapoport 2010, and Beine et al. 2011), but not in all
possible interesting situations to test it. Specifically, whether those are complements or
substitutes have been mainly tested only in the international literature, but not in the
domestic (internal) migration literature. One can argue that the latter is similar in a way to
international migration because going from the mountain village to a big coastal city has
aspects of costs and threshold investments just like going from a Mexican village to Los
Angeles would have. But it also might be a different situation, at least in degree, as the
poorest do not tend to internationally migrate while they do tend to migrate internally,
and there have been few tests of their network capital and education in that setting.
But even controlling for whether a study of education and network capital effects has
been done in domestic or international settings, it has seldom been done where specific
destination cities have been used as the focus of the âwhereâ (instead of the general issue
of whether to just migrate domestically vs internationally as in Yunez-Naude and Taylor
(2001)). The city-specific âwhereâ is interesting in analysis of domestic migration
because one would expect heterogeneity of cities (size, economies, amenities) to
significantly affect both the individual coefficients on education and on network capital,
but also on their interaction as substitutes or complements.

63

Therefore there are two major research questions this analysis addresses. First, how
do network effects alone impact rural-urban migrantsâ destination choices, once moving
costs are controlled for? Usually the literature proxies search and moving costs by
whether the migrant has a network to help him/her move and to receive at the destination.
But here I separate the network from an explicit representation of the search and moving
costs, as a contribution to the literature. I hypothesize that rural-urban migrants prefer a
city where they have a more developed network (from their sending area), once search
and moving costs are controlled for.
Second, in what ways do rural migrantsâ networks and education (individually and
jointly) affect their urban destination choice? The hypothesis on this is ambiguous.
Networks can help less-educated rural people to migrate if the network helps the wouldbe migrant find and land a job requiring low skills in the destination city. But a network
can also help well-educated rural people to find commensurate jobs in the destination
cities. For either pool of migrants, poorly or well educated, a network could reduce
asymmetric information risks in the job market.
These two research questions will be addressed with a detailed data set for rural
migrants in nine destination cities in Guangdong province.
The paper is organized as follows. Section 2 explains the models. Section 3 summarizes
the data and the statistics. Section 4 shows the estimations. Section 5 discusses the
results. Section 6 concludes.

64

2. The model
Following the migration literature, the utility of a rural migrant choosing city j from
all available cities he could choose from depend on the expected net incomes person i
could obtain in location j, networks of i in j, migration cost from origin to destination j,
and a vector of location-specific amenities đżđ .

đ!" = đ! đ¸đĽđđźđđđđđ!" , đđ!" , đđś!" , đżđ  Â  = đ!" + đ!" ,

(1)

where đ!" is an indirect utility and

đđđ!" =đ˝! đđđ¸đĽđđźđđ!" + đ˝!" đđ!" â đ˝!" đđś!" + đˇđż đđđżđ + đ!" ,

(2)

Two factors should be emphasized here. First, I separate the network effects (NW)
and physical migration cost (MC) apart, which two factors are usually lumped together in
the literature (Bayer et al., 2009, Fafchamps and Shilpi, 2013). But I assume that the two
variables could have differently signed impacts on destination choice. I will use travel
distance in the model to measure moving cost. Second, the net income of migrant i
obtained in city j is only observed in chosen city j, while we cannot observe the incomes
that this individual would have earned if he gone to the other cities. I will generate
counterfactual expected net incomes for the individual in all destinations, following the
semi-parametric method in Dahl (2002). The detailed estimation methods of all the
independent variables will be discussed in section 4.

65

In equation (1), đ!" captures a vector of unobservables. Suppose the error term đ!"
follows independent and identical distribution type I extreme value distribution, the
probability of migrant i choosing city j to is,

đ!" =

!! !"!"#$%&!" !!!" !"!" !!!" !"!" !đˇđż đđđżđ !!!"
!
!!! !! !"!"#$%&!" !!!" !"!" !!!" !"!" !đˇđż đđđżđ !!!"

,

(3)

To recover the estimators in (3), a LOGIT estimation will be used. To explore the
joint effects of network effects and other controls, later in the analysis the interaction
terms between individual characteristics, city level amenities and network effects will be
included to allow for the inclusion of individualâs characteristics.

3. Data
3.1. Data and sampling
The primary data set used in this study is collected in the China Labor-force
Dynamics Survey (CLDS) by Sun Yat-sen University in 2009. The survey interviewed
rural-urban workers with college degrees or lower, who migrated from another province
or from a rural area within Guangdong province into one of the nine cities4 in
Guangdong. The survey collected origin and destination places of the migrants, detailed
information on employment and living situation, including wage and rents, and
demographics of the migrants.

4

The nine cities include Guangzhou, Shenzhen, Zhuhai, Foshan, Zhaoqing,
Dongguan, Huizhou, Zhongshan, and Jiangmen cities.
66

The sampling method of the CLDS is as follows. First, they used the 2000 census to
obtain the number of rural migrants (including migrants from another province or within
Guangdong province but from another city. Second, they generated share of migrants in
city j over total number of migrants in the nine cities to obtain the distribution of
population migrants over sample cities. Third, they used the stratified sampling method to
randomly sample migrants in the nine cities. In total 1,544 migrants are used in my
analysis5.
CLDS selected the sample migrants in nine cities, then the sample is likely to be
nonrandom. Since the probability of selecting this migrant in city j is conditional on the
probability of this migrantâs choosing city j, which will result in biased estimators (Shaw,
1988). To address this dress this endogenous stratification concern, I will use weighted
exogenous stratification maximum likelihood estimation in the later analysis.

3.2. Characteristics of the sample
Sample migrantsâ birth provinces are spatially concentrated. 28% of them are from
rural Guangdong. The other migrants are mainly from four nearby provinces, including
provinces inland from the coast and in the north-south middle belt of China, composed of
relatively poor provinces (Hunan, Guangxi, Sichuan, and Hubei), which account for
another 47%. The remaining 25% are mainly from the North. The migrants tend not to be
from the richer coastal provinces.

5

22 observations from urban areas (with Chinese urban hukou registration) collected by
CLDS are excluded in the analysis, since I will only study the destination choice of ruralurban migrants.

67

The demographics of the sample are shown in Table 2.1. The average age of the
sample migrant is 29. 54% of them are male, and 48% are married. Only 42% have
attended high school or received higher-level degrees. This is not surprising, since the
survey only interviewed migrants with education levels lower than the collage.
Table 2.2 shows that the average monthly wage is 1730 RMB (around $251 in 2017
USD), which, according to Guangdong Statistical Yearbook, is lower than the average
wage earned by laborers in Guangdong from the manufacture sector (2126 RMB) and
service sector (3456 RMB), but higher than that from the agriculture sector (1087 RMB).
Only small amount of rents is paid by the migrants due to the poor living conditions or
that the accommodations are provided by the employers. Average net income (after
subtracting the small housing rent paid from the gross income) is 1647 RMB.
4. Estimations
To recover the coefficients in (3), I discuss the specifics of the independent variables
in the model, including network effects, counterfactual incomes, migration cost, and city
amenities.

4.1. Networks
The information to calculate network effects is obtained from the 2010 population
census of Guangdong province6. The census shows the number of migrants in each of the
cities (prefectural level) of Guangdong province by origin province and number of
migrants in each of the cities from another city.
6

Note population census is available every five years, thus we use the information of 2009, which is
collected in 2010, though the survey launched in 2009 and collected information of 2008. And I assume the
distribution of migrants from a certain city over total migrants in the destination city will remain similar in
2008 and 2009.

68

I calculate the network using two methods. First, the network of each individual from
region r (r refers to a province if the individual is not from Guangdong, and refers to a
prefectural level city within Guangdong province if the individual is from Guangdong
province) and living in city j is the share of people in city j from origin r over total
migrants in 21 cities of Guangdong from region r, which can be specified as,

!
đđ!"
=

!"#$%"!"
!" !"#$%"
!"
!!!

,

j=1,..,9, r=1,âŚ,23.

(4)

In total, there are 23 origin provinces and nine destination cities. The networks vary
by origin-destination but not by individual within the same origin. The total number of
cities (k) does not equal to the maximum number of destination city (j), because there are
21 cities in Guangdong, while only nine cities are covered in the survey. Although only 9
cities are included in the survey, 93% of the rural-urban out-province migrants have
chosen one of these 9 cities as their destinations.
The second type of network is calculated as follows. The network of each individual
from region r and living in city j is the share of people in city j from origin r over total
migrants in city j from all other regions o, which can be specified as,

!
đđ!"
=

!"#$%"!"
!" !"#$%"
!"
!!!

,

j=1,..,9, r=1,âŚ,23.

(5)

Same as the first type of network effect, the second one also varies by origindestination but not by individual within the same origin. Region o equals to 32, including

69

31 provinces other than Guangdong province in China, and Hong Kong, Macao and
Taiwan regions.
Table 2.3 reports the statistics for the network variables. Sample migrants are from 23
provinces and thus the networks have 207 different values (23 origins*9 destinations).
The results from the upper panel show that the average value is 10.42%, and it varies
from 0.37% to 46.75%. The higher the value is, the higher level of networks the sample
migrants are faced with. Specifically, the highest value in Shenzhen city (46.75%) is from
Jilin province, that is to say 46.75% of rural migrants from Jilin province in Guangdong
province is concentrated in Shenzhen city.
The results from the lower panel show the results of the second measure of networks.
The average value is 5.67%, and it varies a lot from 0.15% to 76.56%. Still, the higher
the value is, the higher level of networks the sample migrants are faced with. Specifically,
the lowest value in Zhongshan city (0.15%) is from Shanxi province, that is to say only
0.15% of out-of-province rural-urban migrants in Zhongshan city is from Shanxi
province.

4.2. Calculation of counterfactual incomes
The income of migrant i obtained in city j is only observed in city j, while of course
we cannot observe the incomes that this individual would have earned in other cities.
Unobserved heterogeneity of the individual (such as risk aversion) affects whether he has
chosen city j and has thus entered our sample for city j. To solve this self-selection
problem in a choice model, I generate counterfactual expected net incomes for the
individual in all destinations, following the semi-parametric method in Dahl (2002).

70

First, I use the following function form to regress the location-specific realized
income on individual attributes,

đđđ¸đĽđđźđđđđđ!" = đź + đź!"#,! đđđ´đđ + đź!"#$,! đđđđ + đź!"##$%&,! đđđđđđđ +
!
đź!"#_!!"! , đđ¸đđ˘_âđđâ + đź!! ! đ r, j Edu + đź!! ! [đ r, j Edu ]! + đ!"
,

(6)

The term đ r, j Edu measures the percentage of individuals with different education
levels, born in region r, and living in city j. The reason to include đ r, j Edu is to
improve the accuracy of estimators on independent variables and thus counterfactual
incomes. Including these terms, we assume that unobserved heterogeneity is relatively
small of migrants migrating from the same birth location to to same destination j with the
same level of education.
The equation to obtain đ r, j Edu is as follows,

đ(r, j|Edu) = đ¸đđ˘_đżđđ¤! â P(r, j|Edu_low) + đ¸đđ˘_âđđâ! â P(r, j|Edu_high),

(7)

To calculate this, we need to divide the observations into mutually exclusive cells
based on education level. For simplicity, I combine some provinces, which gives three
regions of origin: Guangdong province, Neighbor provinces (including four provinces
sharing borders with Guangdong province: Guangxi, Hunan, Jiangxi, and Fujian), and
Non-neighbor provinces.
In total 54 cells are generated. The average size of the cell contains 54 migrants, with
the size ranges from 1 to 104. The distribution of the sample migrants by types is shown

71

in Table 2.4. We can then generate the percentage of individuals with different certain
education level from region r to city j following (7).
The calculated results of đ(r, j|Edu) are shown in Table 2.5. Using these obtained
probabilities in regression (6) allows us to obtain the estimated coefficients for each
variable. With these coefficients, I can then impute the expected incomes for individual
in all cities, even if the migrant did not migrate to that city. The imputed incomes for the
sample can be found in Table 2.6.

4.3. Migration cost
To calculate the migration cost, I use Google Map to get the travel distance and time
from the migrantâs origin to the nine destination cities. Since we know from which
location to what city each migrant goes, I first extract geographic coordinates of the
urban centers of these cities. Then I use Google Map to calculate a pairwise travel time
from 174 origin cities to nine destination cities.
The summary of the results can be found in Table 2.7. It is not surprising that the
travel distance and time are similar among sending locations to each destination city,
since these nine cities are geographically close by. But the distances and travel time vary
a lot from the center of each of the nine study cities to the origins of the migrants as these
origins stretch across China. The maximum travel distance is 3555 km and the longest
travel time by car is 36 hours.

72

4.4. Amenities of cities
City amenities hypothesized to affect rural-urban migration are obtained from
Guangdong Bureau of Statistics, the 2009 Guangdong Statistical Yearbook. The
summary of the amenities is in Table 2.8.
Urban population shows the size of the city (the urban center of the prefecture city).
Per capita expenditure from the government is controlled for to capture the investment in
local amenities. I expect these two would positively affect the migration choice to city j
since these two variables are pull effects.
Numbers of hospitals, and primary and middle schools per person are included. This
is of course important for the children and elderly family members of migrants as well as
migrantsâ work injuries.
Other employment- related variables are included. The average unemployment rate is
2.3% with small variation among nine cities, which is expected to be negatively
correlated with the destination choice. The numbers of industrial and construction firms
are included in the regressions, which vary a lot among the nine cities. Since rural-urban
migrants are likely to be involved in the industrial and construction-related occupations, I
expect that cities with greater number of these firms attract more rural-urban migrants.

4.5. Weighting used to correct for on-site sampling
The last estimation issue to consider in the estimation is the on-site sampling problem.
If it is ignored it will bias the estimators. Following Wooldridge (2002), I use weighted
exogenous stratification maximum likelihood estimation to address this endogenous
stratification problem. The weights used in the analysis are shown in the last column of

73

Table 2.9, which is calculated as the ratio of population share over the sample share
(Weight (column (5))= column (2)/column (4)).
The distribution of city populations and sample migrants in the nine city destinations
are in Table 2.9. All the nine destination cities are in Guangdong, and the distributions
are quite concentrated in a few cities. Around 31% of the migrants are located in
Dongguan city, followed by Shenzhen (23%), Guangzhou (17%), and Foshan (9%) cities,
four of which have accommodated 80% of total sample migrants. The remained 20% of
sample migrants work in other five cities, with fewer than 100 sampled migrants in each
city.

5. Estimation results
5.1. Impacts of network effects
I estimate the probability that rural migrant i choses city j in Guangdong province.
The results are estimated using conditional Logit. To allow for the correlations within
individual, the error terms are clustered by individual level. The results are shown
starting from Table 2.10. The regression is in several variations in Table 2.10, 2.11, and
2.12: I begin by estimating the equation with city attributes, and then control for city
fixed effects. Both results with income and unemployment rate interactions are reported.
The summary statistics for city attributes included in the estimation are shown in Table
2.8. Results in Table 2.10 and Table 2.11 do not include migration costs (hours of travel)
and report the results using two different measures of network effects, respectively. Table
2.12 has all the same models as above tables, but also includes travel time from the
migrantâs origin to the destination city as an independent variable.

74

Recall that our first research question concerns the network effect on rural-urban
migrantsâ destination choice. The regression results from Table 2.10 and Table 2.11 show
that the network effects are positively significant, controlling for other expected income
and city amenities. The estimator is somewhat smaller after controlling for city level
fixed effects in Table 2.10, but the significance level and sign of the estimator are
remained the same. The positive effect of network effects on migrantsâ destination choice
is consistent with the existing literature discussed in the introduction, but this finding is
about the âmeso networkâ (cluster of people they do not know but from the same birth
province) rather than the focus of the earlier literature that was on the âmicro networkâ of
the family and local community of the migrant to a domestic destination. Our results are
similar to the results in Fafchamps and Shilpi (2013), who test impacts of social
proximity on choice of migration destination in Nepal. The results suggest that rural
migrants would prefer to work and live in a city with higher share of population coming
from the same origin (rural location in the province or outside province). The reason
behind this is that the cities with larger size of networks might reduce the moving costs,
and might provide the migrants better access to jobs, settlements, and other information.
Surprisingly, model (1) and (2) in Table 2.10 and Table 2.11 do not show significant
impacts of income no matter if I control for city fixed effects or not. I expected that the
income would be a key variable that positively impact migrantsâ choices, as found in
Bertoli (2010). However these insignificant results might be due to the fact that the
sampled migrants are rural-urban migrant, who are more likely or able to undertake
occupations with poorer conditions or with lower wage. This assumption can be partially
confirmed by the following. The estimator of the unemployment rate is positively

75

significant: rural-urban migrants prefer cities with higher unemployment rates. It implies
that rural migrants may have higher opportunity to be employed by the occupations with
higher risks to be fired and higher temporary turnover like construction jobs. I further test
the hypothesis by interacting income and unemployment. The results are shown in
columns (3) and (4) of Table 2.10 and Table 2.11. The estimators of income as well as
the interaction term are significant. Rural-urban migrants prefer cities where they can
earn higher incomes, while this effect is decreasing by the unemployment rate of the city.
This is reasonable, since in cities with low unemployment rates, seeking a job with a
higher wage is the goal of the rural-urban migrant. However, in cities where
unemployment is high, the first goal of the migrants is to find a job to guarantee some
earnings, and then the second step is to seek a job with higher incomes.

5.2. Impacts of the migration cost
The second research question addressed by this paper is how migration cost will
impact the choice of city destinations. I used travel time as the index for physical
migration cost, and the results are robust to travel distance. The results with two measures
of networks are shown in Table 2.12. As expected, a longer travel time thus higher
migration cost significantly decreases the probability of the destination choice in all the
models. Comparing the results in Table 2.12 with Table 2.10 and Table 2.11, although
ignoring the migration cost leaves the first type of network estimator almost the same, the
estimators of second type of network are not significant once controlling for the city
effects (models 2 and 4 in the lower panel). Two implications are obtained from this
result. First, it emphasizes the importance to include both physical migration cost and

76

social networks separately in the migration equation, which is generally lumped together
in the previous literature. Second, it shows that the different measurements of the mesolevel networks will differ the results. The calculation from the origin-orientated side (first
measure: share of migrants in destination city j from origin r over total migrants from r)
seems to have more significant impacts than destination-orientated side (second measure:
share of migrants in city j from origin r over total migrant in destination city j), when we
control for the physical migration cost.
The estimators on city amenities in Table 2.12, which are not reported due to spaces,
are also different from Table 2.10 and Table 2.11 when the migration cost is controlled
for. This finding is consistent with the finding in Bayer et al. (2009), who showed that
ignoring the mobility cost would make a large difference in the estimated value of clean
air (city amenity) of the destination.

5.3. Results with individual characteristics interactions
Table 2.13 and Table 2.14 show the results with interactions terms between network
effects and education level, and micro-level nonfarm experience, respectively. All the
models from here include city fixed effects and migration cost. For comparison, results
without interaction terms are shown in column (1) and (4), which are the same as the
results in column (4) in Table 2.12.
Recall that the third research question concerns whether the network effect could
condition the effect of education on city destination. As discussed in McKenzie and
Rapoport (2010), networks could shape self-selection through decreasing migration cost.
However, they use the migrantâs network as an index for migration cost to test this

77

hypothesis. In this paper, I can and do separate the network effect from the moving cost
(travel time), and compare the effects of the two on selection.
Columns (3) and (6) in Table 2.13 show that the estimator of the interaction term of
education and migration cost are positive, while the estimator of the migration cost is
negative, no matter which measurement is used to calculate networks. Taking both
together shows that although being close to a city would increase the probability of
choosing this city, this effect can be decreased by having a higher education level. This is
consistent with high migration costs resulting in positive self-selection, as in Chiquiar
and Hanson (2005). The reason for this positive selection is that education can increase
the information accessibility and decrease the other nonmonetary costs to settle and to
find a job.
Interestingly, column (2) and (5) also show that the estimators of education-network
interaction are significantly positive. In particular, since the estimator of the network
effect is significantly positive, it says that cities with greater level of network effects
would attract more highly educated rural-urban migrants. This result is opposite to that of
McKenzie and Rapoport (2010), who find that education and networks have substitution
effects, meaning that lower-educated migrants in communities with stronger networks for
Mexico-U.S. migration are more likely to migrate.
The difference with the Mexico result can be explained in two ways. First, it might be
due to their not separating networks and migration costs. McKenzie and Rapoport (2010)
suggest that migration cost is a function of networks, and use network as the index for
migration cost in their estimation model, without controlling for further physical
migration costs. However, our results show that once we controlled for the physical

78

migration costs (travel time), the network effect alone does result in a positive selection
effect for education.
Second, the difference may be due to the different characteristics of Chinese ruralurban domestic migration and Mexico-U.S. international migration. Since the MexicoU.S. international migrants undertake employment with lower skills, such as wage
employment in the agricultural sector or the service sector, they do not need a high level
of education; networks of Mexican immigrants could help them to find these types of
jobs requiring low skills. However, in the case of Chinese internal migration, recall that
the average income of sample migrants, shown in the statistical table, is much higher than
that of agricultural sector, and similar to the manufactures industry. It is thus possible that
the jobs undertaken by these rural-urban migrants require some skill. Thus even with the
help of the network, education plays a multiplier effect. That implies that the networks
may help the laborers to overcome the asymmetric information problem in the job market,
and thus education is still a requirement asked for by the employers in the cities.
Again, the results suggest that it may not be appropriate to use the network effects as
the index for migration cost directly. Highly educated rural-urban migrants could
decrease the credit constraints and choose cities being far from the origins. Meanwhile,
they would choose cities with higher network effects, once controlling for the travel time.
Both of these show positive selections in terms of education as education could decrease
both types of costs, including physical and psychological costs. The results are consistent
with that in Chiswick (1999).
Table 2.14 shows the results with interactions between networks and whether the first
nonfarm job is referred by a family or friend member in columns (1) and (4), whether the

79

current nonfarm job is referred by a family or friend member in columns (2) and (5), and
years of undertaking nonfarm employment in columns (3) and (6), and we do not find any
significant effects of these micro-level social networks, once we controlled for the mesolevel networks.

5.4. Results with city amenities interactions
I further explored how the results might differ if we allow the networks effects to be
varied by city amenities. The idea of including these terms is to see whether network
effects could generate some substitute or multiplier effects in terms of city attributes.
Specifically, columns (1) and (4), columns (2) and (5), and columns (3) and (6) in Table
2.15 show the results include network interactions with city size (urban population),
number of industrial firms, and number of construction firms, respectively. The Wald
tests show that all the terms with interactions are also jointly significant. Several results
are standing out.
First, the significantly opposite results on networks and network-city size interaction
suggest that network effects are more important for a rural-urban migrant to choose a city
when the city is small. For rural migrant facing with a same level of networks in two
cities, it is much easier for him to get contact with someone in the networks or to be
benefit from the externalities from the networks in a smaller city. While, when the
network effects are low, people prefer larger cities where the urban push effects should
be larger.
Second, the results in columns (2) and (3) using origin-orientated networks (measured
by the first method) show substitute effects between networks and number of industrial

80

and construction firms. Specifically, migrants prefer cities with higher level of networks,
yet these effects are diminishing if cities have greater number industrial and construction
firms. These results make sense as migrants are more likely to be hired in cities with
more firms if other factors are the same, and it is likely that they can find a job more
easily even with no or low level of networks in these cities.
The results in columns (5) and (6) using destination-orientated networks (measured
by the second method) show opposite results. Similar to the results using originorientated networks, migrants still prefer cities with more industrial and construction
firms, however, networks play complementary effects in these two models, meaning they
prefer cities with more firms, especial for those with more networks. These different
results imply the importance to differ the measurement of networks when we study the
impacts of network effects, as different measures may give very different results.

6. Conclusions
Network effect is important for rural-urban migrations. In this paper, I show
networks, calculated by two ways, will increase the probability of choosing this
destination for rural-urban migrants in Guangdong, China.
This paper help in part to show physical migration costs (travel time) have
significantly negative impact on the choice. Ignoring these physical costs would results in
biased estimators of city attributes. Also, the paper emphasizes the importance to separate
the physical (migration costs) and non-monetary costs (networks) to study the domestic
rural-urban migration in China.

81

I find the migration costs and education play the substitute effects, while education
and network effects play the complement effects on destination choice of rural-urban
domestic migration. My different results compared to international migration suggest that
for different types of employment, education and networks may have different effects.
These results address the importance of decreasing migration costs and increasing
network effects if the city would like to attract more rural-urban migrants. Meanwhile,
developing education skills of rural-urban migrants in China should also help the people
faced with credit constraints, and would multiply the network effects of the cities.
The paper also finds that the effects of origin-orientated networks measured as share
of people from origin r in city j over total migrants in Guangdong from origin r, are
different from that of destination-orientated networks measured as share of people from
origin r in city j over total migrants in city j from all other provinces. The originorientated networks show more significant effects in most of the cases and show
substitute effects for number of industrial and construction firms in the city, while the
destination-orientated networks show the opposite. This suggests the importance to differ
the measurements of network effects in the future studies.

82

APPENDIX

83

APPENDIX

Table 2.1. Summary statistics of demographic variables, Individual level
Variable
Description
Mean Std. Dev.
(N=1,544)
Age
Age (years)
28.55
8.87
Male
Gender (=1 if male, 0 if female)
0.54
0.50
Edu_high
Education level (=1 if equal to or higher than
high school, =0 if lower than high school)
0.42
0.49
Married
Marital status (=1 if married)
0.48
0.50

84

Table 2.2. Average incomes and rents per month
Variable (N=1,544)
Mean Std. Dev. Min
Wage received from the firm
1,730
946
240
Housing subsidies received from the firm
42
95
0
Total income (wage+ housing subsidies)
1,772
952
250
Rental cost
125
163
0
Income (total income-rental cost)
1,647
914
230
Note: RMB in 2009/Month

85

Max
15,000
800
15,000
1,200
15,000

Table 2.3. Origin-destination paired networks (%), by cities and measures
N.
Mean Std. Dev.
Min
Max
Network I
All
207 10.42
11.43
0.37 46.75
By Cities
Guangzhou
23
18.61
6.36
9.74 30.23
Shenzhen
23
33.88
8.04
19.31 46.75
Zhuhai
23
2.35
1.42
0.89
5.21
Foshan
23
7.99
3.65
4.29 19.64
Zhaoqing
23
0.87
0.54
0.37
2.25
Dongguan
23
19.09
7.38
8.01 32.39
Huizhou
23
4.97
1.42
3.04
8.30
Zhongshan
23
4.17
1.88
2.07
9.81
Jiangmen
23
1.90
1.17
0.75
4.73
Network II
All
By Cities
Guangzhou
Shenzhen
Zhuhai
Foshan
Zhaoqing
Dongguan
Huizhou
Zhongshan
Jiangmen

207

5.67

9.25

0.15

76.56

23
23
23
23
23
23
23
23
23

5.77
7.62
6.12
5.57
4.64
6.07
5.23
5.24
4.73

8.68
15.92
9.70
8.94
6.42
9.52
6.52
7.79
7.49

0.48
0.43
0.80
0.20
0.25
0.16
0.26
0.15
0.16

34.45
76.56
44.03
29.67
25.47
39.92
21.22
27.10
33.13

86

Three
Regions (r)
Observations
City (j)
Dongguan
Shenzhen
Guangzhou
Foshan
Zhongshan
Zhuhai
Jiangmen
Huizhou
Zhaoqing
Total

Table 2.4. Distribution of sample migrants by types
Education Level=Low
Education Level=High
Guang
NonGuang
Nondong
Neighbor Neighbor
dong
Neighbor Neighbor
219
344
331
211
244
195
23.74
24.2
10.5
17.35
2.74
6.85
4.11
6.39
4.11
100

28.2
15.12
22.38
12.79
9.01
3.78
4.07
2.33
2.33
100

31.42
22.66
18.43
6.04
5.74
3.63
5.44
3.93
2.72
100

87

24.64
32.7
15.64
9.95
2.84
4.74
2.84
3.79
2.84
100

33.61
23.36
18.44
5.74
8.61
5.74
0.41
0.82
3.28
100

43.08
22.56
14.87
2.05
2.05
5.13
3.08
3.08
4.1
100

Table 2.5. Probability of individual from region r to city j, by education levels
Education Level=Low
Education Level=High
Three
Guang
NonGuang
NonRegions (r)
dong
Neighbor Neighbor
dong
Neighbor Neighbor
City (j)
Dongguan
5.82
10.85
11.63
8.00
12.62
12.92
Shenzhen
5.93
5.82
8.39
10.62
8.77
6.77
Guangzhou
2.57
8.61
6.82
5.08
6.92
4.46
Foshan
4.25
4.92
2.24
3.23
2.15
0.62
Zhongshan
0.67
3.47
2.13
0.92
3.23
0.62
Zhuhai
1.68
1.45
1.34
1.54
2.15
1.54
Jiangmen
1.01
1.57
2.01
0.92
0.15
0.92
Huizhou
1.57
0.89
1.45
1.23
0.31
0.92
Zhaoqing
1.01
0.89
1.01
0.92
1.23
1.23
Total
24.50
38.48
37.02
32.46
37.54
30.00

88

Table 2.6. Imputed incomes for all observations by cities
City (N=1544)
Mean
Std. Dev.
Min.
Dongguan
1,443
215
1,094
Shenzhen
1,706
259
1,215
Guangzhou
1,502
306
972
Foshan
1,311
259
738
Zhongshan
1,419
307
711
Zhuhai
1,571
335
816
Jiangmen
1,657
433
801
Huizhou
1,531
242
1,065
Zhaoqing
1,501
395
855

89

Max.
1,915
2,375
2,104
1,805
2,608
3,006
3,057
2,313
2,381

Table 2.7. Travel distance and time, by destination cities
N=1544
Distance (km)
Time (hour)
City
Mean Min Max
Mean Min Max
Dongguan
800
4.4 3,463
9.3
0.2 34.8
Shenzhen
852 75.7 3,459
9.8
1.3 35.1
Guangzhou
756
8.7 3,438
8.8
0.3 34.5
Foshan
767
1.6 3,459
8.9
0.1 34.7
Zhongshan
833
5.0 3,526
9.5
0.3 35.5
Zhuhai
856
0.0 3,555
9.9
0.0 36.0
Jiangmen
802
0.5 3,518
9.1
0.0 35.3
Huizhou
822
2.3 3,369
9.6
0.1 34.1
Zhaoqing
771
1.6 3,514
8.9
0.1 35.1

90

Table 2.8. Summary statistics of city attributes, city level
Std.
Variable (N=9)
Description
Mean
Dev.
Urban_pop
Population of urban resident
(million)
4.2
3.0
Gov_exp_pc
Government expenditure per capita
(billion in 2009 RMB /person)
4.1
2.6
Hospital_num_pc
Number of hospital and health
institution per capita
197
67
PriMidSch_num_pc Number of primary and middle
school per capita
133
94
Unemployment
Share of labor unemployed (%)
2.3
0.4
Industry_num_pc
Number of industrial firms per
capita
4763 3014
Construc_num_pc
Number of constructional firms per
capita
366
269

91

Min.

Max.

1.3

8.8

1.1

9.1

125

318

40
1.6

322
2.8

1035

8930

123

807

Table 2.9. Distribution of sample migrants in nine destination cities
Population
Pop.
Sample
Sample
Weight
City
N.
Share (%)
N.
Share (%)
(1)
(2)
(3)
(4)
(5)
Dongguan
4,922,608
25.3
471
30.5
0.828
Shenzhen
5,848,539
30.0
350
22.7
1.324
Guangzhou
3,312,887
17.0
268
17.4
0.980
Foshan
2,206,538
11.3
141
9.1
1.240
Zhongshan
1,044,861
5.4
87
5.6
0.952
Zhuhai
581,476
3.0
74
4.8
0.623
Jiangmen
463,298
2.4
54
3.5
0.680
Huizhou
913,038
4.7
51
3.3
1.419
Zhaoqing
186,533
1.0
48
3.1
0.308
Total
19,479,778
100.0
1,544
100.0

92

Table 2.10. Estimation results of Network I on destination choice, without travel time
VARIABLES
Observations
Networks (%)
Income (ln)
Income*Unemployment
Unemployment
Industry_num_th
Construc_num_th

Hospital_num_th
PriMidSch_num_th
Gov_exp_pc
Urban_pop

City Fixed Effects
Log likelihood

W/O Interaction
(1)
(2)
13,896
13,896

W/ Interaction
(3)
(4)
13,896
13,896

0.032***
(0.005)
0.267
(0.224)
/
/
10.443***
(1.759)
-0.400***
(0.068)
17.147***
(2.758)
-6.959***
(1.084)
-6.898***
(1.025)
-1.564***
(0.252)
2.456***
(0.352)

0.029***
(0.005)
0.210
(0.230)
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/

0.030***
(0.005)
3.419***
(0.897)
-1.409***
(0.384)
21.039***
(3.476)
-0.394***
(0.068)
17.458***
(2.784)
-7.148***
(1.087)
-7.073***
(1.036)
-1.607***
(0.255)
2.521***
(0.353)

0.029***
(0.005)
3.163***
(0.924)
-1.320***
(0.396)
3.373
(3.017)
/
/
/
/
/
/
/
/
/
/
/
/

No
-2760

Yes
-2756

No
-2754

Yes
-2751

*** p<0.01, ** p<0.05, * p<0.1

93

Table 2.11. Estimation results of Network II on destination choice, without travel
time
VARIABLES
Observations
Networks (%)
Income (ln)
Income*Unemployment
Unemployment
Industry_num_th
Construc_num_th

Hospital_num_th
PriMidSch_num_th
Gov_exp_pc
Urban_pop
City Fixed Effects
Log likelihood

W/O Interaction
(1)
(2)
13,896
13,896

W/ Interaction
(3)
(4)
13,896
13,896

0.012***
(0.003)
0.289
(0.220)
/
/
9.743***
(1.743)
-0.418***
(0.067)
15.840***
(2.727)
-6.526***
(1.073)
-6.690***
(1.011)
-1.485***
(0.250)
2.464***
(0.351)

0.012***
(0.003)
0.205
(0.229)
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/

0.013***
(0.003)
4.007***
(0.868)
-1.674***
(0.375)
22.410***
(3.424)
-0.409***
(0.068)
16.342***
(2.759)
-6.799***
(1.077)
-6.928***
(1.024)
-1.548***
(0.252)
2.548***
(0.352)

0.013***
(0.003)
3.637***
(0.904)
-1.545***
(0.392)
3.436
(2.982)
/
/
/
/
/
/
/
/
/
/
/
/

No
-2765

Yes
-2759

No
-2757

Yes
-2752

*** p<0.01, ** p<0.05, * p<0.1

94

Table 2.12. Estimation results of networks on destination choice, with travel time
VARIABLES
Observations
First Type of Network
Migration cost
(Travel time in hours)
Network I (%)
Income (ln)
Income*Unemployment
Unemployment

City Fixed Effects
Log likelihood
Second Type of Network
Migration cost
(Travel time in hours)
Network II (%)
Income (ln)
Income*Unemployment
Unemployment
City Fixed Effects
Log likelihood

W/O Interaction
(1)
(2)
13,896
13,896

W/ Interaction
(3)
(4)
13,896
13,896

-0.698***
(0.068)
0.033***
(0.005)
0.334
(0.222)
/
/
11.482***
(1.777)

-0.783***
(0.073)
0.028***
(0.005)
0.216
(0.231)
/
/
/
/

-0.708***
(0.068)
0.032***
(0.005)
3.879***
(0.892)
-1.581***
(0.380)
23.407***
(3.471)

-0.784***
(0.072)
0.028***
(0.005)
3.343***
(0.935)
-1.398***
(0.401)
4.207
(3.053)

No
-2702

Yes
-2690

No
-2695

Yes
-2685

-0.671***
(0.078)
0.005
(0.003)
0.331
(0.218)
/
/
10.379***
(1.767)

-0.784***
(0.082)
0.004
(0.003)
0.185
(0.229)
/
/
/
/

-0.678***
(0.078)
0.005*
(0.003)
4.255***
(0.863)
-1.762***
(0.371)
23.742***
(3.417)

-0.780***
(0.082)
0.004
(0.003)
3.589***
(0.916)
-1.531***
(0.396)
3.620
(3.016)

No
-2720

Yes
-2703

No
-2711

Yes
-2696

*** p<0.01, ** p<0.05, * p<0.1

95

Table 2.13. Estimation results of networks and education selections
VARIABLES
Observations
Networks (%)
Edu_high*Network
Edu_high*Mig.cost
Migration cost
Income (ln)
Income*Unempl.
Unemployment

(1)
13,896

Network I
(2)
13,896

(3)
13,896

(4)
13,896

Network II
(5)
13,896

(6)
13,896

0.028***
(0.005)
/
/
/
/
-0.784***
(0.072)
3.343***
(0.935)
-1.398***
(0.401)
4.207
(3.053)

0.020***
(0.006)
0.016***
(0.006)
/
/
-0.786***
(0.073)
2.854***
(0.965)
-1.258***
(0.409)
3.156
(3.081)

0.028***
(0.005)
/
/
0.299***
(0.106)
-0.905***
(0.087)
3.194***
(0.946)
-1.373***
(0.403)
4.032
(3.062)

0.004
(0.003)
/
/
/
/
-0.780***
(0.082)
3.589***
(0.916)
-1.531***
(0.396)
3.620
(3.016)

-0.000
(0.004)
0.010**
(0.005)
/
/
-0.790***
(0.082)
3.789***
(0.922)
-1.616***
(0.398)
4.277
(3.029)

0.004
(0.003)
/
/
0.286***
(0.108)
-0.900***
(0.096)
3.438***
(0.925)
-1.502***
(0.398)
3.394
(3.023)

Yes
-2681

Yes
-2680

Yes
-2696

Yes
-2694

Yes
-2692

City Fixed Effects
Yes
Log likelihood
-2685
*** p<0.01, ** p<0.05, * p<0.1

96

Table 2.14. Estimation results of networks and micro-level nonfarm experience
VARIABLES
Observations
Networks (NW)
NW*Initial refer
NW *Current
refer
NW *Years
NFA
Migration cost
Income (ln)
Income*Unempl
.
Unemployment

(1)
13,896

Network I
(2)
13,896

(4)
13,896

Network II
(5)
13,896

(3)
13,896

(6)
13,896

0.030***
(0.006)
-0.004
(0.006)

0.028***
(0.006)
/
/

0.036***
(0.006)
/
/

0.002
(0.004)
0.003
(0.005)

-0.000
(0.004)
/
/

0.007
(0.004)
/
/

/
/

-0.000
(0.006)

/
/

/
/

0.008
(0.005)

/
/

/
/
-0.785***
(0.072)
3.320***
(0.936)

/
/
-0.784***
(0.072)
3.342***
(0.935)

-0.001**
(0.000)
-0.781***
(0.072)
3.367***
(0.937)

/
/
-0.779***
(0.082)
3.576***
(0.916)

/
/
-0.781***
(0.082)
3.557***
(0.917)

-0.000
(0.000)
-0.777***
(0.082)
3.592***
(0.916)

-1.393***
(0.402)
4.161
(3.054)

-1.398***
(0.401)
4.205
(3.052)

-1.420***
(0.402)
4.308
(3.044)

-1.526***
(0.397)
3.576
(3.018)

-1.516***
(0.397)
3.504
(3.019)

-1.529***
(0.397)
3.606
(3.016)

Yes
-2685

Yes
-2682

Yes
-2696

Yes
-2695

Yes
-2696

City Fixed
Effects
Yes
Log likelihood
-2684
*** p<0.01, ** p<0.05, * p<0.1

97

Table 2.15. Estimation results with networks and city amenities interactions
VARIABLES
Observations
Networks (NW)
NW*urban pop.
Urban population
NW*Number of
Industrial firms
Industry firm num.
NW* Number of
Construction firms
Constr. firm num.
Migration cost
Income (ln)
Income*Unempl.
Unemployment
City Fixed Effects
Log likelihood

(1)
13,896

Network I
(2)
13,896

(3)
13,896

(4)
13,896

Network II
(5)
13,896

(6)
13,896

0.126***
(0.028)
-0.014***
(0.004)
0.212***
(0.037)

0.049*
(0.028)
/
/
/
/

0.059***
(0.017)
/
/
/
/

0.014*
(0.008)
-0.001
(0.001)
0.176***
(0.022)

-0.008
(0.009)
/
/
/
/

0.003
(0.007)
/
/
/
/

/
/
/
/

-0.003
(0.004)
0.239***
(0.050)

/
/
/
/

/
/
/
/

0.001
(0.001)
0.320***
(0.039)

/
/
/
/

/
/
/
/
-0.729***
(0.073)
3.142***
(0.940)
-1.345***
(0.405)
7.279**
(3.037)

/
/
/
/
-0.780***
(0.072)
3.304***
(0.933)
-1.385***
(0.401)
8.191***
(3.022)

-0.054*
(0.028)
1.702***
(0.368)
-0.759***
(0.072)
3.184***
(0.937)
-1.354***
(0.403)
7.290**
(3.019)

/
/
/
/
-0.765***
(0.083)
3.613***
(0.918)
-1.547***
(0.398)
7.979***
(2.985)

/
/
/
/
-0.792***
(0.082)
3.488***
(0.919)
-1.492***
(0.398)
9.300***
(2.993)

0.002
(0.008)
1.706***
(0.231)
-0.782***
(0.083)
3.581***
(0.916)
-1.526***
(0.397)
8.236***
(2.984)

Yes
-2678

Yes
-2684

Yes
-2683

Yes
-2696

Yes
-2695

Yes
-2696

Note: all the variables with interactions are jointly significant
*** p<0.01, ** p<0.05, * p<0.1

98

REFERENCES

99

REFERENCES

Bauer, T., Epstein, G.S. and Gang, I.N., 2009. Measuring ethnic linkages among
migrants. International Journal of Manpower 30, 56-69.
Bayer, P., Keohane, N., Timmins, C., 2009. Migration and hedonic valuation: The case of
air quality. Journal of Environmental Economics and Management 58, 1â14.
Beine, M., Docquier, F., Ăzden, Ă., 2011. Diasporas. Journal of Development
Economics, Symposium on Globalization and Brain Drain 95, 30â41.
Bertoli, S., 2010. Networks, sorting and self-selection of Ecuadorian migrants. Annals of
Economics and Statistics/Annales dâĂconomie et de Statistique 261â288.
Borjas, G.J., 1987. Self-Selection and the Earnings of Immigrants. The American
Economic Review 531â553.
Chiquiar, D., Hanson, G.H., 2005. International migration, self-selection, and the
distribution of wages: Evidence from Mexico and the United States. Journal of
political Economy 113, 239â281.
Chiswick, B.R., 1999. Are immigrants favorably self-selected? The American economic
review 89, 181â185.
Dahl, G.B., 2002. Mobility and the return to education: Testing a Roy model with
multiple markets. Econometrica 70, 2367â2420.
Durand, J., Kandel, W., Parrado, E.A., Massey, D.S., 1996. International migration and
development in Mexican communities. Demography 33, 249â264.
Fafchamps, M., Shilpi, F., 2013. Determinants of the Choice of Migration Destination*.
Oxford Bulletin of Economics and Statistics 75, 388â409.
Grogger, J., Hanson, G.H., 2011. Income maximization and the selection and sorting of
international migrants. Journal of Development Economics 95, 42â57.
Massey, D.S., Goldring, L. and Durand, J., 1994. Continuities in transnational migration:
An analysis of nineteen Mexican communities. American journal of Sociology 99,
1492-1533.
McKenzie, D., Rapoport, H., 2010. Self-Selection Patterns in Mexico-U.S. Migration:
The Role of Migration Networks. Review of Economics and Statistics 92, 811â
821.

100

Moraga, J.F.-H., 2011. New evidence on emigrant selection. The Review of Economics
and Statistics 93, 72â96.
Pedersen, P.J., Pytlikova, M., Smith, N., 2008. Selection and network effectsâMigration
flows into OECD countries 1990â2000. European Economic Review 52, 1160â
1186.
Shaw, D., 1988. On-site samplesâ regression: Problems of non-negative integers,
truncation, and endogenous stratification. Journal of Econometrics 37, 211â223.
Stark, O. Jakubek, M., 2013. Migration networks as a response to financial constraints:
Onset, and endogenous dynamics. Journal of Development Economics 101, 1-7.
Wooldridge, J.M., 2002. Inverse probability weighted M-estimators for sample selection,
attrition, and stratification. Portuguese Economic Journal 1, 117â139.
Yunez-Naude, A., Taylor, J.E., 2001. The determinants of nonfarm activities and
incomes of rural households in Mexico, with emphasis on education. World
Development 29, 561â572.

101

CHAPTER 3: VALUE-CHAIN CLUSTERS AND AQUACULTURE
INNOVATION IN BANGLADESH

1. Introduction
Creators of economic innovations, such as in technologies, products, production
processes, business models, or new locations for markets, generate an advantage for
themselves that usually translates into higher than average profits. This is called the
âmiddleman modelâ (Lerner 1934, Just et al., 1979). Zilberman et al. (2017, this volume)
note that innovators make choices to design, as much as possible, the value chain that
âimplementsâ their innovation, in order to sustain and protect their middleman advantage.
This involves choosing whether to make or buy inputs or marketing services or indeed
the product itself; what technology to use if the innovator chooses to âmakeâ; and what
institutional and organizational approach to use if the innovator buys upstream or
downstream goods or services.
Sometimes an innovator needs actors upstream or downstream from the innovator to
themselves innovate in order to supply the specific goods or services needed â in volumes
or timing or other attributes such as variety and quality. We can call this âcollaborative
inter-segment innovation.â An example is where McDonalds innovated in Argentina in
the 1990s by standardizing large volumes of cooked French fries; it demanded
collaborative innovations by French fry producing companies (such as McCain colocating in Argentina) and the Argentine farmers supplying them. The innovation of the
processor and farmers was to form contract farming for a new variety of potato, the
Atlantic variety, which produces better processed fries (Ghezan et al. 2002).

102

By extension, farmers in developing countries who innovate either in adoption of a
new input or a new product as output, benefit from suppliers of inputs innovating in terms
of the input provided (e.g., setting up input dealerships to supply the new inputs farmers
then adopt) or the location of market (such as by co-location with the farmers).
Innovating farmers also benefit from suppliers of output marketing services innovating in
terms of adaptation to market new products and also of location of their services to be
near farmers. Moreover, farmers would implicitly benefit from these co-innovators
making their goods and services available in a timely manner and at low transaction cost
and risk of accessing them.
It is evident that the way that input suppliers, farmers, and wholesalers could most
easily co-innovate with the least risk and transaction cost is to take advantage of
economies of agglomeration by forming clusters. That formation can be a conscious
group strategy or an accretive and spontaneous act. Apart from lower transaction costs,
informational advantage also facilitates adoption of new technologies or better practices.
Marshall (1920) notes that technology spillover is easier in clusters than outside clusters
because one can see what others in the cluster are doing and imitate them. The prominent
work by Porter (1998, 2000) draws from Marshallâs (1920) idea of âexternal economiesâ
of co-located firms. Porter notes:
A cluster is a geographically proximate group of interconnected companies and
associated institutions in a particular field, linked by commonalities and
complementaritiesâŚ. The geographic scope of a cluster relates to the distance over
which informational, transactional, incentive, and other efficiencies occur. More than

103

single industries, clusters encompass an array of linked industries and other entities
important to competition. (Porter, 2000:16)
The above conceptual foundation has driven a first, substantial, strand of literature on
the formation of clusters, and clustersâ effects on technology (and product) choice of
actors in the clusters. Much of the literature on this effect is in the non-agricultural sector
(furniture (Sandee and Rietveld, 2001), metalworking (Kelley and Helper, 1999),
software (Forman et al., 2005), garments (Visser, 1996), and shoes (Schmitz, 1995)).
Relatively little has been done on food and agricultural products with a few exceptions
(e.g., Zhang and Hu, 2014 for potatoes in China). This literature shows that clusters affect
the productivity as well as the innovation and upgrading by actors in the cluster (such as
product âvalue ladderâ climbing and technology change to produce higher quality or
more consistent production). This can be in both âprocess upgradingâ (e.g. in the
production technology of an aquaculture farm) and âproduct upgradingâ (such as a shift
from a local niche to commodity non-traditional fish species and newly commercialized
niche species in my case). The inducement and facilitation of upgrading is a function of
the ease of access to needed inputs and services where actors in clusters are
complementary (Porter 2000; Humphrey and Schmitz, 2002; Sonobe and Otsuka, 2011),
and where intra-cluster vertical and horizontal relationships create âcollective efficiencyâ
(Schmitz, 1995). Schmitz (1999) distinguishes between âincidental external economiesâ
and âconsciously pursued joint actionâ such as contracts between traders and
subcontractors, or business associations.
A second strand of literature joins the literatures on clusters and value chains.
Examples include Humphrey and Schmitz (2000) and Pietrobelli and Rabellotti (2006).

104

The latter note that beyond the emphasis of the first strand of literature on local interrelations among firms in the cluster, there should also be a recognition that many of the
actors in the cluster are engaged in domestic or global value chains and are also
responding to price signals and governance institution (such as contracts and standards)
signals from upstream and downstream in the value chain. They present non-agricultural
cases and three agrifood cluster cases in Latin America - the dairy, fruit, and salmon
clusters - and focus on small and medium enterprises and farms linking to large-scale
international processors and wholesalers. The cases combine study of conscious joint
action in the small firm/farm clusters and institutional coordination with contracts and
standards by the large-scale actors.
While this second strand of literature is rich in case studies of cluster actors, there are
several gaps in the analysis. First, the studies tend to be of one or several clusters, but not
of multiple clusters of differential intensities (low versus high) of clustering and distances
from markets. Second, the studies tend to not examine inter-actor differences per
segment, thus not controlling for the effect of the cluster on actors of different asset
profiles. Both of these two sets of factors may substantially affect the effect of the cluster
on actor behavior.
To address these two gaps in the literature, as well as add to a literature with scant
agrifood cases, my paper tests for the effects of clusters of different intensities of
clustering on farm technology and product upgrading, controlling for micro level
variables reflecting differential asset levels, as well as for other (than cluster) meso
variables, such as distance to urban markets, the target markets for the value chain. While
there is a rich literature on farm technology adoption (see Feder et al. 1985; Liu 2013),

105

and some farm technology adoption literature that models the impact of social learning
from a density of adopters in a location or from networks (Bandiera and Rasul 2006,
Conley and Udry 2010), the degree of clustering in the location of the farmer has not
figured in the literature as a determinant of farmer choices in agriculture or aquaculture or
indeed the food sector.
To study that interaction I calculate and use as a determinant of farmer choice an
index to include both horizontal agglomeration and vertical interconnections among
actors in the value chain. The case is aquaculture farmers, feed and other inputs (e.g.
chemicals) traders, and fish traders (wholesalers) in Bangladesh. I rely on a unique data
set collected, that includes meso level information on the density of actors of different
segments over time and over districts, as well as micro level data from a large survey of
aquaculture farm households. To my knowledge this is the first time in the agricultural
economics literature that such a combined data set has been collected.
I study three innovations by farmers. The first is the adoption of âmodern inputsâ
central to intensification of aquaculture (commercial feeds, fuel for generators for pumps,
fertilizer, pesticides, lime, and vitamins), The second is the adoption of non-traditional
(exotic) commodity fish species (tilapia and pangas catfish) that grow very rapidly and
densely in ponds especially in combination with the modern inputs. This corresponds to
the âcommoditizationâ phase of the product cycle. The third is the adoption in
aquaculture production (instead of wild capture as traditionally done) of various niche
species. This corresponds to the âproduct differentiationâ phase of the product cycle.
Both of these are innovations relative to the traditional âextensiveâ, low-input

106

aquaculture production technology that has been focused on indigenous, slow-growing,
and low density carps (Ali et al., 2013).
I proceed as follows. Section 2 briefly provides background on the aquaculture sector
in Bangladesh. Section 3 discusses the survey and sampling. Section 4 explains the
clustering index. Section 5 discusses the econometric approach and provides descriptive
statistics. Section 6 shows the regression results and robustness checks. Section 7
concludes.

2. Background on Aquaculture in Bangladesh
Several points are important background to the subsequent discussion. First, demand
and supply for fish has grown fast. Rapid urbanization and increases in incomes have
spurred a strong increase in fish consumption (Belton et al., 2011; Reardon et al. 2014.
Moreover, the aquaculture sector is growing fast in Bangladesh. Over 1984 to 2014,
Bangladeshâs farmed fish jumped from 124,000 tons to 1.96 million tons, increasing by
1580%. As a result, aquaculture now accounts for 55% of Bangladeshâs fish supply, up
from just 16% three decades ago (The department of fisheries of Bangladesh, 1994; 1997;
2006; 2015). The great majority (94%) of aquaculture production is destined for domestic
consumption. (Nearly all the aquaculture product exports are of shrimp.) The farmed fish
market grew by a factor of 25 times in three decades to nearly 2 million tons today. At
most 10% of farmed fish are home-consumed, the rest are marketed. 42% of marketed
farmed fish is consumed in urban areas and that share is growing quickly (Hernandez et
al. 2017).

107

Second, aquaculture fish product composition has changed substantially over the past
two decades. Carps are the traditional species in Bangladesh. As a response to pond water
resources limitations (for extensive aquaculture) and due to disease outbreaks among
shrimp and carp (Hussain, 2009), two species were introduced into Bangladesh
aquaculture that allowed for intensification based on high densities of fish combined with
use of various external inputs such as feed and chemicals. Pangasius catfish were
introduced in 1989 from Thailand and spread quickly over farms thereafter. Tilapia
production increased more than three fold from 2005 to 2007 due to the rapid expansion
of tilapia hatcheries (Hussain, 2009). Both of these newly introduced species grow fast,
perform well under a range of aquaculture conditions, and tolerate poor water quality.
They thus can be farmed at higher intensities and attain higher yields than carps.
Moreover, niche varieties that used to be wild-capture fish have been introduced
mainly over the past decade into Bangladesh aquaculture and now marketed. This is part
of the long term process of transition from hunting/fishing to farming, and the
short/medium term process of transition over the product cycle, from initial traditional
niche local products in aquaculture (different kinds of carp) to exotic commodities
(tilapia, pangas), to differentiated products that include diverse tasty fish from wild
capture now laboriously âdomesticatedâ into ponds.
At the all-sector level of fish farming in Bangladesh, tilapia and pangasius
experienced dramatic diffusion in the 2000s (Ali et al., 2013). Our survey corroborates
this. Table 3.1 shows that from 2008 to 2013, the share of carp in total farmed fish tons
produced by the sample households decreased largely from 50.8% to 39.8%, while the

108

shares of tilapia, pangas, and niche fish species increased sharply. Carps output doubled,
but tilapia niche varieties tripled, and pangasius quadrupled, only over five years.
Third, farmers have gone from a traditional system in which there is no use of added
feed or merely using manure or fertilizer to induce phytoplankton thence zooplankton
growth on which fish feed, to use of supplementary feeding (with rice bran and other
simple agricultural processing byproducts) and formulated feed concentrates. Mamun-UrRashid et al. (2013) estimate that the production of pelleted feeds increased at 32% per
year over 2008-2012, to reach 1.07 million tons in 2012. Another 0.3-0.4 million tons of
farm-made feeds are estimated to have been produced in 2012. They note that sinking
feeds7 accounted for 81% of feed output, but the rest, the more efficient floating feed, has
grown at 89% yearly over that period.

3. Survey and sample
I draw on three sets of data. The first is secondary data on the geography of the
clusters. The second is meso-level (district level) primary survey data on changes in the
numbers of supply chain segment actors. The third is a primary survey of aquaculture
farm households. Teams from IFPRI and MSU undertook the two primary surveys in
2014.
Five main segments of the aquaculture value chain in Bangladesh are studied at the
meso level (fish farmers, rural fish traders (wholesalers), input dealers, hatcheries, and

7

Sinking feed sinks to the bottom of the pond. Floating feed is produced by extrusion in such a way as

to float. Floating feed is much more efficient than sinking feed as a much higher share of it is consumed
rather than wasted.

109

feed mills). The micro survey of aquaculture farm households in 2014 with recall of 2013
and 2008 is the only segment examined in the micro analysis.
Both the meso and micro data come from six broad aquaculture areas in Bangladesh.
These six areas are the major areas of production and concentration of fish farmers and
fish production in Bangladesh. The sample was chosen in the following steps.
First, the six main aquaculture areas were identified using official statistics and our
own analysis of the 2008 agricultural census. We then selected all the districts (first-level
administrative division of Bangladesh, analogous to the âcountyâ level in the US or UK)
in these six areas. Each area contains several districts, ranging from one to five districts.
The details can be found in the footnote.8
Second, within each district, we eliminated upazilas (townships) that have few
aquaculture ponds and/or few fish farmers. We thus eliminated 46% of the upazilas, but
kept the upazilas representing 84% of the total production area in all clusters.
Third, after eliminating upazilas with few ponds, from the remaining universe we
randomly selected upazilas using proportional probability sampling. Once upazilas were
selected, we listed all mouzas (village level, which is below the township) within the
selected upazilas. We then eliminated from the universe of mouzas all with less than 20
aquaculture farmers. Once the list of mouzas was trimmed, we randomly selected two to
three mouzas per selected upazila.
Fourth, once mouzas (PSUs) were randomly selected, we sent a team of enumerators
to each mouza to do a rapid census of aquaculture farmers within the mouza. Once we
collected the list of farmers in each mouza, we randomly selected 25 farmers per mouza
8

Area 1: Khulna, Jessore, Satkhira, Bagerhat, and Gopalganj; Area 2: Barisal, Bhola, Comilla,
Chandpur, and Noakhali; Area 3: Narsingdi, Gazipur, Brahmanbaria, and Mymensingh; Area 4: Chittagong
and Coxâs Bazaar; Area 5: Bogra, Natore, and Dinajpur; Area 6: Sylhet.

110

(20 farmers for our sample, plus 5 replacement farmers in case we failed to find any of
the 20 selected farmers). In total, we obtained 1,540 aquaculture households from 32
upazilas, and 77 mouzas.
Since I focus on fish farmers, I dropped farmers producing shrimp only, and farmers
producing no fish in 2013 and 2008. The resulting sample is 1,514 fish farming
households as the micro-level information, distributed in 20 districts, 32 upazilas, and 77
mouzas (villages). See Figure 3.1 for the location of the 20 sample districts. The selected
mouzas are representative of 86% of the fish pond areas in the districts selected. In turn,
the districts selected constitute 61% of all fish pond production in the country, based on
data for 2014 from the Bangladesh Bureau of Statistics.
The sample design for the inventory over years of the value chain actors (noted
above) is based on the 77 mouzas, which were selected for the household sample. In each
mouza (village), the survey conducted 2-5 interviews with key stakeholders concerning
the number and size distribution of value chain actors at the upazila and district level over
2003, 2008, and 2013. The survey was conducted in 2014 to collect the data in 2013, and
the same information was recalled for 2003 and 2008. To address the concern over
measurement errors I use the cross-section analysis with information in 2013 in my
regressions, while I show statistics in the descriptives section for both years to compare
the patterns. The meso survey collected three categories for âsize of the actorâ: small,
medium, and large.
Table 3.2 presents the definitions of the size category by type of actor. Based on the
average volume of transactions or size of actors (column 2) and their size categories
(column 3), I generate the district level total transactions/sizes of actors, which are used

111

to calculate the clustering index at the district level. Thus in total I have 20 clustering
levels for 20 sample districts. The results concerning the cluster indexes are discussed
below.

4. Clustering index and its patterns
4.1. Calculation of clustering index
Various indexes are used to characterize economic activity in a given area. The
Hirschman-Herfindahl index is used for concentration; the Krugman Specialization Index
for sectoral specialization; and the Agglomeration Index (Uchida and Nelson, 2009) for a
combination of population size (of a city) and density of the population. None of these
indexes gets at spatial agglomeration of diverse value chain segments and thus potentially
how value chain actors in the same locality have high access to each other for interconnection (Long and Zhang, 2012). To fill this gape I calculate a clustering index at
district level to capture the number as well as the relatedness of actors in the different
segments of the aquaculture value chain in each district. My index is based on Long and
Zhang (2011) and Delgado et al. (2015).
Figure 3.2 shows the share of number of actors by segment per thousand rural people
in the sample districts. One can see that fish farmers, input dealers, and fish wholesalers
tend to co-locate in the same areas (northern and southern areas circled in yellow).
However, feed mills tend to be relatively concentrated in the North, the most
developed aquaculture area in terms of buildup over time. Following industrial
agglomeration economy theory (Henderson, 1974), the geographic concentration of feed

112

mills is due to economies of scale in the feed industry, which is different from the other
actors in the aquaculture value chain.
Hatcheries tend to be spatially scattered, with high densities in three places
(Mymensingh in the North, Coxâs Bazar in the East, and Jessore in the Southwest). This
locational concentration could be explained in two ways. First, hatcheries, especially
shrimp hatcheries depend heavily on natural sources, such as saline seawater, and thus
not many places are suitable for them. Second, path dependency is a key factor in the
development of hatcheries. Hatcheries started in places where there was a history of
harvesting, nursing, and trading wild fry (baby fish captured from rivers, instead of
hatchery-produced).
Based on these dissimilarities over the spatial distribution of segments, to capture
relatedness of actors within the cluster, I do not include hatcheries and feed millers in the
calculation of the cluster index. Instead I include the three segments (farmers,
wholesalers, and input dealers) with similar spatial concentration patterns. Specifically, I
use the following six variables in 2008 (the lagged period five years before the survey)
from the meso-level data to calculate the clustering index. I use Principal Component
Analysis (PCA) to identify the underlying relationship among the three segments.

V1: Total number of fish farmers in each district normalized by district rural population
V2: Total area of fish ponds in each district normalized by total area of the district
V3: Total number of traders in each district normalized by district rural population
V4: Total volume of fish transactions in each district normalized by district rural
population

113

V5: Total number of input dealers in each district normalized by district rural population
V6: Total volume of input transactions in each district normalized by district rural
population

To avoid the concern that absolute values will favor large aquaculture districts, which
naturally have more people working in the aquaculture value chain, the number of actors
and total quantities of transactions are normalized by the districtâs rural population.
From the results of the PCA method, I adopt the first component with an eigenvalue
greater than one to calculate the clustering index in 2008, which is a composite value as
the sum of the weighted average of the six variables. To compare the clustering levels
and changes over years, from 2003, to 2008, and to 2013, I apply the obtained weights for
six variables in 2008 to the same normalized variables in 2003 and 2013. The calculated
composite indicators from the PCA method range from -2.05 to 5.91. To make the
indicators more comparable and meaningful, I first generate the district-level (d) z-scores
based on the three-year (t) pooled indicators,

đ§_đ đđđđ!" =  Â  (đźđđđđđđĄđđ!" â đźđđđđđđĄđđ!"#$ )/đźđđđđđđĄđđ!"#

Then, I normalize the z-scores into 0-1 to avoid the negative value. The clustering
index (CI) is calculated as follows,

đśđź!" =  Â  (đ§_đ đđđđ!" â Â đ§_đ đđđđ!"# )/(đ§_đ đđđđ!"# â đ§_đ đđđđ!"# )

114

The calculated clustering index ranges from zero to one. The greater the value, the
higher is the clustering level. Specifically, the Mymensingh district has a value of one in
2013, which means that the clustering degree in 2013 of this district is the highest among
all other sample districts in the three years. It is a relative value, and does not imply that
the clustering level of Mymensingh cannot be higher if more actors establish in it.
In total, I have 20 district-clusters (called clusters hereafter). Each district is regarded
as a cluster with a unique clustering index.
4.2. Clustering patterns
Table 3.3 reports the patterns and changes of the clustering index by districts over the
three years spread over a decade. The results suggest that the degree of clustering varies
greatly among the 20 districts well known for fish farming. Recalling that the 20 sample
districts were selected from six aquaculture areas in Bangladesh, the results show a great
variation of clustering degree among districts even within the same area. Here, to
simplify the description of the patterns, I map the 20 sample districts into four geographic
zones based on cardinal points (North, Southwest, South center, and East zones) that
correspond to regional development debate in Bangladesh, rather than the six sampling
areas discussed in the sample section.
Specifically, among all the districts, Mymensingh, in the North, always has the
highest degree of clustering for three years (2003, 2008, and 2013), yet the clustering
degrees of other districts in the North are relatively low. This indicates a large and
uneven spatial concentration of fish farming areas in the North. This result is explicable
as Mymensingh is a traditionally important area for aquaculture. With the availability of
numerous hatcheries (Ali et al., 2013), large feed mills (Belton et al., 2011), favorable

115

agro-ecological conditions, and access to aquaculture research institutions (Ahmed,
2009), Mymensingh is ranked first among districts of aquaculture farming in Bangladesh
(Ahmed and Toufique, 2015).
Unlike the spatially uneven concentration of clusters in the North, all four districts in
the Southwest have a relatively high degree of clustering. This is particularly the case for
the district of Khulna, whose degree of clustering is always ranked No.2, following
Mymensingh in the North, during the three-year measures during the decade. The
districts in the East have low degrees of clustering.
In the three traditional fish farming districts (Mymensingh, Khulna, and Gopalganj)
with high levels of clustering, their growth in the degree of clustering is also high. This
implies that the initial fish farm value chain segments clusters are developing their
clustering faster than the more recently established clusters. There are racers even among
the top clustering districts: specifically, one district (Bhola in the Southcenter) recently
(over the decade) rose to become among the top five clustered districts.
To summarize, the clustering indices of 20 districts in 2013 shown in Table 3.3 are
mapped in Figure 3.3. Three zones with districts with higher level of clustering degrees,
including North (Mymensingh) area, Southwest (Khulna) area, and South Center (Bhola),
are circled in red. Based on the measure of clustering, the populations of fish cluster
actors as well as vertical relatedness of fish producers and service providers are greater in
these regions.

116

5. Regression estimations
5.1. Regression specification
The regressions address how the degree of district-level clustering affects fish
farmersâ adoption of innovations, specifically the use of modern inputs (for farming
intensification) and product composition changes â in particular the production of nontraditional commodity species and ânicheâ commercial species. As discussed before, to
address the concerns of measurement errors I use the cross-section analysis with
information in 2013 in my regression analysis, while I show descriptive statistics for both
2008 and 2013 to compare the basic changes.
Apart from clustering, I test several other control variables that are assumed to affect
farmersâ behaviors. The increasing demand for fish particularly in cities in Bangladesh
suggests that urban proximity may affect all three farmersâ behaviors that I am testing.
These hypotheses give us the following reduced form for the adoption decision of
innovation k of farmer i in district j,
!
đ´đđđđĄđđđ!"
= đ(đśđź! , đđ!" , đ!" , đ! , đ, đ! , đ!" )

Modern inputs and new species are regarded as two sets of innovations (k) in this
study. First, as noted above, I define modern inputs as fuel, fertilizer, pesticide, lime,
purchased feed, and vitamins. According to the survey, a sample householdâs share of the
expenditures on the modern inputs over total expenditures during fish production (from
pond preparation to harvesting stages) is 32.6%, without imputing the family labor cost.
In the analysis, in one variation, I also separate the purchased feeds from total modern

117

input. Two input-related variables are used as dependent variables, including adoption of
the modern inputs, and the expenditure on modern inputs per acre. The former captures
the adoption decision directly, and the latter stands for the technology intensification via
addition of modern (variable) inputs.
Second, because 95% of the sample households produced more than one species (on
average, a household farmed four fish species), rather than analyzing the binary adoption
decision, I model the share of output by species sets. The species sets I analyze are carps,
tilapia, pangasius catfish, and (newly commercialized) niche fishes. I expect the
coefficient of the degree of clustering on the carps share to be negative, and the
coefficients on the tilapia and pangasius shares to be positive for the reasons discussed
above.
Moreover, I include a regression explaining the number of species produced by the
household as an index for specialization. I expect that the effect of clustering on the
number of species to be negative: the higher the clustering is, the more speciesspecialized the household, and thus economic gains from specialization to be facilitated
by clustering.
CIj stands for the calculated clustering index in district j. I hypothesize that fish
farmers in areas with greater clustering are more likely to adopt innovations, due to
reduced risks and transaction costs, and due to the âincidental external economiesâ that
de facto equal collaborative innovations over segments in the supply chain.
UPij captures urban proximity. With the geographic coordinates of sample
households, I used Google Maps to calculate travel time from households to the three
biggest cities in Bangladesh (Dhaka, Chittagong, and Khulna). I adopt the minimum

118

travel time of the three to measure the urban proximity of household i in district j. I
assume that the coefficient on adoption is negative, and thus being close to a large city
will induce adoption of innovations. This would be an effect similar to what Fafchamps
and Shilpi (2003) found for proximity to cities on farmerâs fertilizer use. The positive
impact of urban proximity may be because of the payoff to modern input use due to lower
transaction costs to access markets and meet their demand for increasing volumes, and
the payoff to non-traditional species adoption to meet the changing diet patterns in cities
consistent with Bennettâs Law.
Xij contains a vector of household controls including gender and education of the
household head, family size, total area of crop lands and ponds, and nonfarm
employment. The latter is expected to favor adoption of modern inputs as nonfarm
employment is a cash source for inputs in the presence of credit constraints. It also
permits overall income risk management (Reardon et al. 1994). The hypothesis on the
land variable is, however, ambiguous. Wealth can reduce risk aversion and thus spur
adoption of innovations. But it may be the smaller aquaculture farmers who intensify
with modern inputs and maximize revenue with profitable non-traditional species that are
best for densification9 of fish stocks in a pond.
Vj is a district-level variable: it is total length of paved roads, used as a proxy for road
access. To avoid the potential endogeneity of roads (such as unobserved district level
economic factors affecting both adoption and roads), I use the five-year lagged value.

9

Densification is a subset of what in economics is intensification. One cannot increase the density of
fish per acre, but adding non-land inputs is intensification. Holding non-land inputs apart from seed
constant, increasing seed increases density as an outcome, which is also intensification.

119

To capture the dissimilarities among zones, I control for locational fixed effects by
adding zone (Z) dummy variables. đ! contains district level unobservables, and đ!" is an
idiosyncratic error term.
I estimate the adoption regressions with Probit, and the other equations with OLS.
The heteroskedasticity robust standard errors are reported. My results are robust to Tobit
and Fractional analyses, even when the dependent variables are in the fractional (share)
forms.
5.2. Econometric issues and their resolution
There are two main econometric issues and for each I note how I address it.
First, a potential concern in the above analysis is the existence of unobserved district
level variables (đ! ), which may affect the degree of district clustering and the farmer
adoption decision. Unobserved aquaculture-related natural advantages like water quality
and disadvantages, like floods, could affect the number of actors or their linkages in the
cluster, or reduce the incentive for technology adoption by farmers. In that case, my
estimated coefficient on the clustering index would be biased. To address this concern, I
construct a set of district level instrumental variables based on the characteristics of crop
clusters.
As rice and fish are the two fundamental agricultural sectors in Bangladesh, districts
with more aquaculture clustering may have lower (substituted) or higher level (rice-fish
inter-production) crop clusters. That is to say that the district-level crop clusters are
closely correlated with the fish clusters. Moreover, I assume that the level of crop
clustering will not directly impact fish farmerâs technology or species adoption decisions
as it is very unlikely that crop input dealers sell fish feeds (based on they fish trader

120

survey, only 4% of traders also trade some other agricultural products except for fish).
Based on these two conditions, the crop clustering is regarded as the instrumental
variables (IVs) for fish clustering, which is consistent with the exclusion restriction of
using IVs.
Similar to the calculation of the fish clustering index, I use two normalized districtlevel variables as the IVs to capture the degree of crop clustering in a district, including:
1) the quantity of inputs used for crop production, to capture input provider-crop
producer relations; and 2) quantity of crops sold, to capture output trader-crop producer
relations. These variables are generated from the nationally representative household
income and expenditure surveys (HIES) conducted in 2005 and 2010 by the Bangladesh
Bureau of Statistics.
Since the adoption equation is estimated using a Probit (non-linear) model, I use
control function estimation instead of two-stage least squares. However, to first test the
endogeneity of the clustering index and the validation of the IVs, the first stage
regression result is shown and the results using different IV methods are reported and
compared in Table 3.A2. The Hausman tests imply the need to use IVs. The significant
impacts of the two IVs on the clustering index show that the IVs are not weak IVs. The
negative impacts show that the districts with higher clustering of aquaculture value chain
actors have fewer crop inputs and crop outputs transactions or lower levels of crop
clustering. This is as expected, since fish production requires more ponds while crop
production requires more crop lands to produce, thus one district might be hard to be
agglomerated by two types of clusters. The similar coefficients and standard errors of the
estimators from three IV methods, including two-stage least square (2SLS), limited

121

information maximum likelihood (LIML), and generalized method of moments (GMM),
suggest that the IVs are proper. Although I reject the over-identification restriction at the
5% level, meaning that one of the IVs is likely to be invalid when assuming another IV is
valid, I will report results both with and without IVs for all the models in the results
section, so the differences can be compared.
Second, an endogeneity problem could be caused by reverse causality. Ideally, I
would not worry about the reverse causality, since it is less likely that a famerâs adoption
of innovations (micro-level) would affect the clustering of all actors (macro-level). But in
the case of the extreme monopoly situation as described in the middleman model
discussed in the introduction, I provide an additional robustness check by using only the
sub-sample of those who produced fish in 2013 but not in 2008. By doing so, I exclude
the concern that adoption of innovations of farmers happened earlier than the existence of
the fish clusters.
5.3. Descriptive results
Table 3.4 shows the descriptive statistics of household and district variables.
Although I do not use 2008 recall information in the regressions, I report two annual
statistics here to show basic changes.
The first three columns of Table 3.4 show the characteristics of sample households
and districts for two years and significance tests between the two. The right six columns
compare households in districts with lower and higher levels of clustering for two years.
The degree of clustering grew significantly from 2008 to 2013, particularly for
districts that have been important in aquaculture for the past two decades. This implies
that the initial fish farm value chain segments clusters are being developed and

122

strengthened, although the growth rate is also high in the recently established clusters
where the clustering degree is still low. These findings are consistent with what are
shown in Table 3.2, and the reasons are discussed as above.
Households differ by the degree of clustering of their districts in several interesting
ways. First, higher clustering districts have a significantly higher share of participation in
nonfarm activities.
Second, they tend to be within three hours of one of the three large cities (Dhaka,
Chittagong or Khulna); clustering of the district and proximity to a large city are
correlated. Households in the districts with lower clustering are twice as far (in travel
time terms) as those in highly clustered districts. All else equal, it is reasonable that fish
farm clusters are nearer urban demand points.
Third, infrastructure density is greater in higher clustered districts. But in both low
and high clustering districts total paved road length has increased similarly in the past
five years. Economies of agglomeration appear to be a function of commerce and
transport transaction costs, controlling for other factors like natural endowment of water
of the zone.
Fourth, more than 70% of households in higher clustering areas own some ponds
alone (rather than in groups of households), yet that share is only 50% for households in
lower clustering areas. Farmers with sole ownership may perceive a greater ability to
appropriate investments in modernizing their operations. Yet also the share of farmers
renting rather than owning is higher in high cluster districts. Yet I did not find significant
differences in aquaculture farm size (operated pond area) and years of pond operation
between low and high clustering areas. This suggests that in high clustering areas farms

123

are being added and existing farms are undergoing intensification, compared with low
cluster areas.
Table 3.5 provides descriptive statistics on farmersâ adoption of modern inputs. The
adoption rates between high and low clustering districts are significantly different.
Around 90% (average for two years) of farmers in high clustering areas adopt at least one
type of modern input (commercial feed, fuel, fertilizer, pesticide, lime, vitamins) versus
70% for households in low clustering areas. This pattern is also true for farmersâ adoption
of commercial feeds alone and modern inputs outside of commercial feeds.
The shares of expenditures on modern inputs over total input expenditures during fish
production (from pond preparation to harvesting stages) are also greater for farmers in
high degree clusters, no matter whether we include the imputed cost of family labor.
These results make sense for the reasons we discussed above.
For carps, a lower share of households in higher clustering areas grows them, as
expected. On average, the share of carps in output was 63% for households in high
clustering areas, more than 10% lower than those in low clustering areas.
For pangas, the opposite pattern emerges. As farms are multi-species (four on
average), the share of farms growing pangas is actually higher in low cluster districts but
the share in overall output is lower.
The findings for tilapia are, however, contrary to my initial expectations. The table
shows that the share of farms and of total output in higher clustering areas are actually
both somewhat lower than in low clustering areas. There are two possible reasons for
this.

124

On the one hand, Table 3.3 showed that Mymensingh district in the North and
districts in the Southwest have high degrees of clustering. But districts in the Southwest
have a disproportionate focus on shrimp (in Khulna, Satkhira, Bagerhat) and carp
(Jessore); the Mymensingh district is specialized in pangasius, niche species, and carps.
Hence, tilapia is farmed mainly in polyculture in these high cluster districts. Also, around
16% of farmers produce shrimp in high clustering areas, while almost none do so in low
cluster districts.
On the other hand, there are localized pockets of intensive tilapia (as opposed to
polyculture) production that emerged relatively recently in the Northern and Eastern
zones, but mainly on small farms outside our district-level clusters, according to key
informants. I do not have data on these localized pockets.
Finally, the number of species per household, measuring householdâs specialization
level, is not statistically different by degree of clustering. This is consistent with the
finding of species polyculture in all districts.

6. Results of Regressions
6.1. Adoption of modern inputs
Table 3.6 shows the regression results on adoption of modern inputs. The left two
columns show households adopt some modern inputs if they use any commercial feed,
fuel, fertilizer, pesticide, lime, or vitamins. The middle two columns exclude commercial
feeds, and the right two columns show adoption of commercial feeds only. Both results
are shown with and without use of control function estimation to take care of the
endogeneity issues are reported.

125

The salient findings are as follows. First, a major finding of the paper is that
regardless of how I specify the modern input set and estimation method, the degree of
clustering has a positive and significant effect on use of modern inputs, especially
commercial feed. For the coefficients of the clustering variable, three Probit estimations
using IVs show much larger estimators than models that do not use IVs, but the signs and
significant levels of the estimators are consistent between the two methods. The reasons
for these effects were discussed above in the conceptual framework.
Second, another major finding is the significant negative coefficients on the distance
to major city and adoption of modern inputs, especially again on commercial feeds. In
terms of the estimations, I find the estimators of travel time to the nearest big city from
the IV-Probit method are smaller than those from the regular Probit method. The impact
of city distance and clustering degree can also be observed using per acre expenses on
modern inputs, shown in Table 3.7. The three OLS-IV regressions may not provide good
results, as the coefficients of the residuals are not significant. Therefore, I will explain the
results here based on OLS estimations only. Similar to adoption of modern inputs,
households in high clustering areas and close to a big city use intensification technologies
(with more external inputs). As above, the impacts on commercial feeds are greater than
other modern inputs.
Moreover, Table 3.6 and Table 3.7 show that fish farmers with larger areas of crop
land are more likely to adopt modern inputs and spend more on modern inputs per acre
on their fish farms. Larger farms imply more wealth and potentially better access to credit
as well as cash from crop sales, thus are less likely to be constrained by the liquidity
constraints to adopt and intensify in modern technology.

126

6.2. Adoption of nontraditional commodity and newly commercialized aquaculture
niche species
Table 3.8 shows the results of regressions explaining the quantity shares of carp,
tilapia, pangasius catfish and niche fishes. As shown in Table 3.5 that the share of pangas
volume is low, I will combine tilapia and pangas together as modern species.
First, an important finding is that fish farmers in districts with higher clustering have
significantly lower shares of carp production, since carps belongs to the traditional
species in Bangladesh. But the clustering degree significantly increases the output share
of tilapia and pangasius species combined, as well as the newly commercialized niche
fishes. The positive impacts on these non-traditional species are consistent with my
assumption that clusters help adoption of or specialization in non-traditional species,
probably through decreased costs to produce/sell these species through clusters of actors
in the supply chain. The coefficients of residuals of the first two species sets are
significant, so I interpret the OLS-IV results for carp, tilapia and pangas regressions, but
OLS for niche regression.
Interestingly, I find significantly negative impacts of travel time to a big city on
output share of traditional versus non-traditional species. This result is again consistent
with my hypothesis that proximity to big cities would increase the adoption of more
advanced species to meet the changing urban diets. The results thus emphasize the
importance of both supply side (clusters) as well as demand side (urban proximity) on
intensification in non-traditional farmed fish species in Bangladesh.
Additionally, I explore the impacts of the clustering degree on adoption of species
composition conditional on types of farmers. The hypothesis behind this is that although

127

clusters decrease the output share of carps, it is possible that carp producers in areas with
higher degrees of clustering are more likely to adopt advanced inputs. This hypothesis is
confirmed by the results shown in Table 3.9. Specifically, both carp farmers, tilapia and
pangas producers are more likely to adopt modern inputs, especially commercial feeds, in
districts with higher clustering degree, and being close to a big city. This is particularly
the case for carp farmers, the results of which are robust to methods with and without
IVs.

6.3. Specialization versus diversification of fish production
Table 3.10 shows the regression results on specialization. Householdâs specialization
level is measured as the number of species farmed by the household. Several points stand
out.
First, clustering of the district induces farmerâs species specialization significantly. It
could be that output traders persuade farmers to specialize in species they need, or easy
access to feed dealers helps farmers to specialize in high stocking density feed-requiring
non-traditional species, all else equal.
Second, fish farmers living close to a big city are more diversified. Diversified fish
farmers may lose on economies of specialization but gain on flexibility to meet
diversified and evolving urban fish markets.

6.4. Robustness check
In this section, I show the results using a subsample to address the potential concern
about reverse causality, in particular that the adoption of innovations by farmers as a first

128

step might then induce clustering of off-farm actors to meet their needs. To avoid the
situation that adoption happened before development of clusters, I use fish farmers that
produced fish in 2013 only but not in 2008. By doing so, one can safely assume that the
adoption of fish production-related innovation will be affected by the degree of
clustering, instead of vice versa. As before, each model is estimated using two methods:
with IVs and without IVs.
In Table 3.11, I show the results with a full sample and a sub-sample. Most of the key
results (in bold) are consistent between the two samples and the two methods.
Specifically, clusters increase the adoption of commercial feeds for all sample
households, including both carp farmers and more non-traditional farmers who adopt
tilapia and pangasius. Clusters also increase householdsâ production of non-traditional
commodity species and specialization in fish species.
As before, urban proximity plays significantly consistent roles in spurring adoption of
commercial feeds and of non-traditional species.

7. Conclusions
Farmers adopting and implementing innovations, such as new technologies and new
products, often require âcollaborative inter-segment innovationâ by other actors in other
segments of the value chain, such as wholesalers implementing new product innovations
such as supply of commercial fish feed and chemicals and buying and marketing nontraditional fish species. I tested whether the clustering, and thus economies of
agglomeration with implied lower transaction costs, encourage and facilitate farmers to
innovate in technology choice (adoption of modern inputs) and product choice (adoption

129

of non-traditional âcommodityâ species exotic to Bangladesh (tilapia and pangasius
catfish) and of newly commercialized and-pond farmed local niche species). Clustering in
the meso environment of the farm as a potential determinant of farmer choices has not
been studied in agriculture or aquaculture or indeed the food sector.
I used a unique data set from our own primary survey of the aquaculture value chain
in Bangladesh, including micro data for 1500 fish farm households and 20 districts (77
villages) for meso level data. I calculated an index to include both horizontal
agglomeration and vertical interconnections among actors in the value chain.
The analysis gave rise to several key findings. First, I found that being in an area with
a high clustering index is associated with a higher probability of farmers intensifying via
use of more modern inputs (especially commercial feed) and making the fish species
product choices noted above. This emphasizes the importance of value chain off-farm
components for farm modernization.
Second, I found that farmers in clusters nearer to big cities have twice the propensity
to make these innovations compared to farmers far from the big cities.
Third, I found that the degree of clustering spurs specialization (or a reduction of the
widespread species diversification practiced), allowing relative gains from economies of
specialization.
Fourth, controlling for the meso variables of clustering and distance to cities, I found
that while ponds area does not drive adoption of the modern inputs (hence there is not an
âexclusionâ factor in fish farm size unlike what we expected), there is a positive
correlation between lagged crop farm size (part of the overall household farming
operation of fish farmers) and rural nonfarm employment. These two sources of cash to

130

invest in fish farming implies the possibility of local credit market constraints and thus
need for reliance on own cash sources (a common finding in the crop literature, see
Reardon et al. 1994).
The policy and research implications of the findings are as follows.
First, a key implication of the approach and findings is that research would benefit
from increased attention to meso level variables, including clustering in particular and
economies of agglomeration in general, and distance to urban areas, as a way of enriching
study of the microeconomics of product composition and technology choice. This will be
increasingly important as large-scale urbanization as well as local small town
development proceeds apace in developing regions.
Second, the research reaffirms and extends the importance of study of rural nonfarm
employment on farm technology and product choices. Again, this will be increasingly
important as the rural nonfarm labor market develops incessantly in these regions
(Haggblade et al. 2010).
Third, a key implication for policy is that development of value chains in general, and
the local off-farm components of those chains, such as input dealers and output
wholesalers, would best be important parts of strategies of agricultural and commerce
ministries to spur modernization of farming as well as alignment of fish species adoption
with patterns of urban food demand.
Part and parcel of that development is the array of infrastructural investments â and
other âenabling businessâ policies (World Bank, 2017) that facilitate the commercial
environment for clusters and value chain actors to invest. As I found that older
established fish farm areas had more rapid cluster development, it is important for these

131

policies to level the playing field for the newcomer aquaculture districts to catch up and
develop rapidly.
Finally, I see as complementary the findings concerning meso level determinants of
farm modernization, and extant policy prescriptions â such as extension and collective
asset provision - for building capacity of farms to adopt technologies and shift product
mix to take advantage of opportunities in product markets â and take advantage of the
opportunities afforded by the reduction of transaction costs to reach new markets due to
clustering itself.

132

APPENDIX

133

APPENDIX

Table 3.1. Farmed fish output of sample households, by species
2008
2013
2008-2013
Output
Output
Growth Rate
Total Pie
%
%
(Kg)
(Kg)
of Output (%)
Observations
1084
1514
/
Carps
417,411
50.8
869,268
39.8
108
Tilapia
124,967
15.2
310,490
14.2
148
Pangasius
195,608
23.8
873,521
40
347
Shrimp
60,733
7.4
86,656
4
43
Niche
7,905
1
24,457
1.1
209
Other
15,474
1.9
18,896
0.9
22
Total
822,098
100
2183286
100
166
Note 1: the list of fish species in each category can be found in the Table 3.A1.
Note 2: unbalanced sample since farms do not produce fishes in 2008 are
excluded

134

Table 3.2. Distribution of size categories of different types of actors in the study areas
Size
Actor
Defining Characteristic
Definition
Category
Small
Less than 0.1 ha
Hatcheries
Total production area in hectares
Medium 0.1 to 2.0 ha
Large
More than 2 ha
Feed Mills

Total metric tons of feed produced per
month

Small
Medium
Large

Less than 50 MT
50 to 300 MT
More than 300 MT

Input Traders

Total metric tons of feed sold per
month

Small
Medium
Large

Less than 10 MT
10 to 100 MT
More than 100 MT

Farmers

Total pond area in hectares

Small
Medium
Large

Less than 0.5 ha
0.5 to 1.0 ha
More than 1 ha

Traders

Total metric tons of fish traded per
week

Small
Medium
Large

Less than 1 MT
1 to 5 MT
More than 5 MT

135

Table 3.3. Clustering in the 20 districts in four cardinal-points zones
Change
Change
Change rate Change rate
District
Zone
2003 2008 2013
03-13
(1308-13
(1303)/03
08)/08
Mymensingh
0.236
North
0.528 0.809 1.000 0.472 0.894 0.191
North
Natore
0.574
0.078 0.122 0.192 0.114 1.462 0.070
North
Bogra
0.623
0.057 0.106 0.173 0.116 2.035 0.066
North
Gazipur
0.078
0.080 0.115 0.124 0.044 0.550 0.009
North
Dinajpur
0.652
0.014 0.023 0.038 0.024 1.714 0.015
North
Narsingdi
0.226
0.026 0.031 0.038 0.012 0.462 0.007
Khulna
Bagerhat
Jessore
Satkhira

Southwest
Southwest
Southwest
Southwest

0.421
0.410
0.255
0.200

0.542
0.375
0.326
0.280

0.683
0.448
0.396
0.372

0.263
0.038
0.141
0.172

0.625
0.093
0.553
0.860

0.142
0.074
0.071
0.092

0.262
0.197
0.218
0.329

Gopalganj
Bhola
Barisal
Chandpur

Southcenter
Southcenter
Southcenter
Southcenter

0.413
0.114
0.080
0.000

0.484
0.263
0.111
0.005

0.646
0.560
0.158
0.011

0.233
0.446
0.078
0.011

0.564
3.912
0.975
-

0.162
0.297
0.047
0.006

0.335
1.129
0.423
1.200

Brahmanbaria
0.311
East
0.146 0.193 0.253 0.107 0.733 0.060
East
Noakhali
0.432
0.088 0.118 0.169 0.081 0.920 0.051
East
Sylhet
0.438
0.032 0.073 0.105 0.073 2.281 0.032
East
Cox's bazar
0.379
0.058 0.066 0.090 0.032 0.552 0.025
East
Chittagong
0.246
0.059 0.069 0.086 0.027 0.458 0.017
East
Comilla
0.560
0.032 0.050 0.078 0.046 1.438 0.028
Note: The clustering measure in 2008 is generated by principle component analysis on
the six variables explained in the text (Section 4.1). The measures in 2003 and 2013 are
calculated based on the parameters generated from the principle component analysis
based on data in 2008.

136

Table 3.4. Descriptive statistics of household and district characteristics
HH Level

Observations
Clustering Index (CI)

2008

2013

1084
0.27

1514
0.35

Individual & Household Variables
HH Size
4.9
4.9
HHH Male (%)
92.7
92.5
HHH Education. Level
1.4
1.5
(1-5 levels)
% of HH doing any
42.4
54.7
nonfarm activities
Annual income from
nonfarm activities
32.2
42.3
(1,000 Taka/HH)
Hours to nearest large
city (either Dhaka,
2.7
2.7
Chittagong, or Khulna)
HH's fish farming characteristics
Operated crop land area
0.83
0.98
(Acre)
Operated pond area
1.23
1.26
(Acre)
Ownership (% of HH
owned some ponds
65.1
61.5
alone by themselves)
Ownership (% of HH
owned some ponds by
93.8
89.2
themselves or jointly
with other households)
Average years of ponds 19.2
15.8
District Variable
Total length of paved
roads per district (km)
(four-year lag)

By 2008's CI [median] and Year
2008
2013
TLow High
Low High
Test
527
557
752
762
0.08
0.44 ***
0.12 0.58

TTest

5.1
89.6

4.7
95.7

1.5

1.4

***

36.6

47.9

***

34.1

30.4

3.6

1.8

0.89

All HH

271

343

TTest
***

5.1
89.5

4.7
95.5

***
***

1.5

1.4

**

48.7

60.6

***

43.7

40.9

3.5

1.9

***

0.78

1.08

0.89

**

1.33

1.13

1.31

1.20

*

54.5

75.2

***

52.0

70.9

***

***

97.3

90.5

***

92.6

86.0

***

***

19.6

18.7

16.2

15.5

***

318

226

390

296

***

***
***

***

***

***

***

Note: unbalanced sample since farms that do not produce fish in 2008 are excluded
*** p<0.01, ** p<0.05, * p<0.1

137

***

Table 3.5. Patterns in fish farmersâ adoption of technology innovations
All HH

HH Level
2008(1)

2013

TTest

Low

By 2013's CI [median] and Year
2008
2013
THigh
Low
High
Test
557
752
762

Observations
1084
1514
527
I. Inputs
Modern Inputs [include commercial feeds, fuel, fertilizer, pesticide, lime, vitamins, probiotic]
% of HH used any
77.7
84.5
66.6
88.2 ***
74.9
94.0
***
modern inputs
% of HH used modern
68.2
76.9
57.3
78.5 ***
66.8
87.0
***
inputs, excluding
commercial feeds
% of HH used any
64.0
73.8
54.3
73.2 ***
64.2
83.2
***
commercial feeds
(purchased any feeds)
Average expenses on
all modern inputs per
acre per HH
(Taka/Acre)
Average expenses on
commercial feeds per
acre per HH
(Taka/Acre)

6,349

12,182

***

5,805

6,863

10,819

13,527

4,784

9,422

***

4,580

4,977

8,650

10,184

Share (%) of expenses on âŚ over total expenses during fish production per HH (3)
% on modern inputs,
with family labor cost
15.0
17.5
***
11.9
18
***
14.4
imputed
% on modern inputs,
without family labor
28.1
32.6
***
20.7
35.1 ***
25.9
cost imputed
% on commercial
feeds, with family
7.6
9.4
***
7.5
7.8
9.4
labor cost imputed
% on commercial
feeds, without family
14.2
17.2
***
12.7
15.5
**
16.3
labor cost imputed
% on fingerling, with
family labor cost
imputed
36.3
36.0
41.9
31.1 ***
42.0
% on fingerling,
without family labor
cost imputed
60.4
59.0
69.2
52.1 ***
67.4
% on medicines, with
family labor cost
imputed
0.2
0.5
*
0.3
0.2
0.7

138

TTest

***
***
***

20.5

***

39.3

***

9.5
18.2

30.0

***

50.6

***

0.3

Table 3.5. (contâd)
% on medicines,
without family labor
cost imputed
% on hired labor, with
family labor cost
imputed
% on hired labor,
without family labor
cost imputed
% on family labor,
with family labor cost
imputed

0.3

0.7

*

0.4

0.2

0.8

0.5

3.3

4.3

**

1.9

4.6

***

3.1

5.4

***

5.1

6.2

*

3.0

7.0

***

4.4

8.0

***

40.3

41.0

39.0

41.5

39.1

42.9

**

4.2
4.4
***
4.2
4.3
Number of Species
(4)
Share (%) of fish farms, by species
96.2
95.2
98.5
94.1
Carps
61.9
61.4
71.5
52.8
Tilapia
7.7
8.9
10.1
5.6
Pangasius
28.0
25.7
2.3
52.4
Shrimp
4.4
4.5
5.1
3.8
Niche
11.2
10.7
2.8
19.0
Other
(3)
Share (%) of fish volume over total output per HH, by species
70.0
67.4
**
76.9
63.5
Carps
16.3
17.6
*
19.6
13.1
Tilapia
2.7
4.2
**
2.3
3.1
Pangasius
8.8
8.5
0.5
16.7
Shrimp
0.4
0.7
0.3
0.4
Niche
1.8
1.7
0.4
3.1
Other

4.4

4.4

97.6
73.0
10.8
1.9
4.7
2.8

92.8
49.9
7.1
49.2
4.3
18.5

***
***
**
***

72.8
22.5
3.2
0.4
0.6
0.5

62.1
12.7
5.1
16.4
0.7
2.9

***
***
**
***

II. Species
Specialization

***
***
***
***
***
***
***
***
***

Note (1): unbalanced sample since farms do not producing fish in 2008 are excluded
Note (2): the list of fish species in each category can be found in the Appendix
Note (3): Total expenses with family labor imputed = expenses on modern input +
fingerling + medicine + hired labor + imputed family labor
Total expenses without family labor imputed = expenses on modern input + fingerling +
medicine + hired labor
Note (4): this share is an average share over households; it differs from the shares in
Table 3.1, which are shares of an aggregate âpieâ
*** p<0.01, ** p<0.05, * p<0.1

139

***

***

Table 3.6. Regression of clustering degree on adoption of modern inputs
Adoption of Mod.
Adoption of Mod.
Adoption of
Inputs, w/ feeds
Inputs, w/o feeds
commercial feeds
(N=1,514)
Probit
Probit-IV Probit
Probit-IV Probit
Probit-IV
Clustering Degree
0.129***
0.322***
0.097**
0.226*
0.150***
0.435***
HH Level Controls
HH size
HHH male (dummy)
HHH education level
Nonfarm activities
Total area of crop lands
Total area of pond
Pond ownership
Pond years
Hours to nearest big city

(0.041)

(0.094)

(0.044)

(0.116)

(0.048)

(0.119)

0.002
(0.004)
0.038
(0.031)
-0.003
(0.007)
0.000
(0.018)
0.020***
(0.008)
0.001
(0.001)
0.030*
(0.018)
-0.003***
(0.001)
-0.015***
(0.005)

0.001
(0.004)
0.037
(0.030)
0.000
(0.007)
-0.003
(0.018)
0.020***
(0.008)
0.001
(0.001)
0.020
(0.018)
-0.002***
(0.001)
-0.006
(0.007)

0.003
(0.005)
0.076**
(0.037)
0.007
(0.008)
-0.002
(0.021)
0.020**
(0.008)
0.002
(0.002)
0.018
(0.022)
-0.003***
(0.001)
-0.023***
(0.007)

0.002
(0.005)
0.075**
(0.037)
0.009
(0.008)
-0.003
(0.021)
0.020**
(0.008)
0.002
(0.002)
0.011
(0.022)
-0.003***
(0.001)
-0.017*
(0.009)

0.003
(0.005)
0.064*
(0.038)
-0.007
(0.008)
0.007
(0.021)
0.038***
(0.010)
0.001
(0.002)
0.035
(0.022)
-0.005***
(0.001)
-0.045***
(0.007)

0.001
(0.005)
0.061
(0.038)
-0.002
(0.008)
0.004
(0.021)
0.039***
(0.010)
0.002
(0.002)
0.020
(0.022)
-0.005***
(0.001)
-0.031***
(0.009)

-0.015*
(0.009)

-0.010
(0.012)

-0.014
(0.012)

-0.010
(0.012)

-0.019
(0.012)

0.002
(0.032)
0.184***
(0.049)
-0.071*
(0.041)
-0.226**
(0.098)
0.192

-0.031
(0.039)
0.076*
(0.043)
-0.211***
(0.032)

-0.025
(0.040)
0.083*
(0.044)
-0.163***
(0.052)
-0.148
(0.123)
0.134

-0.146***
(0.040)
0.079*
(0.046)
-0.245***
(0.034)

-0.136***
(0.041)
0.090**
(0.046)
-0.142***
(0.054)
-0.331***
(0.125)
0.169

Other District Level Controls
Total paved roads(lag)
-0.009
(0.009)

Zone Fixed Effects
South West Zone
South Center Zone
East Zone
Residual

-0.006
(0.032)
0.177***
(0.050)
-0.141***
(0.025)

/
/

/
/

/
/

0.189
0.133
Pseudo R2
0.165
Note: Marginal effects are reported. Heteroskedasticity robust standard errors are in
parentheses. *** p<0.01, ** p<0.05, * p<0.1. Control function estimation is used of
Probit-IV columns.

140

Table 3.7. Regression of clustering degree on per acre expenditures on modern inputs
Exp. on Mod. Inputs,
Exp. on Mod. Inputs,
Exp. on commercial
w/ feeds per acre
w/o feeds per acre
feeds per acre
(N=1,514)
OLS
OLS-IV
OLS
OLS-IV
OLS
OLS-IV
Clustering Degree 1.846***
1.381
1.601**
-0.377
1.854***
3.504**
HH Level
Controls
HH size
HHH male
HHH education
level
Nonfarm activities
Area of crop lands
Area of ponds
Pond ownership
Pond years

(0.539)

(1.204)

(0.635)

(1.377)

(0.651)

(1.427)

-0.012
(0.080)
1.321*
(0.743)

-0.009
(0.080)
1.325*
(0.744)

0.021
(0.085)
1.744**
(0.787)

0.034
(0.085)
1.761**
(0.785)

-0.021
(0.091)
1.530**
(0.763)

-0.032
(0.091)
1.516**
(0.761)

-0.055
(0.114)
0.009
(0.317)
0.318***
(0.093)
-0.006
(0.027)
0.757**
(0.346)
-0.068***
(0.015)
-0.474***
(0.122)

-0.063
(0.115)
0.012
(0.317)
0.316***
(0.093)
-0.007
(0.027)
0.778**
(0.346)
-0.069***
(0.015)
-0.498***
(0.137)

0.080
(0.124)
0.016
(0.351)
0.312***
(0.104)
0.014
(0.021)
0.301
(0.380)
-0.065***
(0.016)
-0.411***
(0.129)

0.046
(0.125)
0.031
(0.351)
0.307***
(0.104)
0.012
(0.021)
0.390
(0.381)
-0.068***
(0.016)
-0.509***
(0.144)

-0.158
(0.133)
0.128
(0.368)
0.551***
(0.117)
-0.000
(0.032)
0.937**
(0.386)
-0.101***
(0.017)
-0.945***
(0.131)

-0.130
(0.135)
0.115
(0.368)
0.555***
(0.117)
0.001
(0.032)
0.863**
(0.384)
-0.098***
(0.017)
-0.863***
(0.144)

(0.880)

-2.379***
(0.895)

-2.085**
(0.926)

-1.845*
(0.942)

-1.272
(0.996)

-1.472
(1.015)

-1.998***
(0.667)
-0.247
(0.615)
-4.562***
(0.771)
0.147
(0.338)
10.133***

-1.111
(0.702)
0.182
(0.684)
-4.583***
(0.645)

-1.210*
(0.706)
0.031
(0.699)
-5.331***
(0.811)
0.627
(0.399)
7.447***

-4.002***
(0.708)
-0.450
(0.712)
-5.403***
(0.655)

-3.920***
(0.713)
-0.324
(0.724)
-4.778***
(0.848)
-0.523
(0.395)
7.727***

Hours to nearest
big city
Other District Level Controls
Total paved roads -2.435***
Zone Fixed
Effects
South West Zone
South Center Zone
East Zone
Residual

-1.975***
(0.660)
-0.211
(0.601)
-4.386***
(0.619)

/
/

/
/

/
/

Adjusted R2
9.928***
6.574***
8.457***
Note: Marginal effects are reported. Heteroskedasticity robust standard errors are in
parentheses. *** p<0.01, ** p<0.05, * p<0.1

141

Table 3.8. Regression of clustering degree on output share, by species
% of Niche fish
% of Carp output
% of Til.+Pang. output output
(N=1,514)
OLS
OLS-IV
OLS
OLS-IV
OLS
OLS-IV
Clustering Degree -0.058
-0.342*** 0.048
0.274***
0.015*
-0.001
(0.036)

(0.078)

(0.035)

(0.071)

(0.009)

(0.019)

(0.003)
0.016
(0.029)

-0.004
(0.003)
0.019
(0.029)

0.002
(0.003)
-0.047*
(0.026)

0.001
(0.003)
-0.049*
(0.026)

-0.000
(0.001)
-0.005
(0.005)

-0.000
(0.001)
-0.005
(0.005)

-0.005
(0.005)
0.019
(0.015)
0.024***
(0.005)
-0.006**
(0.003)
-0.030**
(0.015)
0.001*
(0.001)
0.027***
(0.005)

-0.009*
(0.005)
0.021
(0.015)
0.023***
(0.005)
-0.006**
(0.003)
-0.017
(0.015)
0.001
(0.001)
0.012**
(0.005)

0.005
(0.005)
-0.012
(0.013)
-0.015***
(0.004)
0.002
(0.001)
0.023*
(0.013)
-0.001***
(0.001)
-0.027***
(0.004)

0.009**
(0.005)
-0.013
(0.013)
-0.015***
(0.004)
0.002
(0.001)
0.013
(0.013)
-0.001*
(0.001)
-0.016***
(0.005)

0.001
(0.001)
0.001
(0.003)
0.000
(0.001)
-0.000**
(0.000)
-0.001
(0.002)
0.000**
(0.000)
0.000
(0.000)

0.001
(0.001)
0.001
(0.003)
0.000
(0.001)
-0.000**
(0.000)
-0.001
(0.002)
0.000**
(0.000)
-0.001
(0.001)

(0.009)

-0.021**
(0.008)

-0.027***
(0.009)

0.001
(0.001)

0.001
(0.001)

-0.214***
(0.026)
-0.062**
(0.028)
0.048**
(0.022)

-0.200***
(0.025)
-0.042
(0.027)
0.135***
(0.028)
-0.256***
(0.067)
0.355***
(0.056)
0.123

-0.009**
(0.004)
-0.006*
(0.004)
-0.005*
(0.003)

-0.010**
(0.004)
-0.007**
(0.004)
-0.011*
(0.006)
0.018
(0.014)
0.009
(0.011)
0.014

HH Level Controls
HH size
-0.005
HHH male
HHH education
level
Nonfarm activities
Area of crop lands
Area of ponds
Pond ownership
Pond years

Hours to nearest
big city
Other District Level Controls
Total paved roads 0.035*** 0.044***
(0.009)

Zone Fixed Effects
South West Zone
-0.023
South Center Zone
East Zone

(0.028)
0.081***
(0.029)
-0.054**
(0.023)

Residual

/
/

Constant

0.503***
(0.056)
0.117

-0.041
(0.027)
0.056**
(0.028)
-0.163***
(0.031)
0.321***
(0.074)
0.643***
(0.062)
0.123

/
/
0.466***
(0.052)
0.118

Adjusted R2
Note: Heteroskedasticity robust standard errors are in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

142

/
/
0.001
(0.008)
0.014

Table 3.9. Regression of clustering degree and urban proximity on modern inputs
adoption, by product-type farmers
Probit
Probit-IV
M1
M2
M3
M1
M2
M3
Carp Producers (N=1,441)
0.130***
0.093**
0.153***
0.314***
0.205*
0.417***
Clustering
(0.042)
(0.045)
(0.048)
(0.096)
(0.117)
(0.118)
Degree
-0.008
-0.019** -0.031***
Hours to nearest -0.016*** -0.025*** -0.044***
(0.005)
(0.007)
(0.007)
(0.007)
(0.009)
(0.009)
big city
/
/
/
-0.214**
-0.129
-0.308**
Residual
/

/

Tilapia & Pangas Producers (N=962)
0.184*** 0.149**
Clustering
(0.070)
(0.060)
Degree
Hours to nearest -0.015** -0.020**
(0.007)
(0.008)
big city
/
/
Residual
/

/

/

0.176**
(0.068)
-0.048***
(0.009)
/
/

(0.098)

0.198
(0.169)
-0.014
(0.011)
-0.016
(0.164)

(0.124)

0.212
(0.169)
-0.016
(0.012)
-0.070
(0.175)

(0.123)

0.299
(0.193)
-0.041***
(0.014)
-0.140
(0.198)

Note: M1 refers to all modern inputs; M2 refers to modern inputs outside of commercial
feed; M3 includes commercial feeds only. Heteroskedasticity robust standard errors are in
parentheses.
*** p<0.01, ** p<0.05, * p<0.1

143

Table 3.10. Regression of clustering degree on specialization
(N=1,514)
OLS
OLS-IV
Clustering Degree
-0.982***
-2.828***
HH Level Controls
HH size
HHH male (dummy)
HHH education level
Nonfarm activities (dummy)
Total area of crop lands (Acre)
Total area of ponds (Acre)
Pond ownership (owned=1)
Pond years
Hours to nearest big city
Other District Level Controls
Total paved roads (km) (lag)
Zone Fixed Effects
South West Zone (dummy)
South Center Zone (dummy)
East Zone (dummy)
Constant
Residual

(0.157)

(0.541)

0.060***
(0.021)
0.320**
(0.143)
0.044
(0.032)
0.095
(0.084)
0.064**
(0.032)
0.023***
(0.008)
-0.039
(0.083)
0.011***
(0.004)
-0.071***
(0.026)

0.073***
(0.021)
0.335**
(0.144)
0.013
(0.033)
0.109
(0.083)
0.060*
(0.031)
0.021**
(0.008)
0.043
(0.085)
0.008**
(0.004)
-0.164***
(0.036)

-0.022
(0.061)

0.034
(0.063)

0.005
(0.166)
-0.260
(0.170)
-0.709***
(0.146)
4.302***
(0.355)

-0.107
(0.170)
-0.422**
(0.177)
-1.418***
(0.240)
5.207***
(0.445)
2.087***
(0.589)
0.082

/
/

Adjusted R2
0.072
Note: Heteroskedasticity robust standard errors are in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

144

Table 3.11. Regression of clustering degree and urban proximity, by different samples
Full sample (N=1,514)
Sub-sample (N=430)
Observations
without IV
with IV
without IV
with IV
(A) Adoption of commercial feeds
0.150***
0.435***
0.137**
0.229**
Clustering Degree
Hours to nearest big city

(0.048)
-0.045***
(0.007)

Residual

/
/

Clustering Degree

0.153***
(0.048)
-0.044***
(0.007)
/
/

Hours to nearest big city
Residual
Clustering Degree
Hours to nearest big city
Residual
Clustering Degree
Hours to nearest big city

0.417***
(0.118)
-0.031***
(0.009)
-0.308**
(0.123)

0.184***
(0.067)
-0.020**
(0.010)
/
/

(0.193)
-0.041***
(0.014)
-0.140
(0.198)

(0.090)
-0.034**
(0.015)

(D) Output Share of Carps (%)

-0.058
(0.036)
0.027***
(0.005)

Clustering Degree

0.048
(0.035)
-0.027***
(0.004)

0.275***
(0.103)
-0.017
(0.011)
-0.137
(0.118)

-0.342***
(0.078)
0.012**
(0.005)
0.321***
(0.074)

-0.158**
(0.064)
0.032***
(0.009)
/

0.274***
(0.071)
-0.016***
(0.005)
-0.256***
(0.067)

0.116*
(0.064)
-0.029***
(0.009)

-2.828***
(0.541)
-0.164***
(0.036)
2.087***
(0.589)

-0.987***
(0.276)
-0.092*
(0.050)

(0.150)
-0.030*
(0.016)
-0.245
(0.177)

-0.374***
(0.102)
0.022**
(0.010)
0.315***
(0.112)

/
(E) Output Share of Tilapia+ Pangas (%)

Residual

/
/

Clustering Degree

-0.982***
(0.157)
-0.071***
(0.026)

Residual

(0.107)
-0.020*
(0.012)
-0.139
(0.130)

(C) Adoption of commercial feeds, Tilapia+Pangas
0.299 producers
0.186**
0.346**

0.176**
(0.068)
-0.048***
(0.009)
/
/

/
/

Hours to nearest big city

(0.067)
-0.023**
(0.011)
/
/

(B) Adoption of commercial feeds, Carp producers

Residual

Hours to nearest big city

(0.119)
-0.031***
(0.009)
-0.331***
(0.125)

/
/
(F) Number of Species per HH

Note: Heteroskedasticity robust standard errors are in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

145

0.308***
(0.089)
-0.019**
(0.009)
-0.279***
(0.090)
-0.880*
(0.489)
-0.087*
(0.052)
-0.156
(0.611)

Figure 3.1. 20 sample districts in six major aquaculture areas of Bangladesh

146

Fish Farmers

Input Dealers

Feed Mills

Hatcheries

Fish Traders

Figure 3.2. Distribution of actors per 1,000 rural people in 2013 by district
Note 1: the scale is not provided due to the space limitation. In each figure, the darker the
blue, the higher the number of actors per 1000 rural people in the district; Note 2: the
concentrated areas of each actor are circled in yellow.

147

Figure 3.3. Clustering degrees of Sample Districts, 2013
Note: We have 20 sample districts and each district is regarded as a cluster. Only
20 sample districts are colored in green. The other districts of Bangladesh are not
colored. The darker the green, the greater is the degree of clustering.

148

Carp
silver carp
grass carp
mirror carp
common carp
black carp
karfu
rui
katla
mrigel
kalibaus
pona
puti/ swarputi
bata
blad cap
blad carp
gunia
big head
hangry
hanri

Table 3.A1. Fishes that mapped into each specie category
Other
Niche
Pangas Tilapia
Shrimp
faishiya
chitol
pangas tilapia
bagda chingri
khalse
koi
nailotica golda chingri
shol/gajar/taki magur
chaka chingri
tangra/baim
shingi
harina chingri
vetki
rup chanda
deshi cingri
mola/dela
rup chnda
desi chinri
ayri
gura chingri
bowal
foli
fosholla
koral mas
mola dhela
sai plane
tablet
tampo
tblet

149

Table 3.A2. IV regression of clustering degree on specialization, by IV methods
Dependent Var.=number of fishes
(N=1,514)
Instrument Variables
Crop input volumes per capita (ton)
Crop output volumes per capita (ton)
Clustering Degree
HH Level Controls
HH size
HHH male (dummy)
HHH education level
HH doing nonfarm activities (dummy)
Total areas of crop lands (Acre)
Total areas of pond (Acre)
Pond ownership (owned=1)
Pond years
Hour to nearest big city
Other District Level Controls
Total paved roads of district (km) (lag)
Zone Fixed Effects
In South West Zone (dummy)
In South Center Zone (dummy)
In East Zone (dummy)
Constant

(1)
First-Stage

(2)
IV-2SLS

(3)
IV-LIML

(4)
IV-GMM

-1.991***
(0.306)
-0.371***
(0.063)
/
/

/
/
/
/
-2.828***
(0.532)

/
/
/
/
-2.889***
(0.548)

/
/
/
/
-2.994***
(0.529)

0.003
(0.003)
0.025
(0.023)
-0.017***
(0.004)
0.006
(0.012)
-0.001
(0.004)
-0.001
(0.001)
0.029**
(0.012)
-0.001**
(0.001)
-0.034***
(0.004)

0.073***
(0.022)
0.335**
(0.151)
0.013
(0.034)
0.109
(0.085)
0.060*
(0.032)
0.021***
(0.008)
0.043
(0.087)
0.008**
(0.004)
-0.164***
(0.037)

0.073***
(0.022)
0.336**
(0.152)
0.012
(0.034)
0.110
(0.086)
0.060*
(0.032)
0.021***
(0.008)
0.046
(0.088)
0.007**
(0.004)
-0.167***
(0.038)

0.077***
(0.022)
0.321**
(0.153)
0.007
(0.034)
0.112
(0.086)
0.052
(0.032)
0.021***
(0.008)
0.054
(0.088)
0.007*
(0.004)
-0.178***
(0.037)

0.009
(0.008)

0.034
(0.064)

0.036
(0.064)

0.062
(0.063)

-0.053**
(0.025)
-0.222***
(0.025)
-0.450***
(0.019)
0.697***
(0.050)

-0.107
(0.181)
-0.422**
(0.188)
-1.418***
(0.237)
5.207***
(0.450)

-0.111
(0.182)
-0.427**
(0.189)
-1.441***
(0.243)
5.237***
(0.455)

-0.141
(0.183)
-0.429**
(0.190)
-1.533***
(0.233)
5.251***
(0.453)

NA

0.000

0.018

0.017

Hausman Test (H0: ci is exogenous)
P-Value
/
0.000
Overidentification Test
P-Value
/
0.017
Note: Heteroskedasticity robust standard errors are in parentheses.

150

REFERENCES

151

REFERENCES
Ahmed, N., 2009. Revolution in small-scale freshwater rural aquaculture in Mymensingh,
Bangladesh. World Aquaculture, 40 (4), 31.
Ahmed, N., Toufique, K.A., 2015. Greening the blue revolution of smallâscale freshwater
aquaculture in Mymensingh, Bangladesh. Aquaculture Research, 46 (10), 2305-2322.
Ali, H., Haque, M.M., Belton, B., 2013. Striped catfish (Pangasianodon hypophthalmus,
Sauvage, 1878) aquaculture in Bangladesh: an overview. Aquaculture Research, 44 (6),
950-965.
Bandiera, O., Rasul, I., 2006. Social networks and technology adoption in northern
Mozambique. The Economic Journal, 116 (514), 869-902.
Belton, B., Karim, M., Thilsted, S., Collis, W., Phillips, M., 2011. Review of aquaculture
and fish consumption in Bangladesh. Studies and Reviews 2011-53. Penang: The
WorldFish Center, November 2011.
Conley, T.G., Udry, C.R., 2010. Learning about a new technology: Pineapple in
Ghana. The American Economic Review, 100 (1), 25-69.
Delgado, M., Porter, M.E., Stern, S., 2015. Defining clusters of related industries. Journal
of Economic Geography, 16 (1), 1-38.
Fafchamps, M., Shilpi, F., 2013. Determinants of the choice of migration
destination. Oxford Bulletin of Economics and Statistics, 75 (3), 388-409.
Feder, G., Just, R.E., Zilberman, D., 1985. Adoption of agricultural innovations in
developing countries: A survey. Economic Development and Cultural Change, 33 (2),
255-298.
Forman, C., Goldfarb, A., Greenstein, S., 2005. Technology adoption in and out of major
urban areas: When do internal firm resources matter most? NBER Working Paper No.
11642: Cambridge, Massachusetts.
GhezĂĄn, G., Mateos, M., Viteri, L. 2002. Impact of supermarkets and fast-food chains on
horticultural supply chains in Argentina. Development Policy Review 20 (4), 389-408.
Haggblade, S., P.B.R. Hazell, and T. Reardon. 2010. The Rural Nonfarm Economy: Prospects
for Growth and Poverty Reduction, World Development 38(10), 1429â1441.

Henderson, J.V., 1974. The sizes and types of cities. The American Economic Review,
64 (4), 640-656.

152

Hernandez, R., Belton, B., Reardon, T., Hu, C., Zhang, X. and Ahmed, A., 2017. The
âquiet revolutionâ in the aquaculture value chain in Bangladesh. Aquaculture.
http://dx.doi.org/10.1016/j.aquaculture.2017.06.006
Humphrey, J. and Schmitz, H., 2000. Governance and upgrading: linking industrial
cluster and global value chain research. IDS Working Paper No.120: Institute of
Development Studies (IDS), University of Sussex, Brighton.
Humphrey, J. and Schmitz, H., 2002. How does insertion in global value chains affect
upgrading in industrial clusters?. Regional Studies, 36 (9), 1017-1027.
Hussain, M.G., 2009. A future for the tilapia in Bangladesh. Aquaculture Asia Pacific
Magazine, 5(4), 38â40.
Just, R.E., Schmitz, A. and Zilberman, D., 1979. Price controls and optimal export
policies under alternative market structures. The American Economic Review, 69
(4),706-714.
Kelley, M.R. and Helper, S., 1999. Firm size and capabilities, regional agglomeration,
and the adoption of new technology. Economics of Innovation and New Technology, 8
(1-2), 79-103.
Lerner, A.P., 1934. The Concept of Monopoly and the Measurement of Monopoly Power.
Review of Economic Studies, 1 (3), 157â175.
Liu, E.M., 2013. Time to change what to sow: risk preferences and technology adoption
decisions of cotton farmers in China. The Review of Economics and Statistics. 95 (4),
1386-1403.
Long, C. and Zhang, X., 2011. Cluster-based industrialization in China: Financing and
performance. Journal of International Economics, 84 (1), 112-123.
Long, C. and Zhang, X., 2012. Patterns of China's industrialization: Concentration,
specialization, and clustering. China Economic Review, 23 (3), 593-612.
Mamun-Ur-Rashid, M., Belton, B., Phillips, M. and Rosentrater, K.A., 2013. Improving
aquaculture feed in Bangladesh: From feed ingredients to farmer profit to safe
consumption. Working Paper: 2013-34. WorldFish, Penang, Malaysia.
Marshall, A., 1920. Principles of Economics. An Introductory Volume. Eighth Edition.
London: Macmillan.
Pietrobelli, C., and Rabelloti, R. 2006. âClusters and Value Chains in Latin America: In
Search of an Integrated Approach.â In Upgrading to Compete: Global Value Chains,
Clusters, and SMEs in Latin America, edited by Pietrobelli, C., and Rabelloti, R. 1â40.
Washington, DC: Inter-American Development Bank.

153

Porter, M.E., 1998. Clusters and the new economics of competition. Harvard Business
Review, 76 (6), 77â90
Porter, M.E., 2000. Location, competition, and economic development: Local clusters in
a global economy. Economic Development Quarterly, 14 (1), 15-34.
Reardon, T., Crawford, E. and Kelly, V., 1994. Links between nonfarm income and farm
investment in African households: adding the capital market perspective. American
Journal of Agricultural Economics, 76 (5), 1172-1176.
Reardon, T., Tschirley, D., Dolislager, M., Snyder, J., Hu, C. and White, S., 2014.
Urbanization, diet change, and transformation of food supply chains in Asia. East
Lansing, Michigan: Global Center for Food Systems Innovation. Michigan State
University.
Sandee, H. and Rietveld, P., 2001. Upgrading traditional technologies in small-scale
industry clusters: Collaboration and innovation adoption in Indonesia. Journal of
Development Studies, 37 (4), 150-172.
Schmitz, H., 1995. Small shoemakers and Fordist giants: tale of a supercluster. World
Development, 23 (1), 9-28.
Schmitz, H., 1999. Collective efficiency and increasing returns. Cambridge Journal of
Economics, 23 (4), 465-483.
Otsuka, K. and Sonobe, T., 2011. A cluster-based industrial development policy for lowincome countries. Policy Research Working Paper No. 5703: World Bank, Washington,
DC.
Uchida, H., A. Nelson. 2009. Agglomeration Index: Towards a new measure of urban
concentration. Washington, D.C.: World Bank, Washington, DC.
Visser, E. J., 1996. Local sources of competitiveness. Spatial clustering and
organisational dynamics in small-scale clothing in Lima, Peru. Tinbergen Institute Ph.D
Thesis. University of Amsterdam, Amsterdam.
World Bank. 2017. Enabling the Business of Agriculture 2017. Washington DC, The
World Bank.
Zhang, X. and Hu, D., 2014. Overcoming successive bottlenecks: The evolution of a
potato cluster in China. World Development, 63 (2014), 102-112.
Zilberman, D., Lu, L. and Reardon, T., 2017. Innovation-induced food supply chain
design. Food Policy. https://doi.org/10.1016/j.foodpol.2017.03.010

154