-
Full text
-
i
DEVELOPMENT IN SUB-SAHARAN AFRICA
NEW MICRO-LEVEL EVIDENCE ON EDUCATION,
GEOGRAPHY AND TRADE
Dissertation
zur Erlangung des Grades eines Doktors der Wirtschaftswissenschaften
der Rechts- und Wirtschaftswissenschaftlichen Fakultät
der Universität Bayreuth
Vorgelegt von
Frederik Wild aus
BAB 8 Stuttgart-München bei km 10.7
ii
Dekan:
Prof. Dr. André Mayer
Erstberichterstatter:
Prof. Dr. David Stadelmann
Zweitberichterstatter:
Prof. Dr. Mario Larch
Tag der mündlichen Prüfung:
18.04.2024
iii
Für meine Familie
iii
Acknowledgments
First and foremost, I am deeply indebted to my supervisor, David Stadelmann, for affording
me the opportunity to pursue this endeavor. Without your trust and support, I could not
imagine having developed the deep passion and satisfaction for academic work that I now enjoy.
You gave me guidance when needed, yet also allowed me the freedom to explore the topics I
am passionate about. I will remember the many insightful, often lively debates we had in the
office, on our many travels, or simply between doors. Your intellectual acumen is truly inspiring.
Secondly, I would like to thank Mario Larch, who took on the role of second supervisor
without hesitation and was always available for advice. I also won’t forget Mario’s summertime
seminars, which were academically fruitful and, above all, encouraging environments. Not to
mention the great barbecues! I am also very grateful to Benno Torgler for inviting me to QUT,
where I spent an amazing two months during my research visit.
I am appreciative of the many colleagues I have met during my time in Bayreuth and of
the faculty for providing a supportive and constructive environment. I would like to thank the
team of VWL II, and especially Heidi, for always being there with an open ear and for a
delightful morning talk. Not to forget Helge Wiechmann, with whom I have had the pleasure
of sharing the office and who became a true friend during our dissertation chapters.
Importantly, I would like to thank Julia, who, in comparison to all of the above, could
not escape the pleasure of listening to countless sermons on my projects. I don’t take your
move to Bayreuth for granted one bit. Thank you for your understanding, your limitless
kindness, and for centering my mind. I have learned to count on your unwavering support. I
also want to thank my best friend Dominik for always keeping me sane, be it with solid life
advice or with some cheerful banter.
Last and certainly not least, my greatest gratitude is directed to my parents. I am deeply
indebted to you for instilling in me the respect, devotion, and work ethic necessary for such an
endeavor, all while not having had the chance to pursue an academic path yourselves. Your
unshakeable support and your seemingly naïve trust in my doing is second to none and certainly
to the surprise of my two brothers, Philipp and Maximilian, who have had the pleasure of
watching the “Nesthäkchen” receive one or the other additional break throughout our lives. I
thank you for always keeping a watchful eye.
iv
Contents
1 Introduction
1
2 Coastal Proximity and Individual Living Standards: Econometric Evidence from
Geo-Referenced Household Surveys in Sub-Saharan Africa
5
2.1 Introduction ................................................................................................................ 6
2.2 Data ............................................................................................................................ 7
2.2.1. Dependent Variables and Channels of Influence ................................................. 8
2.2.2. Main Independent Variables............................................................................... 9
2.3 Empirical Strategy .................................................................................................... 13
2.4 Results ...................................................................................................................... 13
2.4.1. Robustness Checks and Extensions .................................................................. 15
2.4.2. Mechanisms Explaining the Relevance of Coastal Proximity ............................ 16
2.4.3. Bad Controls and Mediation Analysis .............................................................. 19
2.5 Conclusion ................................................................................................................. 23
2.6 References ................................................................................................................. 25
Appendix A
29
3 Heterogeneous Effects of Women’s Schooling on Fertility, Literacy and Work:
Evidence from Burundi’s Free Primary Schooling Policy
52
3.1 Introduction .............................................................................................................. 53
3.2 Institutional Background and Identification Strategy ................................................ 56
3.2.1. Institutional Background .................................................................................. 56
3.2.2. Identification Strategy ...................................................................................... 57
3.3 Data and Descriptive Statistics ................................................................................. 59
3.3.1. Dependent Variables ........................................................................................ 60
3.3.2. Independent Variables ...................................................................................... 62
3.4 Results ...................................................................................................................... 63
3.4.1. The Effect of FPE on Schooling: First-Stage RDD Evidence ............................ 63
3.4.2. The Effects of Schooling: Second-Stage Evidence ............................................. 66
3.4.3. Further Evidence on the Heterogeneity ............................................................ 71
3.5 Robustness Checks and Validity Tests ...................................................................... 73
3.6 Conclusion ................................................................................................................. 79
3.7 References ................................................................................................................. 81
Appendix B
87
4 Regional Market Integration and Household Welfare: Spatial Evidence from the
East African Community
114
4.1 Introduction ............................................................................................................ 115
4.2 Related Literature ................................................................................................... 117
4.2.1. Trade and Household Welfare ........................................................................ 117
4.2.2. Spatial Effects of Trade .................................................................................. 119
4.3 Institutional Background ......................................................................................... 123
v
4.4 A Four Region Economy ......................................................................................... 128
4.4.1. Consumer Preferences and Behavior ............................................................... 129
4.4.2. Transport Costs and Heterogeneous intra-national Space ............................... 132
4.4.3. Producer Behavior.......................................................................................... 136
4.4.4. Short-run Equilibrium .................................................................................... 137
4.4.5. Spatial Equilibria and Regional Trade Liberalization ..................................... 142
4.5 Empirical Strategy .................................................................................................. 151
4.6 Data ........................................................................................................................ 153
4.6.1. Dependent Variables ...................................................................................... 155
4.6.2. Independent Variables .................................................................................... 159
4.7 Results .................................................................................................................... 160
4.7.1. Main Results .................................................................................................. 160
4.7.2. Robustness Checks, Validity Tests and Extensions ........................................ 167
4.8 Conclusion ............................................................................................................... 182
4.9 References ............................................................................................................... 184
Appendix C
193
5 Concluding Remarks
303
1
Chapter 1
Introduction
Income differences across the world remain strikingly large, particularly along the North-South
divide. For instance, as a region, sub-Saharan Africa’s GDP per capita amounts to $3,767
compared to the European Union’s $45,917 (World Bank 2022).1 Importantly, these differences
in aggregate economic performance are associated with deprivation across many vital
dimensions of individual human development, such as health and education, along with broader
individual capabilities and freedoms (Sen 1985). As such, the research on the causes and
consequences of these persistent disparities remains significant. Particularly so in sub-Saharan
Africa, home to the world’s poorest countries.
This dissertation contributes to the understanding of development in sub-Saharan Africa
by presenting three independent and self-contained scientific articles that address distinct
developmental challenges of the continent, namely education, geography, and trade. More
specifically, they provide new micro-level evidence on these broadly discussed aspects of
development, which allows for a direct study of these factors in their contribution to individual
economic welfare. Thus, while the chapters address distinct research questions, they are aligned
in analyzing economic development from the individual- and household perspective.
Notes to the Reader. The individual chapters of this dissertation are structured as singular
scientific papers and can be read independently from one another and in any sequence preferred
by the reader. As such, there exists no general list of figures and tables nor a collated list of
references. Each of these items is provided in the respective chapters of the dissertation.
Modified versions of Chapter 2 and Chapter 3, on geography and education, have been
published in peer-reviewed Journals, namely the Review of Development Economics and the
Journal of African Economies, respectively. Chapter 4 is designed as a typical job market
paper, which presents an empirical as well as a theoretical treatment of regional market
integration in East Africa. As is common, this paper, in particular, was presented at numerous
national and international conferences and workshops. In what follows, I provide a brief non-
technical overview of each of the three chapters, individually.
1 GDP figures are expressed in constant 2017 international (PPP) USD.
2
Chapter 2: Coastal Proximity and Individual Living Standards:
Econometric Evidence from Geo-Referenced Household Surveys
in Sub-Saharan Africa
This chapter re-visits the relevance of physical geography in economic development. In
particular, we analyze coastal access as a distinct predictor of individual welfare, which has
been suggested as influential for comparative development in as early as the “Wealth of
Nations” (Smith 1776). Cross-country studies have confirmed these observations and find
coastal access as a robust indicator for income differences across countries (e.g. Bloom et al.
1998; Gallup et al. 1999). However, the inability to isolate such features from other country-
specific factors, such as institutions or culture, has prompted studies that analyze the link from
a regional perspective (Gennaioli et al. 2013; Henderson et al. 2018).
This chapter systematically extends this latter line of research by drawing evidence from
a granular large-scale, geo-referenced household-level survey covering 28 sub-Saharan African
countries over 20 years. Analyzing individual-level data allows us to test whether the insights
from cross-country and cross-regional contexts also apply at the individual level. Moreover, we
can utilize the comprehensiveness of our dataset to explore a large set of indicators and
potential channels of influence through which coastal access may matter for individual living
standards.
The following main results emerge from the analysis. First, coastal proximity is confirmed
as a relevant indicator of within country income disparities. Second, while coastal proximity
remains robust for visually all controls entered, there are distinct factors that mediate the
results. In particular, the inclusion of human capital, infrastructure, and urbanization seem to
account for a large part of the identified welfare gap. And lastly, our study thereby also
strengthens the notion that geography need not be “destiny”, as policy may cater towards
influencing these three factors.
Chapter 3: Heterogeneous Effects of Women’s Schooling on
Fertility, Literacy and Work: Evidence from Burundi’s Free
Primary Education Policy
Education has long been regarded as one of the primary instruments in fostering equitable
economic development, as prominently outlined in the Millenium Development Goals (United
Nations 2001). Particularly for women, schooling is widely forwarded as an effective means to
tackle challenges connected to gender equality, and particularly for reducing teenage
pregnancies and the negative consequences linked to them (Lloyd and Young, 2009; World
Bank, 2017).
3
However, several empirical issues complicate studying the education-fertility nexus in a
causal manner, as educational attainment is jointly influenced by factors such as family
background, wealth, or selection effects by skill. As such, recent research has employed quasi-
experimental methods to provide a more robust estimate that can be interpreted causally (e.g.
Keats 2018).
This chapter contributes to this recent and now growing literature by exploiting
Burundi’s Free Primary Education Policy of 2005 as a natural experiment. To identify causal
variation in education, we employ a Regression Discontinuity Design and analyze the outcomes
affected by schooling through an instrumental variable (IV) approach. Contrary to the extant
studies in this field, we are able to provide distinct and relevant evidence on the differential
treatment effects of education. While poor women profit in terms of increases in literacy,
remunerated employment opportunities as well as a reduction in teenage births, none of these
effects of additional education are observed for women from the wealthier households of our
sample. The evidence of such a marked heterogeneity helps to evaluate under which conditions
the literature’s findings may generalize.
Chapter 4: Regional Market Integration and Household Welfare:
Spatial Evidence from the East African Community
The role of trade in economic development has been widely studied and, from a country-level
perspective, usually points to positive effects on income and growth (e.g. Frankel and Romer
1999). However, donor agencies have emphasized the potentially inequality-enhancing impact
of trade within countries (World Bank 2009), and there is a growing body of literature dedicated
to examining these distributional effects (for an overview see Engel et al. 2021). One aspect
that has garnered particular attention is whether trade increases regional inequalities within
countries, i.e. across space (see for an overview Brülhart 2011). However, in Africa, the
literature on such concerns is scant and mainly relies on the use of lights emitted by night as
a proxy of economic development (e.g. Eberhard-Ruiz and Moradi 2019).
In this chapter, I provide novel evidence on the distributional effects of trade
liberalization in Africa by combining the spatial considerations of market integration with a
household-level analysis. I thereby treat the re-establishment of the East African Community
(EAC) in 2001 as a regional policy intervention having differential effects on individual
households governed by their geo-spatial location within the countries. I derive these predictions
from an extension of a canonical New Economic Geography (NEG) model which accommodates
the key features of the EAC’s spatial layout. Contrary to the model’s prediction, the reduced-
form empirical results do not show evidence of increased household welfare in regions closer to
the newly accessed markets. Rather, I observe strengthened agglomeration tendencies in the
pre-existing economic hubs of Nairobi, Dar es Salaam and Kampala.
4
The empirical results are in contrast to the ones found on similar integration processes
in other contexts and thereby provide relevant insights for policy makers. The findings may
also contribute to the understanding of the causes and consequences of regional disparities in
developing countries, particularly on the increasingly studied and peculiar urbanization
tendencies.
5
Chapter 2
Coastal Proximity and Individual Living
Standards: Econometric Evidence from Geo-
Referenced Household Surveys in Sub-
Saharan Africa1
Abstract. We investigate geo-referenced household-level data consisting of up to 128,609
individuals living in 11,261 localities across 17 coastal sub-Saharan African countries over 20 years.
We analyze the relevance of coastal proximity, measured by geographic distance to harbors, as a
predictor of individual economic living standards. Our setting allows us to account for country-time
fixed effects as well as individual-specific controls. Results reveal that individuals living further
away from the coast are significantly poorer measured along an array of welfare indicators. Our
findings are robust to the inclusion of other geographic covariates of development such as climate
(e.g. temperature, precipitation) or terrain conditions (e.g. ruggedness, land suitability). We also
explore mechanisms through which coastal proximity may matter for individual welfare and
decompose the estimated effect of coastal proximity via formal mediation analysis. Our results
highlight the role of human capital, urbanization as well as infrastructural endowments in explaining
within-country differences in individual economic welfare.
JEL Classification: O15, O18, R12, O55
Keywords: Geography, Coastal Proximity, Sub-Saharan Africa, Mediation Analysis
1 A modified version has been published under the identical title in the Review of Development Economics,
2022, 26 (4), 1883-1901. DOI: 10.1111/rode.12901 (joint work with David Stadelmann). The version presented
here systematically expands on the article to include additional evidence, robustness tests and discussions.
Acknowledgements. Funding from the DFG (project number 390713894) is gratefully acknowledged. We thank
two anonymous referees and the handling editor Andy McKay. We also thank the participants at the European
Public Choice Society Annual Meeting 2022 (Braga, Portugal) and at the Bayreuth Young Economics Research
Seminar 2022 (Bayreuth, Germany) for fruitful comments and discussions. All remaining errors are my own.
6
2.1 Introduction
“As by means of water-carriage a more
extensive market is opened to every sort of industry than
what land-carriage alone can afford it, so it is upon the sea-
coast, and along the banks of navigable rivers, that industry
of every kind naturally begins to subdivide and improve itself
[…]. There are in Africa none of those great inlets, such as the
Baltic and Adriatic seas in Europe, the Mediterranean […] to
carry maritime commerce into the interior parts of that great
continent: […]”
Adam Smith (1776)2
Cross-country studies investigating the link between physical geography and economic
development consistently provide evidence of a positive and statistically significant association
between coastal access and national income (e.g. Bloom et al. 1998; Radelet and Sachs 1998;
Gallup et al. 1999; Easterly and Levine 2003; Sala-I-Martin et al. 2004; UN-OHRLLS 2013).
More recent literature analyzing subnational variation in economic activity also suggests coastal
access as well as coastal proximity as relevant indicators of within-country differences in income
and related developmental outcomes (e.g. Rappaport and Sachs 2003; Gennaioli et al. 2013;
Motamed et al. 2014; Mitton 2016; Flückiger and Ludwig 2018; Henderson et al. 2018; Jetter
et al. 2019).
To systematically complement the literature that focused on outcomes at the national or
regional level, this paper analyzes the relevance of coastal proximity in predicting individual
economic welfare. We employ a repeated cross-sectional dataset from the Afrobarometer
spanning almost 20 years and consisting of up to 128,609 individuals living in 11,261 geo-
referenced localities across 17 coastal sub-Saharan African countries. Particularly in Africa,
countries and regions with coastal access have had higher levels of economic development
compared to more remote areas, which has been attributed to factors such as lower costs of
trade, the distribution of natural resources, as well as the amplifying forces of urbanization and
agglomeration (e.g. Bloom et al. 1998a; Radelet and Sachs 1998; Gallup et al. 1999; Limão and
Venables 2001; Atkin and Donaldson 2015; Storeygard 2016; Henderson et al. 2018). Spatial
inequalities such as these have been shown to persist even when the initial advantages of
(coastal) regions may have declined in relevance (Bleakley and Lin 2012; Jedwab et al. 2017).
Our results confirm coastal proximity as a robust indicator of individual economic welfare
across African countries: living further away from the coast is associated with a significant and
meaningful reduction in the likelihood of having cash employment (income), increases in the
occurrence of cash-, food-, water- and medicinal droughts (deprivation), as well as lower overall
household wealth (possessions). Our results are robust to the inclusion of relevant individual-
2 As cited in (Smith and Campbell 2009; 19,21).
7
level covariates, country-time specific influences via fixed-effects, as well as an extensive set of
further geographic variables related to development such as latitude, elevation, climatic factors
(e.g. temperature, precipitation) and features of the terrain (e.g. ruggedness, land suitability).
We also explore potential mechanisms on how coastal proximity may matter for
individual living standards and investigate several candidate factors shown to contribute to
spatial disparities in the literature (for an overview see Breinlich et al. 2014). In particular, we
analyze the relevance of human capital (e.g. Skoufias and Katayama 2011; Gennaioli et al.
2013; Flückiger and Ludwig 2018), urbanization-agglomeration (e.g. Young 2013; Motamed et
al. 2014; Chauvin et al. 2017; Gollin et al. 2017; Henderson et al. 2018), institutions (e.g.
Acemoglu et al. 2001; Nunn and Wantchekon 2011; Radeny and Bulte 2012; Michalopoulos and
Papaioannou 2014; Mitton 2016), infrastructure (e.g. Calderón and Servén 2010; Dinkelman
2011; Jedwab and Moradi 2016; Bluhm et al. 2018; Donaldson 2018; Jetter et al. 2019) as well
as market access and trade (e.g. Brülhart 2011; Bosker and Garretsen 2012; Hirte et al. 2020;
Jedwab and Storeygard 2020). We consider these factors in turn and assess their power in
mediating the relationship between coastal proximity and individual welfare through a formal
mediation analysis. The results highlight human capital, urbanization as well as infrastructural
endowments as the predominant channels via which the presented within-country differences
in individual economic welfare may be explained.
Our findings at the individual level emphasize the relevance of coastal proximity as an
indicator of economic development and lend further support to the previously discussed
interrelation between first- and second-nature causes of development (see Rodrik et al. 2004;
Breinlich et al. 2014; Lessmann and Seidel 2017).
The remainder of this paper is structured as follows: Section 2.2 presents the data, and
Section 2.3 the estimation strategy. Our results are given in Section 2.4, where we also present
the insights from the mediation analysis. Concluding remarks are offered in Section 2.5.
2.2 Data
We employ the complete set of the geo-referenced Afrobarometer survey rounds, spanning a
timeframe of 20 years (from 1999 to 2018) across seven survey waves (Afrobarometer 2019).3
Afrobarometer surveys are representative at the national level and respondents are adults of
the sampled households. They carry individual- and household-level information on basic
characteristics such as living conditions and household assets, and additionally, provide
information on individuals’ sentiments as well as opinions towards the economy, democracy,
governance and society. Afrobarometer fits geo-coordinates (latitude and longitude) to
respondents at the level of their respective enumeration area (EA) (BenYishay et al. 2017).
The sampling procedure aims for eight individuals/households per EA. Our main (extended)
3 Surveys were sampled in 1999-2001, 2002-2004, 2005-2006, 2008-2009, 2011-2013, 2014-2015 and 2016-2018,
respectively.
8
sample of countries consists of 128,609 (212,037) individuals living in 11,261 (17,319) geo-
referenced localities across 17 (28) coastal sub-Saharan African countries (see Figure 2.1). We
chose to restrict the main sample to coastal countries, so as to separate the distance effect from
a more general “landlockedness” effect which potentially confounds distance with other
influences such as administrative dependencies on transit countries (see Faye et al. 2004; UN-
OHRLLS 2013). We investigate the extended sample, including individuals living in landlocked
countries in our robustness tests.
Figure 2.1: Sample Coverage
2.2.1. Dependent Variables and Channels of Influence
We employ three main dependent variables as indicators of individual economic welfare: 1) The
dichotomous variable Cash Employment {0,1}, which indicates whether survey respondents
9
currently have part- or full-time cash employment, i.e. serving as a measure of individuals’
income. 2) The index How often gone without enough: (Water / Food / Cash Income / Medical
Care) [0,4] which serves as a measure for individual- as well as household deprivation and is
constructed by averaging individuals’ responses in these four categories. 4 3) The index
Possessions: (Radio / TV / Motor Vehicle) [0,1] which serves as a measure for survey
respondents’ wealth.5 To explore the potential channels through which coastal proximity may
matter for individual living standards, we make use of Afrobarometer’s opinion polling and
first investigate individuals’ sentiments regarding the most important issue of their respective
country, Most Important Issue: (Education/Institutions/Infrastructure) {0,1}.6 Thereafter, we
directly investigate the lived realities around these concerns via Education Level {0,1}, two
further composite indices, Institutions Score [1,4]7 and Infrastructure Present in Enumeration
Area: (Electricity Grid / Piped Water / Sewage / School / Paved Road / Health Clinic) [0,1]8,
as well as the dummy variable Urban {0,1}. We also analyze individuals’ opinions towards
supranational organizations aimed at increasing political as well as economic integration, Helps
your Country: (AU or ECOWAS/SADC/EAC/IGAD…) {0,3} to directly relate coastal
proximity with (regional) trade considerations.9 To further explore a potential trade channel,
we use information on individuals’ occupation and test for a differential effect of distance using
Commercial Farmer {0,1}, a dichotomous indicator for individuals working as farmers who
grow their produce mainly for sale.10
2.2.2. Main Independent Variables
To construct our main explanatory variable of interest, log(Distance to Harbor), we measure
the shortest geodesic (ellipsoidal) within-country distance from each respondent’s enumeration
4 The four different questions read: “Over the past year, how often, if ever, have you or anyone in your family
gone without: Enough clean water for home use” / “[…]: Enough food to eat” / “[…]: A cash income” / “[…]:
Medicines or medical treatment?”. Answers are: “Never”, “Just once or twice”, “Several times”, “Many times”,
“Always”, ranging from 0 through 4, respectively. These questions are consistently available in all
Afrobarometer survey rounds. Using each question separately does not affect our main insights as shown in
Table A.2 of the Appendix.
5 The questions read “Which of these things do you personally own? Radio” / “[…]? Television” / “[…]? Motor
Vehicle-Car-Motorcycle”. Wealth possessions were surveyed from Round 3 and onwards. Using each question
separately does not affect our main insights as shown Table A.2 of the Appendix.
6 The question reads: “In your opinion, what are the most important problems facing this country that
government should address?”.
7 To measure the quality of institutions, we construct an “Institutions Score” similar to Mitton (2016), which
is based on an array of questions regarding local- authorities, processes and government. The score is
constituted of 21 questions measuring individuals’ trust in (local) courts, police and government, their
experience with the procedures of local authorities, especially regarding bribery (corruption), the enforcement
of crime, and the ease of handling administrative matters. Higher values indicate fewer negative
experiences/better judgments of (local) institutions.
8 Nunn and Wantchekon (2011) use an identical measure für the provision of public goods, excluding roads.
9 Regional Economic Communities have the proclaimed aim to foster the movement of goods and people, and
to improve living standards. The question reads: “In your opinion, how much do each of the following do to
help your country, or haven’t you heard enough to say”.
10 We thank an anonymous referee for pointing out this additional extension.
10
area to the respective country’s major harbor(s).11 Similar to Rappaport and Sachs (2003), we
define all large and medium-sized ports listed in the World Port Index (WPI) as “major
harbors” (NGA 2019). We also employ alternative conceptions of coastal proximity for
robustness checks, namely shortest within-country distance to the coastline, log(Distance to
Coastline), distance to major harbors using beelines (as the crow flies), log(Beeline Distance to
Harbor), as well as distance to the coastline using beelines, log(Beeline Distance to Coast).12
Shapefile data for country administrative areas, the boundaries of which we use to calculate
within country distances – and also from which we construct the coastline – come from the
Center for Spatial Sciences at the University of California (GADM 2020).
Further Covariates. To isolate coastal proximity from other, potentially correlated,
geographic influences of development, we closely follow Henderson et al. (2018) and add an
extensive set of geographic covariates. We include Elevation (Farr et al. 2007), (Abs.) Latitude,
Ruggedness (Nunn and Puga 2012) and Malaria Ecology (Sachs et al. 2004) as well as
agricultural characteristics such as Land Suitability (Ramankutty et al. 2002), Growing Days
(FAO and IIASA 2019), Monthly Temperature and Monthly Rainfall (Fick and Hijmans 2017).
We also include seven dummy variables indicating the dominant natural vegetation of the area
according to Olson et al. (2001).13 We account for individuals’ access to rivers or lakes by
adding two dummy variables indicating whether individuals live within 25 kilometers of a
navigable river or major lake, i.e. Navigable River {0,1} and Major Lake {0,1}, and thereby
analyze an extended set of trade-related covariates together with our main explanatory variable,
log(Distance to Harbor) (see Henderson et al. 2018).14 We also add the individual-level
covariates Age, squared Age and a dichotomous indicator of gender, Female {0,1}. The
importance of urbanization-agglomeration aspects, argued to be particularly relevant in African
contexts (see Young 2013; Motamed, Florax, and Masters 2014; Chauvin et al. 2017; Jedwab,
Kerby, and Moradi 2017; Gollin, Kirchberger, and Lagakos 2017; Flückiger and Ludwig 2018;
Henderson et al. 2018), is encapsulated by three distinct indicators of urbanization: Primate
City {0,1}, a dummy indicating whether individuals live within 25 kilometers of a capital or
primate city, Population Density (CIESIN 2017), a continuous measure of population density
(per sq. kilometer), as well as Urban {0,1}, a dichotomous indicator included in the
Afrobarometer survey.
Descriptive statistics for all variables are presented in Table 2.1 panels a) and b).
11 We measure distances using the projection of coordinates along the earth’s ellipsoid (using WGS 84, EPSG
7030). We add +1 (kilometer) to our distance measure prior to taking the logarithm.
12 Beeline distances disregard country boundaries, i.e. cross country borders for shorter distances.
13 Following Henderson et al. (2018) for the definition of those dummies leaves us with seven indicators relevant
to our sample: Mediterranean {0,1}, Desert {0,1}, Mangroves {0,1}, Tropical Forest {0,1}, Tropical Grassland
{0,1}, Temperate Grassland {0,1} and Montane Grassland {0,1}.
14 The inclusion criteria for both rivers, i.e. “navigability” as well as lakes, i.e. “major”, is defined as in
Henderson et al. (2018): we select all natural rivers within size categories 1-5 (scale 1-7) as defined in Natural
Earth (2019) and lakes with a surface area of over 5,000 sq. kilometers (Lehner and Döll 2004).
11
Table 2.1: Summary Statistics
Panel a)
Mean St.Dev. Min. 1st Quartile Median 3rd Quartile Max.
N
Basic Characteristics
Age
36
14
18
25
33
44
130
127,462
Female {0,1}
0.5
0.5
0.0
0.0
1.0
1.0
1.0
128,747
Educational Level {0,9}
3.4
2.2
0.0
2.0
4.0
5.0
9.0
128,211
Trade-related Covariates
Distance to Harbor (in km)
345
289
0
82
282
564
1346
128,804
Distance to Coast (in km)
278
276
0
26
188
468
1328
128,804
Distance to Navigable River (in km)
296
229
0
127
234
422
1111
128,868
Distance to Major Lake (in km)
814
683
0
230
623
1309
2749
128,868
Urbanization Covariates
Urban {0,1}
0.5
0.5
0.0
0.0
0.0
1.0
1.0
128,656
Primate City ≤ 25km {0,1}
0.2
0.4
0.0
0.0
0.0
0.0
1.0
128,868
Population Density (per sq. km)
2.0
4.8
0.0
0.1
0.2
1.3
125.4
128,848
Geographical Covariates
Absolute Latitude
12
9
0
6
8
16
35
128,868
Elevation (in m)
543
595
0
48
276
1094
3914
128,860
Terrain Ruggedness (standardized)
0.0
1.0
-0.7
-0.6
-0.3
0.1
17.1
128,860
Land Suitability [0,1]
0.5
0.2
0.0
0.3
0.4
0.6
1.0
128,802
Average Monthly Temperature (in Celsius)
26
4
8
24
28
29
33
128,860
Average Monthly Rainfall (in mm)
108
63
1
68
94
127
384
128,860
Growing Days {0,365}
231
82
0
178
244
296
365
128,868
Malaria Ecology Index
12
10
0
0
13
23
33
128,802
Mediterranean {0,1}
0.0
0.1
0.0
0.0
0.0
0.0
1.0
128,868
Desert {0,1}
0.0
0.2
0.0
0.0
0.0
0.0
1.0
128,868
Mangroves {0,1}
0.0
0.2
0.0
0.0
0.0
0.0
1.0
128,868
Tropical Forest {0,1}
0.3
0.5
0.0
0.0
0.0
1.0
1.0
128,868
Tropical Grassland {0,1}
0.5
0.5
0.0
0.0
1.0
1.0
1.0
128,868
Temperate Grassland {0,1}
0.0
0.1
0.0
0.0
0.0
0.0
1.0
128,868
Montane Grassland {0,1}
0.1
0.3
0.0
0.0
0.0
0.0
1.0
128,868
Distribution across Sample
Notes: The table depicts summary statistics corresponding to the main sample used in the estimations across the paper.
Data encompasses individual-level responses from 17 coastal countries in sub-Saharan Africa and come from the 1999-
2001, 2002-2004, 2005-2006, 2008-2009, 2011-2013, 2014-2015 and 2016-2018 (i.e. Round 1 through Round 7) geo-
referenced Afrobarometer survey rounds. Variation in the number of observations size stem from differences in response
rates of variables as well as changes in questions asked across surveys. Geographic covariates come from an array of sources
described in the section on data (2.2).
12
Panel b)
Mean St.Dev. Min.
1st Q.
2nd Q.
3rd Q.
4th Q.
N
Dependent Variables
Cash Employment {0,1}
0.39
0.49
0
0.40
0.38
0.40
0.38
123,857
How often: Gone without (...) [0,4]
1.28
0.99
0
1.20
1.33
1.28
1.30
128,673
How often: Gone without Food {0,4}
0.97
1.16
0
0.94
1.03
0.93
0.99
128,420
How often: Gone without Water {0,4}
1.14
1.36
0
1.04
1.21
1.16
1.15
128,491
How often: Gone without Cash Income {0,4}
1.90
1.36
0
1.80
1.95
1.90
1.94
122,427
How often: Gone without Medical Care {0,4}
1.15
1.26
0
1.05
1.22
1.16
1.18
128,169
Possessions: (...) [0,1]
0.51
0.34
0
0.58
0.49
0.50
0.45
103,953†
Possessions: TV {0,1}
0.48
0.50
0
0.65
0.46
0.46
0.33
103,646†
Possessions: Radio {0,1}
0.75
0.43
0
0.77
0.74
0.74
0.73
103,902†
Possessions: Motor Vehicle {0,1}
0.30
0.46
0
0.32
0.28
0.31
0.29
103,452†
Pathways
Most Important Issue: Education {0,1}
0.06
0.24
0
0.06
0.07
0.06
0.06
122,062
Most Important Issue: Institutions {0,1}
0.09
0.29
0
0.10
0.07
0.10
0.10
122,062
Most Important Issue: Infrastructure {0,1}
0.06
0.24
0
0.05
0.08
0.06
0.05
122,062
Present in EA: (…) [0,1]
0.56
0.30
0
0.67
0.54
0.54
0.47
117,554‡
Present in EA: Electricity Grid {0,1}
0.64
0.48
0
0.82
0.61
0.61
0.53
116,989‡
Present in EA: Piped Water {0,1}
0.55
0.50
0
0.74
0.50
0.55
0.40
116,304‡
Present in EA: Sewage {0,1}
0.29
0.45
0
0.41
0.26
0.30
0.18
115,159‡
Present in EA: Paved Road {0,1}
0.43
0.49
0
0.54
0.41
0.42
0.33
116,975‡
Present in EA: School {0,1}
0.86
0.35
0
0.88
0.88
0.83
0.84
116,668‡
Present in EA: Health Clinic {0,1}
0.57
0.49
0
0.63
0.58
0.54
0.53
115,179‡
Institutions Score [1,4]
2.81
0.53
1
2.72
2.75
2.85
2.91
128,347
Helps your Country: REC {0,4}
1.80
0.97
0
1.72
1.79
1.81
1.86
35,710'
Helps your Country: AU {0,4}
1.68
0.99
0
1.64
1.64
1.71
1.71
47,726*
Occupation: Commercial Farmer {0,1}
0.01
0.10
0
0.01
0.01
0.01
0.01
37,349**
Distribution across Distance (within Quartiles)
Notes: The table depicts summary statistics corresponding to the main sample used in the estimations across the paper.
Data encompasses individual-level responses from 17 coastal countries in sub-Saharan Africa and come from the 1999-
2001, 2002-2004, 2005-2006, 2008-2009, 2011-2013, 2014-2015 and 2016-2018 (i.e. Round 1 through Round 7) geo-
referenced Afrobarometer survey rounds. Variation in the number of observations size stem from differences in response
rates of variables as well as changes in questions asked across surveys. Geographic covariates come from an array of sources
described in the section on data (2.2).
† Not asked in survey rounds 1 and 2. ‡ Not asked in survey round 1. ' Only asked in survey rounds 2, 4 and 6. * Only
asked in survey rounds 2, 4, 5 and 6. ** Only asked in survey rounds 1, 2 and 3.
13
2.3 Empirical Strategy
We employ the following regression control approach to analyze the link between coastal
proximity and individual economic welfare:
𝑌𝑌𝑖𝑖,𝑐𝑐,𝑡𝑡= 𝛼𝛼+ 𝛽𝛽log(𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡𝑡𝑡 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻)𝑖𝑖+ 𝛾𝛾𝑿𝑿′𝑖𝑖+ 𝛿𝛿𝑐𝑐,𝑡𝑡+ 𝜀𝜀𝑖𝑖,𝑐𝑐,𝑡𝑡
(2.1)n
𝑌𝑌𝑖𝑖,𝑐𝑐,𝑡𝑡 represents the respective welfare indicator of individual 𝑖𝑖 in country 𝑐𝑐, surveyed at survey-
sampling period 𝑡𝑡. 𝛽𝛽 captures the influence of the logged (within-country) distance to major
harbors such that the link between distance and the respective welfare indicator can be
interpreted as a semi-elasticity. Standard errors are clustered at the level of the survey
enumeration area, i.e. at the survey cluster level. Binary dependent variables are estimated
with a simple Linear Probability Model (LPM) specification.15 𝑿𝑿 represents a vector of control
variables which allows us to account for all influences potentially conflating the relationship
between coastal distance and individual economic welfare. In contrast to the cross-country
(cross-regional) literature, our setting allows us to account for country-time fixed effects 𝛿𝛿𝑐𝑐,𝑡𝑡
such that we can explore a within-country estimate of distance to harbor on (individual)
outcomes net of time-specific influences as well as country-specific influences at specific points
in time, such as the Kenyan Post-Election Crisis of 2007-2008. 𝜀𝜀𝑖𝑖,𝑐𝑐,𝑡𝑡 is an idiosyncratic error
term.
We explore potential mechanisms and factors affecting the link between coastal proximity
and individual living standards both via a “bad control” approach as well as a formal mediation
analysis, after establishing the relevance of coastal proximity for individual living standards.
Numerous robustness checks for the persistence of the observed links are offered (mostly
relegated to the Appendix).
2.4 Results
Table 2.2 presents the main estimation results employing our three distinct individual welfare
indicators as dependent variables. We report a parsimonious specification including country-
time fixed effects in the odd-numbered columns. Even numbered columns include the full set
of controls and represent our stringent setting.16
The results systematically indicate that distance to harbors is inversely related to
individual economic welfare throughout all specifications.17 To facilitate the interpretation of
15 Results for binary dependent variables estimated via Probit yields qualitatively identical and quantitatively
similar marginal effects. Results can be obtained from the authors.
16 See Table A.1 for coefficient estimates of all (geographic) control variables.
17 The variation in the number of observation stems from missing values. Holding the sample size constant by
eliminating observations for which not all dependent/independent variables are available does not change our
main insights as shown in Table A.8 of the Appendix.
14
Table 2.2: Coastal Proximity and Individual Living Standards
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.018***
-0.009***
0.073***
0.035***
-0.037***
-0.019***
(0.001)
(0.002)
(0.004)
(0.006)
(0.001)
(0.002)
Discrete Change of
Distance from Harbor to
the 3rd Quartile (564km)
-0.115
-0.055
0.460
0.223
-0.234
-0.120
Sample Mean of Dep. Var.
Basic Controls
Age
0.031***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.106***
0.020***
-0.083***
(0.003)
(0.004)
(0.002)
Urbanization Controls
Urban {0,1}
0.055***
-0.284***
0.134***
(0.004)
(0.010)
(0.003)
Primate City {0,1}
0.029***
-0.084***
0.029***
(0.007)
(0.016)
(0.006)
Population Density
-0.001
0.004***
-0.001*
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.013
-0.121***
0.029***
(0.009)
(0.021)
(0.008)
Major Lake {0,1}
0.010
0.033
-0.023***
(0.012)
(0.024)
(0.008)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,793
122,238
128,609
126,982
103,889
102,990
R-Squared
0.09
0.14
0.17
0.20
0.15
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.28]
[0.51]
Notes: Results in each column come from separate regressions and are estimated using the main
sample of coastal sub-Saharan African countries included in survey rounds 1 through 7 of the
Afrobarometer. Changes in the number of observations across columns stem from differences in the
response rates of dependent/independent variables. The samples used in columns (5) and (6) do not
include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of
household items were not asked in these rounds. We also report estimated interquartile differences
in the respective dependent variables between minimum- and third quartile harbor distances within
the sample. Binary dependent variables are estimated through a simple linear probability model
specification. The standard errors reported are clustered at the survey enumeration area level. ***,
**, * represents significance at the 1, 5 and 10 percent level, respectively.
15
the quantitative relevance of the main explanatory variable of interest, log(Distance to Harbor),
we report the predicted change of the respective dependent variable when moving from the
minimum distance in the sample (i.e. effectively living by a major port) to living as far as 564
kilometers away from the harbor (3rd quartile of sample) and compare the predicted change of
each individual welfare indicator to the respective sample mean reported in brackets. The
results show that distance to harbors is statistically significantly and negatively related to cash
employment (columns 1 and 2) and positively, statistically significantly, related to deprivation
(columns 3 and 4). Quantitatively, increasing individuals’ distance to major ports to the 3rd
quartile in the sample translates into a 5.5 percentage-point decrease in the probability of
having part- or full-time cash-employment (column 2) and can explain 17% of the occurrence
of monetary-, medicinal- as well as food- and water-related shortages compared to the mean in
the sample (column 4). Coastal remoteness is significantly related to having fewer (wealth)
possessions: increasing individuals’ distance to the 3rd quartile corresponds to a 12 percentage-
point decrease in the probability of owning a radio, a tv or a motor vehicle, accordingly, a 23%
reduction compared to the mean in the sample (column 6). Importantly, the results for our
indices of deprivation (columns 3 and 4) and possessions (columns 5 and 6) also hold when
analyzing the variables that compose our indices separately (see Table A.2).
2.4.1. Robustness Checks and Extensions
We conduct a large array of robustness checks on our main results and summarize them in
Table A.3 of the Appendix. All interpretations regarding the relevance of coastal proximity for
individual living standards remain robust. (a) We re-estimate our main results by altering the
distance specification to a simple “beeline” (“as the crow flies”) measure. (b) We use a different
conceptualization of coastal proximity by regressing our outcome variables on individuals’
distance to the coastline instead of port locations, using log(Distance to Coastline). (c)
Accordingly, we test beeline distances to the coastline with log(Beeline Distance to Coastline).
We add dummies for living within 25 kilometers to a major harbor (d) or the coast (e), Harbor
{0,1} and Coast {0,1} to separate the distance effect from a pure “coastal access” effect. We
also investigate differing effects at distinct distance increments in Figure A.1. Distance to
harbors is consistently, and importantly, increasingly related to lower living standards, negating
a potential “binary effect” of merely living at these major harbors or not. (f) We keep
observations constant across rows and columns. (g) We exclude distances larger than the 80th
percentile (629 kilometers) from the sample. (h) We exclude localities marked with a precision
code of 2 and larger in the Afrobarometer survey (scale 1-8) from our sample. (i) We include
survey sampling weights. (j) We employ clustering at the country-sample level. Moreover, we
check the main coefficient’s stability to potentially excluded controls via Oster tests (Oster
2019) in Table A.13. All robustness checks corroborate our general findings of a negative,
independent, statistically significant relationship between coastal proximity and individual
16
living standards. The results reiterate the relevance of coastal proximity, in varying
conceptualizations, in predicting individual economic welfare.
Next to the above-mentioned robustness tests, we extend our analysis and (a) expand
our main sample to include individuals living in landlocked countries (see “Extended Sample”
in Figure 2.1) and also (b) analyze the persistence of our estimated effects over time. For (a),
we include individuals living in landlocked countries, to explore a potential “placebo” group
compared to individuals living in coastal countries. This allows us to compare the effect of
sheer coastal distance within countries from a landlockedness-effect, the one often explored in
the literature (UN-OHRLLS 2013). The idea is that differences in individual coastal proximity
within landlocked countries should influence individual welfare to a lower degree given that
national borders need to be crossed, creating other, potentially large restrictions unrelated to
sheer distance (Faye et al. 2004).18 As expected, Table A.14 suggests that the relevance of
individual distance to harbors tends to be less pronounced for individuals living in landlocked
countries. (b) The relative importance of trade-related factors of geography might be expected
to change along a country’s developmental path (see Henderson et al. 2018). Hence, we estimate
differential effects using an interaction effect constituted of log(Distance to Harbor) and Young
{0,1}, which indicates respondents below the median age in the sample (33). The results in
Table A.15 of the Appendix show a clear pattern. The negative effect of distance becomes less
stark for younger generations, potentially hinting at a reduction in the relevance of trade-
related aspects over time (see Henderson et al. 2018).19
2.4.2. Mechanisms Explaining the Relevance of Coastal Proximity
Table 2.3 explores potential mechanisms through which coastal proximity may influence
individual economic welfare. Following the literature, we focus on the link between coastal
proximity and human capital, urbanization, institutional quality, infrastructural development
and the perceived relevance of trade to investigate potential (indirect) channels that explain
the spatial economic disparity given by individual geographic distance to harbors (see Breinlich
et al. 2014 for an overview).
Individual educational attainment has been linked to economic welfare at the cross-
regional level (e.g. Skoufias and Katayama 2011; Gennaioli et al. 2013; Chauvin et al. 2017;
Flückiger and Ludwig 2018). Individuals’ opinions regarding education, as shown in column
(1), do not mirror these findings, as respondents living in more remote locations do not report
education as the most important issue (facing the country/government) more often than
individuals living closer to the coast. However, we do find that individuals’ actual educational
attainment decreases substantially along coastal distance (column 2): moving from the
18 Empirically, we add an interaction term constituted of log(Distance to Harbor) and a binary variable
indicating whether the country is Landlocked. The sum of the coefficients log(Distance to Harbor) and the
interaction term represents the total effect of distance to coast for individuals living in landlocked countries.
19 We thank an anonymous referee for pointing out this additional extension.
17
Table 2.3: Exploring Potential Mechanisms of Influence of Coastal Proximity
Most important
Issue:
Education
{0,1}
Education:
Educational
Level
{0,9}
Type of
Residence:
Urban
{0,1}
Most important
Issue:
Institutions
{0,1}
Institutions
Score
[1,4]
Most important
Issue:
Infrastructure
{0,1}
Infrastr. in EA:
(Elec.Grid/Pip.Wa
ter/Sewage/School
/Pave.Road/Clinic)
[0,1]
Helps your
Country:
REC
{0,3}
Helps your
Country:
AU
{0,3}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
log(1+Distance to Harbor)
0.001
-0.106***
-0.071***
0.002
0.025***
0.005***
-0.009***
0.034***
0.046***
(0.001)
(0.012)
(0.007)
(0.001)
(0.003)
(0.001)
(0.003)
(0.009)
(0.008)
Discrete Change of
Distance from Harbor to
the 3rd Quartile (564km)
0.007
-0.670
-0.449
0.010
0.158
0.030
-0.056
0.214
0.293
Sample Mean of Dep. Var.
[0.06]
[3.40]
[0.46]
[0.09]
[2.81]
[0.06]
[0.56]
[1.80]
[1.68]
Basic Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Urbanization Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Trade-related Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Full Geographic Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
YES
YES
YES
Observations
120,461
126,555
127,075
120,461
126,683
120,461
116,174
35,409
47,298
R-Squared
0.03
0.29
0.30
0.06
0.21
0.04
0.43
0.11
0.07
Dependent Variable
Notes: Results in each column come from separate regressions and are estimated using the main sample of coastal sub-Saharan African countries included in survey rounds 1
through 7 of the Afrobarometer. Changes in the number of observations across columns stem from differences in the response rates of dependent/independent variables. The
sample used in column (7) includes individuals from rounds 2 through 7 of the Afrobarometer, columns (8) and [9] includes data from rounds 2, 4, [5], and 6. We also report
estimated interquartile differences in the respective dependent variables between minimum- and third quartile harbor distances within the sample. Binary dependent variables
are estimated through a simple linear probability model specification. The standard errors reported are clustered at the survey enumeration area level. AU, African Union; REC,
Regional Economic Community. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
18
minimum distance to the 3rd quartile within the sample reduces the level of education by .670,
which corresponds to about 20% of the sample mean.
As shown by recent literature (Motamed et al. 2014; Henderson et al. 2018), levels of
urbanization are negatively correlated with increased distance to coast in Africa, and can
negatively impact individual economic welfare directly, or indirectly, via agglomeration
economies (e.g. Skoufias and Katayama 2011; Bosker and Garretsen 2012; Young 2013; Gollin,
Kirchberger, and Lagakos 2017; Flückiger and Ludwig 2018; Henderson et al. 2018). Consistent
with this literature, we also find a negative, statistically significant, and quantitatively large
relationship between (coastal) remoteness and living in urban environments (column 3). Given
the strong interconnection between coastal proximity and urbanization, we also explore the
differential effects of distance for individuals living in urban environments in Table A.16
separately, by estimating the interaction term log(Distance to Harbor) × Urban {0,1}. The
results show that, while less pronounced, the distance penalty remains for two of our three
outcomes for respondents from urban settings.
Regarding institutions, we proceed similarly to Mitton (2016; 107) and construct an
index, Institutions Score [1,4], which combines responses concerning individuals’ experiences
with- and opinions on local authorities, offices and government. The results suggest that
individuals living further away from the coast do not report institutions to be at the top of
their concerns (column 4), nor is the institutional score negatively affected when living further
way from major harbors (column 5). Recent literature has also suggested a weak link between
institutions and differences in subnational development within countries (Radeny and Bulte
2012; Michalopoulos and Papaioannou 2014; Mitton 2016). In fact, individuals living in interior
regions even seem to evaluate institutions more positively compared to coastal areas in our
findings (column 5), a result which mirrors the one in Radeny and Bulte (2012) as well as Nunn
and Wantchekon (2011), whereby distance to coast positively influenced levels of trust via a
lower intensity of slave trade.
Infrastructure has been highlighted as a relevant factor for regional development (e.g.
Calderón and Servén 2010; Dinkelman 2011; Jedwab and Moradi 2016; Storeygard 2016; Bluhm
et al. 2018; Donaldson 2018; Jetter, Mösle, and Stadelmann 2019). Consistent with this
literature, our results at the individual level show that coastal proximity is negatively associated
with respondents’ sentiments that infrastructure needs are issues of concern (see column 6 and
7). The actual access to basic infrastructure (as measured by our composite infrastructure
index), is also negatively associated with distance to major ports.
Increased trade costs and reduced market access have been shown to be an inherent issue
of remote areas in Africa (e.g. Bosker and Garretsen 2012; Atkin and Donaldson 2015;
Henderson et al. 2018; Jedwab and Storeygard 2020). While trade volumes are necessarily an
aggregate phenomenon, we find that survey respondents further away from the coast exhibit a
higher tendency to report their respective Regional Economic Communities (RECs) or the
19
African Union (AU) as helpful to their country, which is consistent with them wishing to
improve trade opportunities. Moreover, Table A.17 shows that the distance penalty is
significantly increased for commercial farmers, i.e. farmers who mainly grow their produce for
sale. Commercial farming is likely to depend on access to markets and trade opportunities,
leaving commercial farmers more vulnerable to a distance penalty.
2.4.3. Bad Controls and Mediation Analysis
All results highlight coastal proximity as a statistically as well as an economically meaningful
indicator of individual living standards and as a relevant predictor for diverse mechanisms that
systematically relate and contribute to economic development and spatial inequalities. As
coastal remoteness need not be destiny (Motamed et al. 2014), we aim to gauge the empirical
importance of our controls as well as the potential mechanisms on our main explanatory
variable by investigating the relevance of a bad controls problem and by performing a formal
mediation analysis.
Bad Controls. We add in all of our baseline covariates and the explored mechanisms in step-
wise fashion and report the corresponding changes to our main coefficient, log(Distance to
Harbor), as well as changes in the residual variance. Results are presented in Table 2.4. Row
(a) shows the coefficient of log(Distance to Harbor) in a regression including country-time fixed
effects only. Row (b) proceeds to add in our basic controls, i.e. Age, Age squared and Female
{0,1}, as is done in Table 2.2. Row (c) adds our three urbanization controls to the specification,
and so on.21 The results show that, while the coefficient size of log(Distance to Harbor)
diminishes, as is expected, coastal proximity remains a statistically relevant predictor of
individual living standards throughout all rows and columns. The covariates contributing most
to the specifications, as seen by changes in the coefficient (odd column numbers) as well as
changes in the R-squared (even column numbers), are Urbanization Controls, Educational Level
and Infrastructure, which are the ones we will explore as potential mediators next.
21 We do not add sentiments of RECs or the AU to the list of covariates as their availability across survey
rounds is sparse, observations size would drop by ~50%.
20
Table 2.4: Bad Controls, Relevance of included Covariates
log(1+Distance
to Harbor)
∆ R-
Squared
log(1+Distance
to Harbor)
∆ R-
Squared
log(1+Distance
to Harbor)
∆ R-
Squared
(1)
(2)
(3)
(4)
(5)
(6)
(a) No Controls
-0.020***
0.076***
-0.037***
(0.001)
(0.004)
(0.001)
(b) = (a) + Basic Controls
-0.020***
0.075***
-0.037***
(0.001)
(0.004)
(0.001)
(c) = (b) + Urbanization Controls
-0.009***
0.024***
-0.014***
(0.002)
(0.004)
(0.001)
(d) = (c) + Trade-related Controls
-0.009***
0.024***
-0.014***
(0.002)
(0.004)
(0.001)
(e) = (d) + Geographic Controls
-0.009***
0.024***
-0.019***
(0.002)
(0.004)
(0.002)
(f) = (e) + Educational Level
-0.006**
0.029***
-0.015***
(0.002)
(0.006)
(0.002)
(g) = (f) + Institutions Score
-0.006**
0.036***
-0.015***
(0.002)
(0.006)
(0.002)
(h) = (g) + Infrastructure
-0.005**
0.032***
-0.014***
(0.002)
(0.006)
(0.002)
Country-Time FE
Observations
R-Squared
YES
114,857
115,307
102,287
0.18
0.25
0.28
[0.010]
[0.000]
[0.000]
[0.000]
[0.001]
[0.000]
[0.003]
[0.017]
[0.000]
[0.000]
[0.035]
[0.016]
[0.011]
[0.056]
[0.000]
Dependent Variable
Cash
Employment
{0,1}
How often gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
-
-
-
[0.054]
[0.003]
[0.026]
[0.004]
[0.021]
[0.036]
YES
YES
Notes: Odd columns present the coefficient (changes) of our main explanatory variable log(1+Distance to Harbor)
when subsequently adding seven distinct (sets of) control variables to a parsimonious baseline regression,
constituted of our main regressor and country-time fixed effects. Even columns report the corresponding changes
in the total R-squared values compared to the previous specification. The results in each row come from separate
regressions, and observations are held constant across rows. Inclusion of mediating factors in (h), variables on
infrastructure, limits the sample to rounds 2 to 7 of coastal sub-Saharan African countries included in the
Afrobarometer. The remaining changes in the number of observations across columns stem from differences in the
response rates of dependent variables (see notes in Table 2.2). Binary dependent variables are estimated through
a simple linear probability model specification. Test statistics at the bottom of the table are produced from the
full regression, that is, specification (h). The standard errors reported are clustered at the survey enumeration area
level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
21
Mediation Analysis. To further evaluate the link between coastal proximity and individual
economic welfare, as well as its potential channels of influence, we conduct a formal mediation
analysis. We empirically decompose the total effect of coastal proximity and individual welfare
into indirect effects, i.e. effects which run through the proposed mediating factors, and direct
effects, i.e. effects of coastal proximity that are unrelated to the proposed channels.
𝑌𝑌𝑖𝑖,𝑐𝑐,𝑡𝑡= 𝛼𝛼1 + 𝛽𝛽1 log(𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡𝑡𝑡 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻)𝑖𝑖+ 𝜃𝜃𝜃𝜃𝑖𝑖+ 𝛾𝛾1𝑿𝑿′𝑖𝑖+ 𝛿𝛿𝑐𝑐,𝑡𝑡+ 𝜀𝜀𝑖𝑖,𝑐𝑐,𝑡𝑡
(2.2)n
𝑀𝑀𝑖𝑖,𝑐𝑐,𝑡𝑡= 𝛼𝛼2 + 𝛽𝛽2 log(𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡𝑡𝑡 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻)𝑖𝑖+ 𝛾𝛾2𝑿𝑿′𝑖𝑖+ 𝛿𝛿𝑐𝑐,𝑡𝑡+ 𝜇𝜇𝑖𝑖,𝑐𝑐,𝑡𝑡
(2.3)n
𝛽𝛽1 measures the direct effect of coastal proximity on our different welfare indicators 𝑌𝑌, and 𝛽𝛽2
measures the effect of distance to harbor on the respective mediator 𝑀𝑀 (e.g. education,
urbanity, infrastructure). 𝜃𝜃 represents the direct effect of the mediator 𝑀𝑀 on the outcome
variable such that the indirect effect is retrieved by multiplying 𝛽𝛽2 × 𝜃𝜃 (Alwin and Hauser 1975;
MacKinnon et al. 2007). The total effect is then given by a summation of the direct (𝛽𝛽1) and
indirect effects (𝛽𝛽2 × 𝜃𝜃).22 Figure A.2 provides a visual representation of the mediation analysis.
As before, 𝑿𝑿 is a vector including all of our usual controls. We keep country-time fixed effects
𝛿𝛿𝑐𝑐,𝑡𝑡 to evaluate a stringent setting.
Table 2.5 reports the coefficients of the total, direct and indirect effects of coastal
proximity on individual economic welfare. Estimations are performed via structural equation
modelling (SEM). To save space, we present the mechanisms on which distance to harbor had
the largest impact in Table 2.3 and Table 2.4, Education Level, Urbanization Controls and
Infrastructure and estimate their mediating effect on our three main outcome variables (results
for our proxies of Institutions and Trade are relegated to Table A.18 in the Appendix).
The results in Table 2.5 suggest that a substantial part of the total effect of distance to
harbors is mediated by educational attainment. Including respondents’ level of schooling in the
main specification (equation (2.2)) reduces the coefficient size of the direct effect of coastal
proximity by 28% (see proportion mediated at the bottom of the table) on average, i.e. across
outcome variables. In other words, coastal proximity matters for educational outcomes, and
through education, it matters for individuals’ living standards, subsequently. The direct effect
of education on living standards is quantitively large and statistically significant throughout
all estimations, indicating a relevant effect of education on economic welfare on its own. These
results are in line with cross-country and subnational evidence, identifying educational
differences as an important factor for explaining disparities in economic development (see
Skoufias and Katayama 2011; Gennaioli et al. 2013; Chauvin et al. 2017; Flückiger and Ludwig
2018).
22 Slight deviations in coefficients between the total effect in Table 2.2 and Table 2.5 arise because of missing
values of the respective mediator variables introduced.
22
Table 2.5: Mediation Analysis: Direct and Indirect Links of Distance to Harbor
Mediator:
Education
{0,9}
Mediator:
Urban
{0,1}
Mediator:
Infrastructure
[0,1]
Mediator:
Education
{0,9}
Mediator:
Urban
{0,1}
Mediator:
Infrastructure
[0,1]
Mediator:
Education
{0,9}
Mediator:
Urban
{0,1}
Mediator:
Infrastructure
[0,1]
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
-0.009***
-0.012***
-0.009***
0.035***
0.055***
0.037***
-0.019***
-0.028***
-0.019***
(0.002)
(0.002)
(0.002)
(0.006)
(0.006)
(0.006)
(0.002)
(0.002)
(0.002)
-0.005**
-0.009***
-0.008***
0.027***
0.035***
0.032***
-0.015***
-0.019***
-0.017***
(0.002)
(0.002)
(0.002)
(0.006)
(0.006)
(0.006)
(0.002)
(0.002)
(0.002)
Distance to Harbor via Mediator
-0.003***
-0.004***
-0.001***
0.009***
0.020***
0.005***
-0.004***
-0.009***
-0.002***
(0.000)
(0.000)
(0.000)
(0.001)
(0.002)
(0.002)
(0.001)
(0.001)
(0.001)
0.033***
0.055***
0.089***
-0.082***
-0.284***
-0.582***
0.044***
0.134***
0.210***
(0.001)
(0.004)
(0.008)
(0.002)
(0.010)
(0.019)
(0.001)
(0.003)
(0.006)
[0.39]
[0.30]
[0.09]
[0.25]
[0.36]
[0.14]
[0.21]
[0.33]
[0.10]
Basic Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Urbanization Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Trade-related Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Full Geographic Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
YES
YES
YES
Observations
121,823
122,238
115,480
126,401
126,982
116,084
102,514
102,990
102,990
Possessions:
(Radio / TV / Motor Vehicle)
[0,1]
How often gone without:
(Water / Food / Cash Inc. / Med.)
[0,4]
Dependent Variable
Indirect Effect:
Direct Effect of Mediator:
Cash
Employment
{0,1}
(Education / Urban / Infrastructure)
Proportion Mediated
Baseline (Total) Effect:
log(1+Distance to Harbor)
Direct Effect:
log(1+Distance to Harbor)
Notes: This table presents results from a formal mediation analysis testing the influence of three potential mediators of log(1+Distance to Harbor) on our three main
outcome variables, respectively. Row 1 presents the baseline effect; its counterpart is shown in Table 2.2. Row 2 shows the direct effect of our main explanatory variable,
that is, the effect of distance not attributable to the mediating factor, while row 3 depicts the part of the effect that runs precisely via its influence on the mediator. Row
4 shows the direct effect of the mediator on the respective outcome variable. Row 5 provides the proportion mediated, which is given by dividing the indirect effect by the
total effect. Results in each column come from a separate SEM regression. The sample used is comprised of coastal sub-Saharan African countries included in survey
rounds 1 through 7 of the Afrobarometer. Columns 3, 6, [7], [8], and [9] do not include individuals surveyed in rounds 1 [and 2] of the Afrobarometer, as questions on
infrastructure (household items) were not asked in this round. Binary dependent variables are estimated through a simple linear probability model specification. The
standard errors reported are clustered at the survey enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
23
Similar insights arise for the role of urbanity in explaining relevant parts of distance’s effect on
living standards, mediating 33% of the effect on average. The mediator Urban {0,1} is therefore
picking up a substantial part of the total effect of coastal distance, in similar magnitude as do
educational differences. It is important to note that while both education and urbanization
absorb variation in explaining individual living standards on their own (Table 2.4), as well as
through their mediation of coastal proximity (Table 2.5), empirically, we cannot fully separate
them. Indeed, existing literature has provided evidence suggesting that they are interrelated
and mutually reenforced (e.g. Skoufias and Katayama 2011; Chauvin et al. 2017; Flückiger and
Ludwig 2018).
Infrastructure, proxied by our composite measure Infrastructure Present in
Enumeration Area: (Electricity Grid / Piped Water / Sewage / School / Paved Road / Health
Clinic) [0,1], while relevant, does not show for an influence in similar magnitudes as do
education or urbanization, mediating only an average of 11% of the effect. Table A.18 explores
the role of institutions as a mediator in explaining the distance penalty. Contrary to human
capital and urbanity, the pronounced gap in individual living standards across distances does
not seem to be associated with perceived differences in (local) institutional quality when
controlling for country-time fixed effects. Also, our evidence for a positive, direct effect of
institutional quality on individual economic welfare is mixed (see row 4), consistent with other
findings from subnational (regional) contexts (see Michalopoulos and Papaioannou 2014;
Gennaioli et al. 2013; Mitton 2016).
The relevance of coastal proximity on economic development has often been ascribed to
trade-related factors, especially among “late developers” (see Henderson et al. 2018). Table
A.18 explores this link, estimating the direct and indirect effect of regional and supra-regional
institutions fostering trade, as measured by respondents’ evaluation of the African Union (AU)
and their “corresponding” Regional Economic Community (REC), respectively. The results
show that more positively perceived trade organizations correlate positively with individual
living standards (row 4), which emphasizes a potential need for trade facilitation independent
of individuals’ distance to harbors, i.e. (global) markets.
2.5 Conclusion
We systematically investigate the role of coastal proximity in explaining intra-national
differences in individual living standards across sub-Saharan Africa economies using an
extensive dataset covering up to 128,609 observations distributed across 11,261 localities over
20 years. We employ geo-referenced individual-level data to complement the existing literature
that focuses on outcomes at the national or regional level. Analyzing individuals’ distance to
harbors and their corresponding living standards allows us to test whether the insights of the
cross-country and cross-regional contexts also apply at the individual level. Moreover, we can
utilize the comprehensiveness of our dataset to explore a large set of indicators and potential
24
channels of influence to gauge the relevance of coastal proximity and to investigate the
mechanism through which it may matter for individual living standards.
Our results show that coastal proximity, as measured by geographical distance to harbors,
predicts a relevant part of individual living standards and remains a strong predictor of
individual economic welfare controlling for individual-level covariates, country-time specific
influences via fixed-effects, as well as an extensive set of other established geographical
influences of development.
Exploring potential channels, we find that human capital, urbanization, as well as access
to infrastructure mediate relevant parts of the link between coastal proximity and economic
development. This highlights that even though coastal proximity is a relevant indicator for
individual living standards across Africa, coastal proximity need not be “destiny”. Fostering
education as well as infrastructural outlays might help in mitigating problems associated with
coastal remoteness. Nevertheless, the systematic robustness of coastal proximity as a predictor
for individual living standards, even in stringent settings, suggests that there are relevant
development costs of remoteness alone that need to be addressed (see also UN-OHRLLS 2013).
25
2.6 References
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2001. ‘The Colonial Origins of
Comparative Development: An Empirical Investigation’. American Economic Review
91 (5): 1369–1401.
Afrobarometer, [Kenya, Tanzania, Uganda] [Round 1, 2, 3, 4, 5, 6, 7]. 2019. ‘Afrobarometer
Data’.
Alwin, Duane F., and Robert M. Hauser. 1975. ‘The Decomposition of Effects in Path Analysis’.
American Sociological Review 40 (1): 37.
Atkin, David, and Dave Donaldson. 2015. ‘Who’s Getting Globalized? The Size and
Implications of Intra-National Trade Costs’. w21439. Cambridge, MA: National Bureau
of Economic Research.
BenYishay, Ariel, Renee Rotberg, Jessica Wells, Zhonghui Lv, Seth Goodman, Lidia Kovacevic,
and Dan Runfola. 2017. ‘Geocoding Afrobarometer Rounds 1-6: Methodology & Data
Quality’. Williamsburg, VA: AidData at William & Mary.
Bleakley, Hoyt, and Jeffrey Lin. 2012. ‘Portage and Path Dependence *’. The Quarterly Journal
of Economics 127 (2): 587–644.
Bloom, David E., Jeffrey D. Sachs, Paul Collier, and Christopher Udry. 1998. ‘Geography,
Demography, and Economic Growth in Africa’. Brookings Papers on Economic Activity
1998 (2): 207.
Bluhm, Richard, Axel Dreher, Andreas Fuchs, Bradley Parks, and Austin Strange. 2018.
‘Connective Financing: Chinese Infrastructure Projects and the Diffusion of Economic
Activity in Developing Countries’. In AidData Working Paper #64. Williamsburg, VA:
Data at William & Mary.
Bosker, Maarten, and Harry Garretsen. 2012. ‘Economic Geography and Economic
Development in Sub-Saharan Africa’. The World Bank Economic Review 26 (3): 443–
85.
Breinlich, Holger, Gianmarco I.P. Ottaviano, and Jonathan R.W. Temple. 2014. ‘Regional
Growth and Regional Decline’. In Handbook of Economic Growth, 2:683–779.
Amsterdam.
Brülhart, Marius. 2011. ‘The Spatial Effects of Trade Openness: A Survey’. Review of World
Economics 147 (1): 59–83.
Calderón, César, and Luis Servén. 2010. ‘Infrastructure and Economic Development in Sub-
Saharan Africa’. Journal of African Economies 19 (suppl_1): i13–87.
Chauvin, Juan Pablo, Edward Glaeser, Yueran Ma, and Kristina Tobio. 2017. ‘What Is
Different about Urbanization in Rich and Poor Countries? Cities in Brazil, China, India
and the United States’. Journal of Urban Economics 98: 17–49.
CIESIN. 2017. ‘Gridded Population of the World, Version 4 (GPWv4): Population Density,
Revision 11’. Center For International Earth Science Information Network-CIESIN-
Columbia University. Socioeconomic Data and Applications Center (SEDAC).
Dinkelman, Taryn. 2011. ‘The Effects of Rural Electrification on Employment: New Evidence
from South Africa’. American Economic Review 101 (7): 3078–3108.
Donaldson, Dave. 2018. ‘Railroads of the Raj: Estimating the Impact of Transportation
Infrastructure’. American Economic Review 108 (4–5): 899–934.
Easterly, William, and Ross Levine. 2003. ‘Tropics, Germs, and Crops: How Endowments
Influence Economic Development’. Journal of Monetary Economics 50 (1): 3–39.
Eberhard-Ruiz, Andreas, and Alexander Moradi. 2019. ‘Regional Market Integration in East
Africa: Local but No Regional Effects?’ Journal of Development Economics 140
(September): 255–68.
26
Engel, Jakob, Deeksha Kokas, Gladys López Acevedo, and Maryla Maliszewska. 2021. The
Distributional Impacts of Trade: Empirical Innovations, Analytical Tools, and Policy
Responses. Trade and Development. Washington, DC: World Bank.
FAO, and IIASA. 2019. ‘Global Agro Ecological Zones Version 4 (GAEZ V4)’.
Faye, Michael L., John W. McArthur, Jeffrey D. Sachs, and Thomas Snow. 2004. ‘The
Challenges Facing Landlocked Developing Countries’. Journal of Human Development
5 (1): 31–68.
Fick, Stephen E., and Robert J. Hijmans. 2017. ‘WorldClim 2: New 1‐km Spatial Resolution
Climate Surfaces for Global Land Areas’. International Journal of Climatology 37 (12):
4302–15.
Flückiger, Matthias, and Markus Ludwig. 2018. ‘Geography, Human Capital and Urbanization:
A Regional Analysis’. Economics Letters 168: 10–14.
Frankel, Jeffrey A, and David Romer. 1999. ‘Does Trade Cause Growth?’ American Economic
Review 89 (3): 379–99.
GADM. 2020. ‘Global Administrative Areas [Digital Geospatial Data]’. Center for Spatial
Sciences at the University of California, Davis.
Gallup, John Luke, Jeffrey D. Sachs, and Andrew D. Mellinger. 1999. ‘Geography and
Economic Development’. International Regional Science Review 22 (2): 179–232.
Gennaioli, Nicola, Rafael La Porta, Florencio Lopez-de-Silanes, and Andrei Shleifer. 2013.
‘Human Capital and Regional Development’. The Quarterly Journal of Economics 128
(1): 105–64.
Gollin, Douglas, Martina Kirchberger, and David Lagakos. 2017. ‘In Search of a Spatial
Equilibrium in the Developing World’. w23916. Cambridge, MA: National Bureau of
Economic Research.
Henderson, Vernon J., Tim Squires, Adam Storeygard, and David Weil. 2018. ‘The Global
Distribution of Economic Activity: Nature, History, and the Role of Trade’. The
Quarterly Journal of Economics 133 (1): 357–406.
Hirte, Georg, Christian Lessmann, and André Seidel. 2020. ‘International Trade, Geographic
Heterogeneity and Interregional Inequality’. European Economic Review 127 (August):
103427.
Jedwab, Remi, Edward Kerby, and Alexander Moradi. 2017. ‘History, Path Dependence and
Development: Evidence from Colonial Railways, Settlers and Cities In Kenya’. The
Economic Journal 127 (603): 1467–94.
Jedwab, Remi, and Alexander Moradi. 2016. ‘The Permanent Effects of Transportation
Revolutions in Poor Countries: Evidence from Africa’. Review of Economics and
Statistics 98 (2): 268–84.
Jedwab, Remi, and Adam Storeygard. 2020. ‘The Average and Heterogeneous Effects of
Transportation Investments: Evidence from Sub-Saharan Africa 1960-2010’. w27670.
Cambridge, MA: National Bureau of Economic Research.
Jetter, Michael, Saskia Mösle, and David Stadelmann. 2019. ‘Cursed by No Coast: How
Regional Landlockedness Affects Income within Countries’. Economics Letters 181: 70–
73.
Keats, Anthony. 2018. ‘Women’s Schooling, Fertility, and Child Health Outcomes: Evidence
from Uganda’s Free Primary Education Program’. Journal of Development Economics
135: 142–59.
Lehner, Bernhard, and Petra Döll. 2004. ‘Development and Validation of a Global Database
of Lakes, Reservoirs and Wetlands’. Journal of Hydrology 296 (1–4): 1–22.
Lessmann, Christian, and André Seidel. 2017. ‘Regional Inequality, Convergence, and Its
Determinants – A View from Outer Space’. European Economic Review 92 (February):
110–32.
27
Limão, Nuno, and Anthony J. Venables. 2001. ‘Infrastructure, Geographical Disadvantage,
Transport Costs, and Trade’, 29.
MacKinnon, David P., Amanda J. Fairchild, and Matthew S. Fritz. 2007. ‘Mediation Analysis’.
Annual Review of Psychology 58 (1): 593–614.
Michalopoulos, Stelios, and Elias Papaioannou. 2014. ‘National Institutions and Subnational
Development in Africa’. The Quarterly Journal of Economics 129 (1): 151–213.
Mitton, Todd. 2016. ‘The Wealth of Subnations: Geography, Institutions, and within-Country
Development’. Journal of Development Economics 118: 88–111.
Motamed, Mesbah J., Raymond J. G. M. Florax, and William A. Masters. 2014. ‘Agriculture,
Transportation and the Timing of Urbanization: Global Analysis at the Grid Cell
Level’. Journal of Economic Growth 19 (3): 339–68.
Natural Earth. 2019. ‘Natural Earth. Free Vector and Raster Map Data.’
NGA, Natioal Geospace-Intelligence Agency. 2019. ‘World Port Index 27th Edition’. United
States Government.
Nunn, Nathan, and Diego Puga. 2012. ‘Ruggedness: The Blessing of Bad Geography in Africa’.
Review of Economics and Statistics 94 (1): 20–36.
Nunn, Nathan, and Leonard Wantchekon. 2011. ‘The Slave Trade and the Origins of Mistrust
in Africa’. American Economic Review 101 (7): 3221–52.
Olson, David M., Eric Dinerstein, Eric D. Wikramanayake, Neil D. Burgess, George V. N.
Powell, Emma C. Underwood, Jennifer A. D’amico, et al. 2001. ‘Terrestrial Ecoregions
of the World: A New Map of Life on Earth’. BioScience 51 (11): 933.
Oster, Emily. 2019. ‘Unobservable Selection and Coefficient Stability: Theory and Evidence’.
Journal of Business & Economic Statistics 37 (2): 187–204.
Radelet, Steven, and Jeffrey D. Sachs. 1998. ‘Shipping Costs, Manufactured Exports, and
Economic Growth’. National Bureau of Economic Research Working Paper 19933.
Radeny, Maren, and Erwin Bulte. 2012. ‘Determinants of Rural Income: The Role of
Geography and Institutions in Kenya’. Journal of African Economies 21 (2): 307–41.
Ramankutty, Navin, Jonathan A. Foley, John Norman, and Kevin McSweeney. 2002. ‘The
Global Distribution of Cultivable Lands: Current Patterns and Sensitivity to Possible
Climate Change: Cropland Sensitivity to Climate Change’. Global Ecology and
Biogeography 11 (5): 377–92.
Rappaport, Jordan, and Jeffrey D. Sachs. 2003. ‘The United States as a Coastal Nation’.
Journal of Economic Growth 8: 5–46.
Rodrik, Dani, Arvind Subramanian, and Francesco Trebbi. 2004. ‘Institutions Rule: The
Primacy of Institutions Over Geography and Integration in Economic Development’.
Journal of Economic Growth 9 (2): 131–65.
Sachs, Sonia Ehrlich, Andrew Mellinger, Pia Malaney, Jeffrey Sachs, Anthony Kiszewski, and
Andrew Spielman. 2004. ‘A Global Index Representing the Stability of Malaria
Transmission’. The American Journal of Tropical Medicine and Hygiene 70 (5): 486–
98.
Sala-I-Martin, Xavier, Gernot Doppelhofer, and Ronald I. Miller. 2004. ‘Determinants of Long-
Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach’.
American Economic Review 94 (4): 813–35.
Sen, Amartya. 1985. Commodities and Capabilities. Lectures in Economics 7. Amsterdam New
York Oxford: North-Holland.
Skoufias, E., and R. S. Katayama. 2011. ‘Sources of Welfare Disparities between and within
Regions of Brazil: Evidence from the 2002-2003 Household Budget Survey (POF)’.
Journal of Economic Geography 11 (5): 897–918.
Smith, Adam. 1776. An Inquiry into the Nature and Causes of the Wealth of Nations. London:
W. Strahan and T. Cadell.
28
Smith, Adam, and R. H. Campbell. 2009. The Glasgow Edition of the Works and
Correspondence of Adam Smith. 2 1: An Inquiry into the Nature and Causes of the
Wealth of Nations / Adam Smith. General Editors R. H. Campbell. 11. print.
Indianapolis, IN: Liberty Fund.
Storeygard, Adam. 2016. ‘Farther on down the Road: Transport Costs, Trade and Urban
Growth in Sub-Saharan Africa’. The Review of Economic Studies 83 (3): 1263–95.
United Nations. 2001. ‘Road Map towards the Implementation of the United Nations
Millennium Declaration: Report of the Secretary-General.’ New York: United Nations
Department of Public Information.
UN-OHRLLS, United Nations Office of the High Representative for the Least Developed
Countries, Landlocked Developing Countries and Small Island Developing States. 2013.
‘The Development Economics of Landlockedness: Understanding the Development
Costs of Being Landlocked’. New York, NY: United Nations - OHRLLS.
Wild, Frederik, and David Stadelmann. 2022. ‘Coastal Proximity and Individual Living
Standards: Econometric Evidence from Georeferenced Household Surveys in sub‐
Saharan Africa’. Review of Development Economics 26 (4): 1883–1901.
World Bank. 2009. World Development Report 2009: Reshaping Economic Geography. 2009.
Washington, D.C.: World Bank.
———. 2022. ‘Data.Worldbank.Org’. 2022.
Young, Alwyn. 2013. ‘Inequality, the Urban-Rural Gap, and Migration’. The Quarterly Journal
of Economics 128 (4): 1727–85.
29
Appendix A
Coastal
Proximity
and
Individual
Living
Standards: Econometric Evidence from Geo-
Referenced Household Surveys in Sub-Saharan
Africa
30
Figure A.1: Distance Bins
Notes: The plot depicts point estimates as well as their corresponding 95% confidence intervals produced from
three separate regressions of the three main outcome variables - distinguished by black, grey and dashed-grey
figures - on all harbor distance increments shown on the X-axis (the omitted category is 0-50 km), and including
country-sample fixed effects. Therefore, coefficients are interpreted as the average change in the outcome variable
for individuals living within the distance increments to living within 50km to the harbor. Results are produced
using the main sample of coastal, sub-Saharan African countries from round 1 through round 7 of the geo-coded
Afrobarometer surveys, except for the dashed estimates (Possessions), given that this question is available only
from round 3 and onwards. Binary dependent variables are estimated through a simple LPM (Linear Probability
Model) specification. Standard errors are clustered at the survey-enumeration level.
31
Figure A.2: Structure of Mediation Analysis
Notes: The figure illustrates the envisioned mediation structure discussed in the manuscript. The estimate
depicted represents the results of column (1) of Table 2.5. ***, **, * represents significance at the 1, 5 and 10
percent level, respectively.
32
Table A.1: Coastal Proximity and Individual Living Standards: Full Table
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.018*** -0.009***
0.073***
0.035***
-0.037***
-0.019***
(0.001)
(0.002)
(0.004)
(0.006)
(0.001)
(0.002)
Discrete Change of Distance from
Harbor to the 3rd Quartile (564km)
-0.115
-0.055
0.460
0.223
-0.234
-0.120
Sample Mean of Dep. Var.
Basic Controls
Age
0.031***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.106***
0.020***
-0.083***
(0.003)
(0.004)
(0.002)
Urbanization Controls
Urban {0,1}
0.055***
-0.284***
0.134***
(0.004)
(0.010)
(0.003)
Primate City {0,1}
0.029***
-0.084***
0.029***
(0.007)
(0.016)
(0.006)
Population Density
-0.001
0.004***
-0.001*
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.013
-0.121***
0.029***
(0.009)
(0.021)
(0.008)
Major Lake {0,1}
0.010
0.033
-0.023***
(0.012)
(0.024)
(0.008)
Full Geographic Controls
Abs. Latitude
-0.002
-0.015***
0.001
(0.001)
(0.002)
(0.001)
Elevation (km)
-0.016
-0.115***
0.038***
(0.013)
(0.028)
(0.012)
Ruggedness (Standardized)
-0.004
0.015*
-0.006**
(0.004)
(0.008)
(0.003)
Land Suitability [0,1]
0.021*
0.189***
0.011
(0.012)
(0.027)
(0.009)
Monthly Temperature (Celsius)
-0.009***
0.020***
-0.001
(0.002)
(0.005)
(0.002)
Monthly Rainfall (Standardized)
-0.003
0.003
-0.001
(0.004)
(0.009)
(0.003)
Growing Days {0,365}
0.000
-0.001***
0.000**
(0.000)
(0.000)
(0.000)
Malaria Index
0.000
0.004***
-0.001***
(0.000)
(0.001)
(0.000)
Mediterraenean {0,1}
0.014
0.038
0.038
(0.037)
(0.084)
(0.031)
Desert {0,1}
0.028
0.170**
0.024
(0.035)
(0.077)
(0.030)
Dependent Variable
[0.39]
[1.28]
[0.51]
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
Notes: Table continued on next page.
33
Mangroves {0,1}
0.017
0.129*
-0.005
(0.034)
(0.074)
(0.028)
Tropical Forest {0,1}
0.041
0.037
0.025
(0.032)
(0.072)
(0.027)
Tropical Grassland {0,1}
0.029
0.070
0.017
(0.032)
(0.071)
(0.027)
Temperate Grassland {0,1}
0.046
0.146*
0.016
(0.039)
(0.089)
(0.032)
Montane Grassland {0,1}
0.032
0.229***
-0.007
(0.034)
(0.076)
(0.029)
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,793
122,238
128,609
126,982
103,889
102,990
R-Squared
0.09
0.14
0.17
0.20
0.15
0.22
Notes: Results in each column come from separate regressions and are estimated using the main
sample of coastal, sub-Saharan African countries included in survey rounds 1 through 7 of the
Afrobarometer. Changes in the number of observations across columns stem from differences in
the response rates of dependent/independent variables. The sample used in columns (5) and (6)
do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on
ownership of household items were not asked in these rounds. We also report estimated
interquartile differences in the respective dependent variables between minimum-, and 3rd quartile
harbor distances within the sample. Binary dependent variables are estimated through a simple
LPM (Linear Probability Model) specification. The standard errors reported are clustered at the
survey enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively. See full notes below Table 2.2.
34
Table A.2: De-Composed Indices
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
log(1+Distance to Harbor)
0.095***
0.051***
0.069***
0.015*
0.036***
0.023***
0.096***
0.053***
-0.019***
-0.017*** -0.085***
-0.039***
-0.008***
0.000
(0.004)
(0.007)
(0.005)
(0.009)
(0.004)
(0.007)
(0.004)
(0.007)
(0.001)
(0.002)
(0.002)
(0.003)
(0.001)
(0.002)
Discrete Change of Distance from
Harbor to the 3rd Quartile (564km)
0.599
0.322
0.433
0.093
0.230
0.143
0.604
0.334
-0.121
-0.109
-0.535
-0.248
-0.050
-0.003
Sample Mean of Dep. Var.
Basic Controls
Age
0.007***
0.010***
0.010***
0.015***
0.014***
0.011***
0.010***
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
0.000***
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
Female {0,1}
0.043***
-0.002
0.041***
0.002
-0.119***
-0.038***
-0.091***
(0.006)
(0.006)
(0.006)
(0.006)
(0.003)
(0.002)
(0.003)
Urbanization Controls
Urban {0,1}
-0.287***
-0.298***
-0.214***
-0.350***
0.062***
0.259***
0.083***
(0.012)
(0.015)
(0.011)
(0.012)
(0.004)
(0.006)
(0.004)
Primate City {0,1}
-0.144***
-0.069***
-0.045***
-0.079***
0.004
0.051***
0.033***
(0.021)
(0.023)
(0.016)
(0.018)
(0.006)
(0.009)
(0.008)
Population Density
0.007***
0.002*
0.004***
0.004***
-0.001*
0.002***
-0.003***
(0.001)
(0.001)
(0.001)
(0.001)
(0.000)
(0.001)
(0.001)
Trade-related Controls
Navigable River {0,1}
-0.125***
-0.137***
-0.152***
-0.075***
0.021**
0.044***
0.022***
(0.023)
(0.030)
(0.024)
(0.025)
(0.009)
(0.011)
(0.008)
Major Lake {0,1}
0.060*
-0.001
0.053*
0.031
-0.008
-0.040***
-0.017**
(0.032)
(0.039)
(0.027)
(0.029)
(0.010)
(0.013)
(0.009)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Observations
122,363
120,792
128,427
126,808
128,356
126,761
128,105
126,508
103,838
102,940
103,582
102,687
103,388
102,493
R-Squared
0.17
0.19
0.09
0.12
0.07
0.09
0.11
0.14
0.04
0.08
0.23
0.30
0.23
0.25
Possessions:
(Motor Vehicle)
{0,1}
[0.48]
[0.30]
Dependent Variable
Possessions:
(Radio)
{0,1}
How often gone
without:
(Medical Care)
{0,4}
[1.15]
[0.75]
How often gone
without:
(Cash Income)
{0,4}
How often gone
without:
(Water)
{0,4}
How often gone
without:
(Food)
{0,4}
[1.90]
[1.14]
[0.97]
Possessions:
(TV)
{0,1}
Notes: This table reports results for all items of our two main composite living-standards indices tested in Table 2.2. Results in each column come from separate regressions and are
estimated using the main sample of coastal, sub-Saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across
columns stem from differences in the response rates of dependent/independent variables. The sample used in columns (9) through (14) do not include individuals surveyed in rounds
1 and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these rounds. We also report estimated interquartile differences in the respective
dependent variables between minimum-, and 3rd quartile harbor distances within the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability
Model) specification. The standard errors reported are clustered at the survey enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
35
Table A.3: Summary of Robustness Tests
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficient
-0.018***
-0.009***
0.073***
0.035***
-0.037***
-0.019***
(0.001)
(0.002)
(0.004)
(0.006)
(0.001)
(0.002)
(a) log(1+Beeline Distance to Harbor)
-0.019***
-0.011***
0.071***
0.033***
-0.037***
-0.019***
(0.001)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
(b) log(1+Distance to Coastline )
-0.014***
-0.008***
0.050***
0.031***
-0.026***
-0.010***
(0.001)
(0.002)
(0.003)
(0.005)
(0.001)
(0.002)
(c) log(1+Beeline Distance to Coastline) -0.014***
-0.011***
0.048***
0.027***
-0.026***
-0.011***
(0.001)
(0.002)
(0.003)
(0.005)
(0.001)
(0.002)
(d) Including Harbor ≤ 25km Dummy
-0.028***
-0.021***
0.038***
0.037***
-0.028***
-0.023***
(0.002)
(0.003)
(0.006)
(0.007)
(0.002)
(0.002)
(e) Including Coast ≤ 25km Dummy
-0.019***
-0.010***
0.060***
0.031***
-0.033***
-0.018***
(0.002)
(0.003)
(0.005)
(0.006)
(0.002)
(0.002)
(f) Keeping Observations Constant
-0.019***
-0.008***
0.074***
0.032***
-0.037***
-0.019***
(0.001)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
(g) Excluding Distances > 629 km
-0.015***
-0.005**
0.074***
0.039***
-0.035***
-0.016***
(0.002)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
(h) Excluding Low-Precision Localities
-0.018***
-0.011***
0.062***
0.029***
-0.034***
-0.018***
(0.002)
(0.003)
(0.005)
(0.008)
(0.002)
(0.003)
(i) Including Survey Weights
-0.018***
-0.006**
0.070***
0.029***
-0.037***
-0.017***
(0.002)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
(j) Country-Sample Clustering
-0.018***
-0.009*
0.073***
0.035***
-0.037***
-0.019***
(0.003)
(0.005)
(0.008)
(0.009)
(0.003)
(0.003)
Basic Controls
NO
YES
NO
YES
NO
YES
Urbanization Controls
NO
YES
NO
YES
NO
YES
Trade-related Controls
NO
YES
NO
YES
NO
YES
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
(see Table A.10)
(see Table A.12)
(see Table A.6)
(see Table A.11)
Coefficient on log(1+Distance to Harbor) - if not indicated otherwise
(see Table A.4)
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
(see Table A.5)
(see Table A.7)
(see Table A.7)
(see Table A.8)
(see Table A.9)
Notes: This table summarizes the robustness checks on our main findings. Results in each row and column come from separate
regressions. a) Replaces the main explanatory variable of (logged) within-country distance used in Table 2.2 with a simple
beeline distance to the nearest harbor. b) Replaces the main explanatory variable with a within-country distance to the
country’s coastline. (c) Replaces the main explanatory variable with a beeline distance to the coastline. (d) Adds a dummy
indicator of living within 25 kilometers to a major harbor to our baseline specification. (e) Includes a dummy indicator of
living within 25 kilometers to the coastline to our baseline specification. (f) Estimates the baseline specification but holds the
number of observations constant across all columns, using the maximum number of individuals for which all in- and dependent
variables are available. Note that this drops rounds 1 and 2 of the Afrobarometer entirely, given that questions on ownership
of household items were not asked. (g) Estimates the baseline specification but excludes distances larger than the 80th
percentile. (h) Estimates the baseline specification but excludes localities marked with a precision code of 2 or larger (scale 1-
8). (i) Estimates the baseline specification but uses the included Afrobarometer survey weights. (j) Estimates the baseline
specification but implements standard error clustering at the country-sample level instead of the survey enumeration area
level. Binary dependent variables are estimated through a simple LPM (Linear Probability Model) specification. The standard
errors reported are clustered at the survey enumeration area level, if not indicated otherwise indicated. ***, **, * represents
significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 2.2.
36
Table A.4: Robustness Test: “beeline” Distance to Harbor
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Beeline Distance to Harbor) -0.019***
-0.011***
0.071***
0.033***
-0.037***
-0.019***
(0.001)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
Discrete Change of Distance
from Harbor to the 3rd
Quartile (506km)
-0.117
-0.067
0.440
0.204
-0.229
-0.115
Sample Mean of Dep. Var.
Basic Controls
Age
0.031***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.106***
0.020***
-0.083***
(0.003)
(0.004)
(0.002)
Urbanization Controls
Urban {0,1}
0.055***
-0.285***
0.135***
(0.004)
(0.010)
(0.003)
Primate City {0,1}
0.026***
-0.087***
0.029***
(0.007)
(0.016)
(0.006)
Population Density
-0.001
0.004***
-0.001
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.012
-0.125***
0.031***
(0.009)
(0.020)
(0.008)
Major Lake {0,1}
0.011
0.033
-0.023***
(0.012)
(0.024)
(0.008)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,793
122,238
128,609
126,982
103,889
102,990
R-Squared
0.10
0.14
0.17
0.20
0.15
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.28]
[0.51]
Notes: This table is equivalent to Table 2.2 but exchanges the within-country distance to harbors
used in Table 1 with a beeline (as the crow flies) distance to the nearest harbor. Results in each
column come from separate regressions and are estimated using the main sample of coastal, sub-
Saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes
in the number of observations across columns stem from differences in the response rates of
dependent/independent variables. The sample used in columns (5) and (6) do not include
individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of
household items were not asked in these rounds. We also report estimated interquartile differences
in the respective dependent variables between minimum-, and 3rd quartile harbor distances within
the sample. Binary dependent variables are estimated through a simple LPM (Linear Probability
Model) specification. The standard errors reported are clustered at the survey enumeration area
level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes
below Table 2.2.
37
Table A.5: Robustness Test: Distance to Coastline
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Coastline)
-0.014***
-0.008***
0.050***
0.031***
-0.026***
-0.010***
(0.001)
(0.002)
(0.003)
(0.005)
(0.001)
(0.002)
Discrete Change of Distance
from the Coast to the 3rd
Quartile (468km)
-0.085
-0.047
0.303
0.188
-0.161
-0.063
Sample Mean of Dep. Var.
Basic Controls
Age
0.031***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.106***
0.020***
-0.082***
(0.003)
(0.004)
(0.002)
Urbanization Controls
Urban {0,1}
0.055***
-0.285***
0.136***
(0.004)
(0.010)
(0.003)
Primate City {0,1}
0.033***
-0.102***
0.043***
(0.007)
(0.014)
(0.005)
Population Density
-0.001
0.004***
0.000
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.011
-0.130***
0.032***
(0.009)
(0.021)
(0.008)
Major Lake {0,1}
0.012
0.026
-0.023***
(0.012)
(0.024)
(0.008)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,793
122,238
128,609
126,982
103,889
102,990
R-Squared
0.09
0.14
0.16
0.20
0.15
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.28]
[0.51]
Notes: This table is equivalent to Table 2.2, but exchanges the within-country distance to harbors
used in Table 1 with within-country distance to the coastline. Results in each column come from
separate regressions and are estimated using the main sample of coastal, sub-Saharan African
countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of
observations across columns stem from differences in the response rates of dependent/independent
variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1
and 2 of the Afrobarometer, as questions on ownership of household items were not asked in these
rounds. We also report estimated interquartile differences in the respective dependent variables
between minimum-, and 3rd quartile harbor distances within the sample. Binary dependent variables
are estimated through a simple LPM (Linear Probability Model) specification. The standard errors
reported are clustered at the survey enumeration area level. ***, **, * represents significance at the
1, 5 and 10 percent level, respectively. See full notes below Table 2.2.
38
Table A.6: Robustness Test: “beeline” Distance to Coastline
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Beeline Distance to Coastline) -0.014***
-0.011***
0.048***
0.027***
-0.026***
-0.011***
(0.001)
(0.002)
(0.003)
(0.005)
(0.001)
(0.002)
Discrete Change of Distance from
the Coast to the 3rd Quartile
(426km)
-0.087
-0.063
0.288
0.162
-0.158
-0.063
Sample Mean of Dep. Var.
Basic Controls
Age
0.031***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.106***
0.020***
-0.082***
(0.003)
(0.004)
(0.002)
Urbanization Controls
Urban {0,1}
0.055***
-0.286***
0.136***
(0.004)
(0.010)
(0.003)
Primate City {0,1}
0.031***
-0.105***
0.043***
(0.007)
(0.015)
(0.005)
Population Density
-0.001
0.004***
0.000
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.010
-0.130***
0.033***
(0.009)
(0.021)
(0.008)
Major Lake {0,1}
0.013
0.028
-0.023***
(0.012)
(0.024)
(0.008)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,793
122,238
128,609
126,982
103,889
102,990
R-Squared
0.09
0.14
0.16
0.20
0.15
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.28]
[0.51]
Notes: This table is equivalent to Table 2.2, but exchanges the within-country distance to harbors
used in Table 1 with a beeline (as the crow flies) distance to the coastline. Results in each column
come from separate regressions and are estimated using the main sample of coastal, sub-Saharan
African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in the
number of observations across columns stem from differences in the response rates of
dependent/independent variables. The sample used in columns (5) and (6) do not include individuals
surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household items were
not asked in these rounds. We also report estimated interquartile differences in the respective
dependent variables between minimum-, and 3rd quartile harbor distances within the sample. Binary
dependent variables are estimated through a simple LPM (Linear Probability Model) specification.
The standard errors reported are clustered at the survey enumeration area level. ***, **, * represents
significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 2.2.
39
Table A.7: Robustness Test: Adding Harbor/Coastal Dummies
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.021***
-0.010***
0.037***
0.031***
-0.023***
-0.018***
(0.003)
(0.003)
(0.007)
(0.006)
(0.002)
(0.002)
Discrete Change of Distance
from Harbor to the 3rd
Quartile (564km)
-0.133
-0.064
0.232
0.195
-0.147
-0.116
Sample Mean of Dependent Var.
[0.39]
[1.28]
[0.51]
Trade-related Controls
Harbor {0,1}
-0.071***
0.008
-0.025***
(0.010)
(0.025)
(0.008)
Coast {0,1}
-0.013*
-0.039**
0.006
(0.008)
(0.018)
(0.006)
Navigable River {0,1}
-0.013
-0.013
-0.121***
-0.122***
0.029***
0.029***
(0.009)
(0.009)
(0.021)
(0.020)
(0.008)
(0.008)
Major Lake {0,1}
0.016
0.010
0.032
0.032
-0.021***
-0.023***
(0.012)
(0.012)
(0.024)
(0.024)
(0.008)
(0.008)
Urbanization Controls
Urban {0,1}
0.056***
0.055***
-0.285***
-0.283***
0.135***
0.134***
(0.004)
0.0043
(0.010)
0.0099
(0.003)
0.0035
Primate City {0,1}
0.044***
0.031***
-0.085***
-0.078***
0.034***
0.028***
(0.008)
(0.007)
(0.016)
(0.016)
(0.006)
(0.006)
Population Density
0.000
-0.001
0.004***
0.004***
-0.001*
-0.001*
(0.000)
(0.000)
(0.001)
(0.001)
(0.000)
(0.000)
Basic Controls
Age
0.031***
0.031***
0.011***
0.011***
0.012***
0.012***
(0.001)
0.0006
(0.001)
0.0009
(0.000)
0.0004
Age2
0.000***
0.000***
0.000***
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.106***
-0.106***
0.020***
0.020***
-0.083***
-0.083***
(0.003)
(0.003)
(0.004)
(0.004)
(0.002)
(0.002)
Full Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
122,238
122,238
126,982
126,982
102,990
102,990
R-Squared
0.15
0.14
0.20
0.20
0.22
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
Notes: This table is equivalent to Table 2.2, but adds dummies indicating individuals living within
25km of the harbor or coast, in turn. We also report estimated interquartile differences in the
respective dependent variables between minimum-, and 3rd quartile harbor distances within the
sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model)
specification. The standard errors reported are clustered at the survey enumeration area level. ***,
**, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table
2.2.
40
Table A.8 Robustness Test: Constant Observations
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.019***
-0.008***
0.074***
0.032***
-0.037***
-0.019***
(0.001)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
Discrete Change of Distance
from Harbor to the 3rd
Quartile (547km)
-0.118
-0.049
0.463
0.201
-0.232
-0.118
Sample Mean of Dep. Var.
Basic Controls
Age
0.035***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.114***
0.022***
-0.083***
(0.003)
(0.005)
(0.002)
Urbanization Controls
Urban {0,1}
0.060***
-0.288***
0.135***
(0.005)
(0.011)
(0.003)
Primate City {0,1}
0.030***
-0.076***
0.029***
(0.007)
(0.018)
(0.006)
Population Density
-0.001**
0.003***
-0.001*
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.018*
-0.143***
0.029***
(0.010)
(0.022)
(0.008)
Major Lake {0,1}
-0.001
0.030***
-0.023***
(0.012)
(0.003)
(0.008)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
102,460
102,460
102,460
102,460
102,460
102,460
R-Squared
0.11
0.17
0.17
0.20
0.15
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.33]
[0.51]
Notes: This table is equivalent to Table 2.2, but holds the number of observations constant across
all columns, using the maximum number of individuals for which all independent and dependent
variables are available. Note that this drops rounds 1 and 2 of the Afrobarometer entirely, given that
questions on ownership of household items were not asked in these rounds. Therefore, the sample
used is comprised of coastal, sub-Saharan African countries included in survey rounds 3 through 7
of the Afrobarometer. Results in each column come from separate regressions. We also report
estimated interquartile differences in the respective dependent variables between minimum-, and 3rd
quartile harbor distances within the sample. Binary dependent variables are estimated through a
simple LPM (Linear Probability Model) specification. The standard errors reported are clustered at
the country-sample level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively. See full notes below Table 2.2.
41
Table A.9: Robustness Test: Excluding Large Distances
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.015***
-0.005**
0.074***
0.039***
-0.035***
-0.016***
(0.002)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
Discrete Change of Distance
from Harbor to the 3rd
Quartile (410km)
-0.093
-0.032
0.444
0.232
-0.207
-0.097
Sample Mean of Dep. Var.
Basic Controls
Age
0.032***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.101***
0.027***
-0.087***
(0.003)
(0.005)
(0.002)
Urbanization Controls
Urban {0,1}
0.050***
-0.296***
0.131***
(0.005)
(0.011)
(0.004)
Primate City {0,1}
0.033***
-0.076***
0.033***
(0.007)
(0.016)
(0.006)
Population Density
-0.001**
0.004***
-0.001**
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.008
-0.108***
0.027***
(0.010)
(0.024)
(0.008)
Major Lake {0,1}
0.042**
0.015***
-0.019
(0.019)
(0.003)
(0.016)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
99,302
98,269
103,029
101,945
84,589
83,929
R-Squared
0.10
0.15
0.18
0.22
0.15
0.21
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.27]
[0.53]
Notes: This table is equivalent to Table 2.2, but excludes distances larger than the 80th percentile.
Results in each column come from separate regressions and are estimated using the main sample of
coastal, sub-Saharan African countries included in survey rounds 1 through 7 of the Afrobarometer.
Changes in the number of observations across columns stem from differences in the response rates
of dependent/independent variables. The sample used in columns (5) and (6) do not include
individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household
items were not asked in these rounds. We also report estimated interquartile differences in the
respective dependent variables between minimum-, and 3rd quartile harbor distances within the
sample. Binary dependent variables are estimated through a simple LPM (Linear Probability Model)
specification. The standard errors reported are clustered at the country-sample level. ***, **, *
represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 2.2.
42
Table A.10: Robustness Test: Excluding Low-Precision Geo-clusters
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.018***
-0.011***
0.062***
0.029***
-0.034***
-0.018***
(0.002)
(0.003)
(0.005)
(0.008)
(0.002)
(0.003)
Discrete Change of Distance
from Harbor to the 3rd
Quartile (542km)
-0.111
-0.068
0.374
0.173
-0.204
-0.108
Sample Mean of Dep. Var.
Basic Controls
Age
0.034***
0.011***
0.017***
(0.001)
(0.001)
(0.001)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.105***
0.022***
-0.088***
(0.004)
(0.006)
(0.003)
Urbanization Controls
Urban {0,1}
0.046***
-0.291***
0.124***
(0.006)
(0.014)
(0.005)
Primate City {0,1}
0.035***
-0.117***
0.044***
(0.010)
(0.021)
(0.007)
Population Density
-0.001
0.006***
-0.001**
(0.001)
(0.001)
(0.001)
Trade-related Controls
Navigable River {0,1}
-0.012
-0.110***
0.031***
(0.012)
(0.027)
(0.010)
Major Lake {0,1}
-0.002
0.002
-0.025**
(0.018)
(0.003)
(0.012)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
61,731
60,869
64,023
63,122
49,067
48,594
R-Squared
0.08
0.14
0.18
0.21
0.15
0.21
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.40]
[1.21]
[0.51]
Notes: This table is equivalent to Table 2.2, but excludes localities marked with a precision code
of 2 or larger (scale 1-8). Results in each column come from separate regressions and are estimated
using the main sample of coastal, sub-Saharan African countries included in survey rounds 1
through 7 of the Afrobarometer. Changes in the number of observations across columns stem from
differences in the response rates of dependent/independent variables. The sample used in columns
(5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions
on ownership of household items were not asked in these rounds. We also report estimated
interquartile differences in the respective dependent variables between minimum-, and 3rd quartile
harbor distances within the sample. Binary dependent variables are estimated through a simple
LPM (Linear Probability Model) specification. The standard errors reported are clustered at the
country-sample level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
See full notes below Table 2.2.
43
Table A.11: Robustness Test: Including Survey Weights
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.018***
-0.006**
0.070***
0.029***
-0.037***
-0.017***
(0.002)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
Discrete Change of Distance
from Harbor to the 3rd
Quartile (425km)
-0.112
-0.039
0.446
0.186
-0.232
-0.106
Sample Mean of Dep. Var.
Basic Controls
Age
0.031***
0.011***
0.012***
(0.001)
(0.001)
(0.000)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.105***
0.021***
-0.083***
(0.003)
(0.005)
(0.002)
Urbanization Controls
Urban {0,1}
0.055***
-0.281***
0.135***
(0.005)
(0.011)
(0.004)
Primate City {0,1}
0.030***
-0.090***
0.030***
(0.008)
(0.017)
(0.006)
Population Density
0.000
0.004***
-0.001
(0.000)
(0.001)
(0.000)
Trade-related Controls
Navigable River {0,1}
-0.009
-0.111***
0.027***
(0.010)
(0.022)
(0.008)
Major Lake {0,1}
0.019
0.023
-0.020**
(0.013)
(0.025)
(0.009)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,793
122,238
128,608
126,982
103,889
102,990
R-Squared
0.10
0.15
0.16
0.20
0.15
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.28]
[0.51]
Notes: This table is equivalent to Table 2.2, but uses the included Afrobarometer survey weights to
produce the estimates. Results in each column come from separate regressions and are estimated
using the main sample of coastal, sub-Saharan African countries included in survey rounds 1
through 7 of the Afrobarometer. Changes in the number of observations across columns stem from
differences in the response rates of dependent/independent variables. The sample used in columns
(5) and (6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions
on ownership of household items were not asked in these rounds. We also report estimated
interquartile differences in the respective dependent variables between minimum-, and 3rd quartile
harbor distances within the sample. Binary dependent variables are estimated through a simple
LPM (Linear Probability Model) specification. The standard errors reported are clustered at the
survey enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively. See full notes below Table 2.2.
44
Table A.12: Robustness Test: Country-Sample Clustering
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.018***
-0.009*
0.073***
0.035***
-0.037***
-0.019***
(0.003)
(0.005)
(0.008)
(0.009)
(0.003)
(0.003)
Discrete Change of Distance
from Harbor to the 3rd
Quartile (564km)
-0.115
-0.055
0.460
0.223
-0.234
-0.120
Sample Mean of Dep. Var.
Basic Controls
Age
0.031***
0.011***
0.012***
(0.002)
(0.001)
(0.001)
Age2
0.000***
0.000***
0.000***
(0.000)
(0.000)
(0.000)
Female {0,1}
-0.106***
0.020***
-0.083***
(0.007)
(0.005)
(0.006)
Urbanization Controls
Urban {0,1}
0.055***
-0.284***
0.134***
(0.007)
(0.018)
(0.007)
Primate City {0,1}
0.029**
-0.084***
0.029***
(0.011)
(0.027)
(0.009)
Population Density
-0.001
0.004***
-0.001
(0.001)
(0.001)
(0.001)
Trade-related Controls
Navigable River {0,1}
-0.013
-0.121***
0.029***
(0.012)
(0.026)
(0.010)
Major Lake {0,1}
0.010
0.033
-0.023**
(0.020)
(0.036)
(0.010)
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,793
122,238
128,609
126,982
103,889
102,990
R-Squared
0.09
0.14
0.17
0.20
0.15
0.22
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
[0.39]
[1.28]
[0.51]
Notes: This table is equivalent to Table 2.2, but implements standard error clustering at the country-
sample level instead of the survey enumeration area level. Results in each column come from separate
regressions and are estimated using the main sample of coastal, sub-Saharan African countries
included in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations
across columns stem from differences in the response rates of dependent/independent variables. The
sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the
Afrobarometer, as questions on ownership of household items were not asked in these rounds. We
also report estimated interquartile differences in the respective dependent variables between
minimum-, and 3rd quartile harbor distances within the sample. Binary dependent variables are
estimated through a simple LPM (Linear Probability Model) specification. The standard errors
reported are clustered at the country-sample level. ***, **, * represents significance at the 1, 5 and
10 percent level, respectively. See full notes below Table 2.2.
45
Table A.13: Robustness Test: Oster (2019) Tests
Explanatory Variable: log(1+Distance to Harbor), Specification: δ=0.5, Rmax=1
Independent Variable
β°
Baseline Effect
β'
Controlled Effect
R°2
Baseline
R'2
Controlled
[β', β*]
Identified Set
(1)
(2)
(3)
(4)
(5)
Cash Employment
-0.003
-0.009***
[0.00]
[0.14]
[-0.027 , -0.009]
{0,1}
(0.002)
(0.002)
How often gone without:
(Water / Food / Cash Inc. / Med.)
0.025***
0.035***
[0.00]
[0.20]
[0.035 , 0.055]
[0,4]
(0.006)
(0.006)
Possessions
(Radio / TV / Motor Vehicle)
-0.030***
-0.019***
[0.02]
[0.22]
[-0.019 , 0.002]
[0,1]
(0.002)
(0.002)
Basic Controls
NO
YES
NO
YES
YES
Urbanization Controls
NO
YES
NO
YES
YES
Trade-related Controls
NO
YES
NO
YES
YES
Full Geographic Controls
NO
YES
NO
YES
YES
Country-Time FE
NO
YES
NO
YES
YES
s
Notes: This table presents result from a formal analysis of coefficient stability and influence of unobservables
according to Oster (2019), analyzing changes in the estimate of our main explanatory variable "log(1+Distance to
Harbor)" when adding the full set of controls as well as fixed-effects, using our three main outcome variables.
Columns (1) and (2) present the uncontrolled β°, as well as the controlled β' and columns (3) and (4) depict their
respective regression’s R-Squared. Column (5) shows the lower- and upper bound estimate of the identified set
assuming Rmax = 1 and δ = 0.5. The bias-adjusted upper bound is calculated using β*= β' - δ((β° - β')*(Rmax
- R'2))/(R'2 - R°2)), the lower bound is given by β'. Results are estimated using the main sample of coastal, sub-
Saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. The sample used row 3 do
not include individuals surveyed in Rounds 1 and 2 of the Afrobarometer, as questions on ownership of household
items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear
Probability Model) specification. The standard errors reported are clustered at the survey-enumeration area level.
***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
46
Table A.14: Extensions: “Landlockedness”
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.018***
0.003
0.073***
0.017***
-0.037*** -0.009***
(0.001)
(0.002)
(0.004)
(0.005)
(0.001)
(0.002)
log(1+Distance to Harbor) × Landlocked {0,1}
0.030**
0.019
0.026
0.057*
0.025*
0.005
(0.013)
(0.013)
(0.039)
(0.033)
(0.014)
(0.010)
Isolated Effect of the Distance to Harbor in Landlocked Countries
log(1+Distance to Harbor) + Interaction
0.011
0.022*
0.099**
0.074**
-0.012
-0.005
[0.39]
[0.08]
[0.01]
[0.03]
[0.40]
[0.65]
Basic Controls
NO
YES
NO
YES
NO
YES
Urbanization Controls
NO
YES
NO
YES
NO
YES
Trade-related Controls
NO
YES
NO
YES
NO
YES
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
204,717
200,128
212,037
207,211
169,590
166,242
R-Squared
0.11
0.15
0.15
0.19
0.16
0.24
Interaction:
Combined Effect:
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
Notes: We analyze the differential effect of our main explanatory variable for individuals living in landlocked
countries. Row one shows the uninteracted effect of log(1+Distance to Harbor), i.e. the distance effect of individuals
in coastal countries, row two shows the differential effect for being landlocked. Row three depicts the combined
effect of the two constituent terms, i.e. the effect of log(1+Distance to Harbor) for individuals living in landlocked
countries together with the corresponding p-value in brackets. Results in each column come from separate
regressions and are estimated using the sample of coastal and landlocked sub-Saharan African countries included
in survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem
from differences in the response rates of dependent/independent variables. The sample used in columns (5) and
(6) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of
household items were not asked in this round. Binary dependent variables are estimated through a simple LPM
(Linear Probability Model) specification. The standard errors reported are clustered at the survey enumeration
area level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
47
Table A.15: Extension: Generations (Young) Interaction
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.023*** -0.014***
0.069***
0.031***
-0.039*** -0.021***
(0.002)
(0.003)
(0.004)
(0.006)
(0.001)
(0.002)
log(1+Distance to Harbor) × Young {0,1}
0.008***
0.010***
0.007*
0.008**
0.003*
0.004***
(0.002)
(0.002)
(0.003)
(0.003)
(0.002)
(0.002)
Isolated Effect of the Distance to Harbor for Individuals below median Age (33)
log(1+Distance to Harbor) + Interaction
-0.015***
-0.004
0.075***
0.039***
-0.036*** -0.017***
[0.00]
[0.11]
[0.00]
[0.00]
[0.00]
[0.00]
Basic Controls
NO
YES
NO
YES
NO
YES
Urbanization Controls
NO
YES
NO
YES
NO
YES
Trade-related Controls
NO
YES
NO
YES
NO
YES
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
122,555
122,238
127,305
126,982
103,083
102,990
R-Squared
0.10
0.15
0.17
0.20
0.16
0.22
Combined Effect:
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
Interaction:
Notes: We analyze the differential effect of our main explanatory variable for individuals above and below the
median age within the sample. Row one shows the uninteracted effect of log(1+Distance to Harbor), i.e. the
distance effect of individuals at and above the median age (33), row two shows the differential effect for being in
the younger strata. Row three depicts the combined effect of the two constituent terms, i.e. the effect of
log(1+Distance to Harbor) for individuals younger than the median age (33) together with the corresponding p-
value in brackets. Results in each column come from separate regressions and are estimated using the sample of
coastal, sub-Saharan African countries included in survey rounds 1 through 7 of the Afrobarometer. Changes in
the number of observations across columns stem from differences in the response rates of dependent/independent
variables. The sample used in columns (5) and (6) do not include individuals surveyed in rounds 1 and 2 of the
Afrobarometer, as questions on ownership of household items were not asked in this round. Binary dependent
variables are estimated through a simple LPM (Linear Probability Model) specification. The standard errors
reported are clustered at the survey enumeration area level. ***, **, * represents significance at the 1, 5 and 10
percent level, respectively.
48
Table A.16: Extension: “Remote Urbanities”
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.023*** -0.025***
0.026***
0.039***
-0.027*** -0.029***
(0.003)
(0.003)
(0.007)
(0.008)
(0.002)
(0.003)
log(1+Distance to Harbor) × Urban {0,1}
0.016***
0.022***
0.010
-0.004
0.010***
0.014***
(0.003)
(0.003)
(0.007)
(0.008)
(0.002)
(0.002)
Isolated Effect of the Distance to Harbor for Individuals living in Urban Environments
log(1+Distance to Harbor) + Interaction
-0.007***
-0.003
0.036***
0.034***
-0.016*** -0.015***
[0.00]
[0.23]
[0.00]
[0.00]
[0.00]
[0.00]
Basic Controls
NO
YES
NO
YES
NO
YES
Urbanization Controls
NO
YES
NO
YES
NO
YES
Trade-related Controls
NO
YES
NO
YES
NO
YES
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
123,584
122,238
128,397
126,982
103,889
102,990
R-Squared
0.10
0.15
0.19
0.20
0.19
0.22
Combined Effect:
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
Interaction:
Notes: We analyze the differential effect of our main explanatory variable for individuals living in urban
environments. Row one shows the uninterested effect of log(1+Distance to Harbor), i.e. the distance effect of
individuals in rural settings, row two shows the differential effect for urban sample participants. Row three depicts
the combined effect of the two constituent terms, i.e. the effect of log(1+Distance to Harbor) for individuals living
in urban environments together with the corresponding p-value in brackets. Results in each column come from
separate regressions and are estimated using the sample of coastal, sub-Saharan African countries included in
survey rounds 1 through 7 of the Afrobarometer. Changes in the number of observations across columns stem from
differences in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do
not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household
items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear
Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level.
***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
49
Table A.17: Extension: Commercial Farmers
(1)
(2)
(3)
(4)
(5)
(6)
log(1+Distance to Harbor)
-0.015***
-0.011**
0.083***
0.045***
-0.032*** -0.023***
(0.004)
(0.006)
(0.008)
(0.012)
(0.003)
(0.006)
log(1+Distance to Harbor) × Commercial Farmer {0,1} -0.047**
-0.047**
0.063
0.088**
0.005
-0.001
(0.022)
(0.022)
(0.041)
(0.040)
(0.018)
(0.018)
Isolated Effect of the Distance to Harbor for Individuals working as Commercial Farmers
log(1+Distance to Harbor) + Interaction
-0.062*** -0.058***
0.146***
0.134***
-0.026*** -0.025***
[0.00]
[0.00]
[0.00]
[0.00]
[0.00]
[0.00]
Basic Controls
NO
YES
NO
YES
NO
YES
Urbanization Controls
NO
YES
NO
YES
NO
YES
Trade-related Controls
NO
YES
NO
YES
NO
YES
Full Geographic Controls
NO
YES
NO
YES
NO
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
Observations
33,084
32,296
37,481
36,637
13,239
13,080
R-Squared
0.06
0.10
0.11
0.15
0.12
0.22
Combined Effect:
Dependent Variable
Cash
Employment
{0,1}
Freq. gone without:
(Water / Food /
Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV /
Motor Vehicle)
[0,1]
Interaction:
Notes: We analyze the differential effect of our main explanatory variable for individuals working as commercial
farmers, i.e. farmers who produce mainly for sale. Row one shows the uninteracted effect of log(1+Distance to
Harbor), i.e. the distance effect of individuals not working as commercial farmers, row two shows the differential
effect for commercial farmers. Row three depicts the combined effect of the two constituent terms, i.e. the effect of
log(1+Distance to Harbor) for individuals working as commercial farmers together with the corresponding p-value
in brackets. Results in each column come from separate regressions and are estimated using the sample of coastal,
sub-Saharan African countries included in survey rounds 1, 2, and 3 of the Afrobarometer for which this detailed
occupational data is available. Remaining changes in the number of observations across columns stem from
differences in the response rates of dependent/independent variables. The sample used in columns (5) and (6) do
not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on ownership of household
items were not asked in this round. Binary dependent variables are estimated through a simple LPM (Linear
Probability Model) specification. The standard errors reported are clustered at the survey enumeration area level.
***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
50
Table A.18: Mediation Analysis: Direct and Indirect Links of Distance to Harbor
Mediator:
Institutions
{1,4}
Mediator:
REC
{0,3}
Mediator:
AU
{0,3}
Mediator:
Institutions
{1,4}
Mediator:
REC
{0,3}
Mediator:
AU
{0,3}
Mediator:
Institutions
{1,4}
Mediator:
REC
{0,3}
Mediator:
AU
{0,3}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
-0.009***
-0.008**
-0.016***
0.035***
0.037***
0.038***
-0.019***
-0.020***
-0.020***
(0.002)
(0.004)
(0.004)
(0.006)
(0.010)
(0.009)
(0.002)
(0.004)
(0.003)
-0.009***
-0.009**
-0.016***
0.041***
0.038***
0.040***
-0.019***
-0.021***
-0.021***
(0.002)
(0.004)
(0.004)
(0.006)
(0.010)
(0.009)
(0.002)
(0.004)
(0.003)
Distance to Harbor via Mediator
0.000
0.000***
0.000***
-0.005***
-0.001***
-0.002***
0.000**
0.001***
0.001***
(0.000)
(0.000)
(0.000)
(0.001)
(0.000)
(0.000)
(0.000)
(0.000)
(0.000)
0.001
0.011***
0.010***
-0.216***
-0.037***
-0.039***
-0.007**
0.016***
0.014***
(0.004)
(0.003)
(0.002)
(0.008)
(0.006)
(0.005)
(0.003)
(0.002)
(0.002)
[0.00]
[0.05]
[0.03]
[0.15]
[0.04]
[0.05]
[0.01]
[0.04]
[0.04]
Basic Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Urbanization Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Trade-related Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Full Geographic Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
Country-Time FE
YES
YES
YES
YES
YES
YES
YES
YES
YES
Observations
121,897
35,216
47,063
126,602
35,380
47,265
102,677
27,769
39,882
Proportion Mediated
Direct Effect:
log(1+Distance toHarbor)
Indirect Effect:
Direct Effect of Mediator:
(Institutions / REC / AU)
log(1+Distance toHarbor)
Dependent Variable
Cash
Employment
{0,1}
How often gone without:
(Water / Food / Cash Inc. / Med.)
[0,4]
Possessions:
(Radio / TV / Motor Vehicle)
[0,1]
Baseline (Total) Effect:
Notes: This table presents results from a formal mediation analysis testing the influence of three potential mediators of log(1+Distance to Harbor) on our three main outcome
variables, respectively. Row one presents the baseline effect, its counterpart is depictable in Table 1. Row two shows the direct effect of our main explanatory variable, i.e. the effect
of distance not attributable to the mediating factor, while row 3 depicts the part of the effect which runs precisely via its influence on the mediator. Row four shows the direct
effect of the mediator on the respective outcome variable. Results in each column come from a separate SEM regression. Changes in the number of observations across columns
stem from differences in the response rates of dependent/independent variables. The sample used is comprised of coastal, sub-Saharan African countries included in survey rounds
1 through 7 of the Afrobarometer. The sample used in columns (7), (8) and (9) do not include individuals surveyed in rounds 1 and 2 of the Afrobarometer, as questions on
ownership of household items (infrastructure) were not asked in this round. Similarly, the sample in columns (2), (5), (8) only include individuals surveyed in rounds 2, 4, and 6
and the sample in columns (3), (6), (9) only include individuals surveyed in rounds 2, 4, 5 and 6. Binary dependent variables are estimated through a simple LPM (Linear
Probability Model) specification. The standard errors reported are clustered at the survey-enumeration area level. * p < .1, ** p < .05, *** p < .01.
51
52
Chapter 3
Heterogeneous Effects of Women’s Schooling
on Fertility, Literacy and Work: Evidence
from Burundi’s Free Primary Schooling
Policy1
Abstract. This article investigates the effect of women’s schooling on fertility as well as on
associated mechanisms by leveraging Burundi’s free primary education policy (FPE) of 2005 as a
natural experiment. Exogenous variation in schooling is identified through a fuzzy regression
discontinuity design (RDD). Our results show that educational attainment was positively influenced
by Burundi’s FPE for women situated at all wealth levels. However, the relevant downstream effects
of schooling – measured by fertility, literacy, and work outcomes – reveal heterogeneous treatment
effects which are moderated by women’s household wealth. While poor women profit in terms of
increases in literacy (6.7 percentage-point increase for each year of policy-induced schooling),
remunerated employment opportunities (5.7 percentage-point increase), as well as a reduction in
desired and actual fertility outcomes (6.9 percentage-point reduction in teenage childbirth), none
of these effects of additional education are observed for women from the wealthier households of
our sample. The evidence of such a marked heterogeneity contributes to the growing literature
examining the nexus between education and fertility in developing countries and helps to evaluate
under which conditions the literature’s findings may generalize.
JEL Classification: I25, I26, J13, O15
Keywords: Female Education, Fertility, sub-Saharan Africa, Regression Discontinuity Design
1 This Chapter is joint work with David Stadelmann. A modified version has been published under the identical
title in the Journal of African Economies, 2024, 33 (1), 67-91. DOI: 10.1093/jae/ejad002. The version
presented here expands on the article to include further evidence on extensions and robustness.
Acknowledgements. Funding from the DFG (project number 390713894) is gratefully acknowledged. We thank
two anonymous referees and the handling editor Brian Dillon for their helpful remarks. I also thank the
participants at the Centre for the Study of African Economies Conference 2023 (Oxford, UK), at the 2020
German Economic Association (VFS) Annual Meeting (Cologne, Germany), and at the Biannual Graduate
Seminar of the faculty of economics at the University of Bayreuth in 2020 for their insightful comments and
remarks at various stages of the project. All remaining errors are my own.
53
3.1 Introduction
“The increase in the education of women and
girls contributes to greater empowerment of women, to
a postponement of the age of marriage and to a
reduction in the size of families.”
(United Nations 1995; 76)
Teenage childbearing is a prevailing problem in many developing countries. Particularly in sub-
Saharan Africa, where 10% of women between the ages of 15 and 19 have already given birth
to at least one child (United Nations 2022). Some of the negative consequences associated with
teenage childbearing are adverse infant and child health, increased maternal mortality, as well
as lower human capital, and less favorable labor market outcomes (see Levine et al. 2009).
Educating women is regarded as an effective policy for tackling many of the developmental
challenges connected to gender equality at large, as prominently outlined in the Millennium
Development Goals (United Nations 2001), and particularly for reducing women’s teenage
pregnancies and the negative consequences linked to them (Lloyd and Young 2009; World Bank
2017).
Indeed, early empirical literature has consistently shown a negative association between
schooling and fertility (e.g. Cochrane 1979; Martin 1995; Ainsworth et al. 1996). However, given
that early childbirth or continued schooling is often jointly determined (e.g. Martínez and
Odhiambo 2018), causal interpretations of these findings are challenging. A growing body of
research therefore explores the education-fertility nexus by employing quasi-experimental
methods (for an overview see Psaki et al. 2019; Evans and Mendez Acosta 2021). In sub-
Saharan Africa, this literature regularly exploits the enactment of free primary education
policies (hereafter FPE) as credibly exogenous shocks to schooling, such as in Ethiopia
(Behrman 2015b; Moussa and Omoeva 2020; Chicoine 2021), Ghana (Adu-Boahen and
Yamauchi 2018), Malawi (Behrman 2015a, 2015b; Grant 2015; Makate and Makate 2016;
Moussa and Omoeva 2020; Zenebe Gebre 2020), Nigeria (Osili and Long 2008), and Uganda
(Behrman 2015a, 2015b; Makate and Makate 2018; Masuda and Yamauchi 2020; Moussa and
Omoeva 2020). These more recent findings identify a negative causal effect of education on
teenage fertility in sub-Saharan African countries. A small set of these studies also started to
explore potential differential effects of FPE on schooling and, importantly, the outcomes
influenced by schooling. Where tested, the results provide evidence that FPE policies almost
exclusively influenced the educational outcomes of the poorer groups in the population (Grépin
and Bharadwaj 2015; Makate and Makate 2016; Keats 2018). However, the absence of
statistically significant increases in FPE-induced education for the wealthy precludes a
comparison of econometrically reliable estimates on downstream effects such as fertility,
54
literacy, or work between poor and wealthy subgroups.2 Still, decisions on schooling and fertility
are dependent on women’s socio-economic background and, correspondingly, their time-use,
particularly in developing contexts (Ilahi 2000; Levine et al. 2009). Thus, identifying and
investigating potential heterogeneities along women’s household wealth, especially in the
downstream effects of female education, represents a relevant research gap.
We contribute to the literature in two ways. First, we add on to the existing evidence by
presenting a statistically robust negative effect of education on women’s teenage fertility and
by providing support for commonly associated mechanisms. Second, we demonstrate that the
estimated downstream effects of policy-induced schooling are distinctly heterogeneous. That is,
there is a marked differential treatment effect governed by women’s household wealth. We
thereby extend previous findings and provide novel (credibly causal) evidence on heterogeneities
materializing in the downstream factors influenced by education. Methodologically, we proceed
similarly to previous studies and consider the implementation of Burundi’s free primary
education (FPE) policy in 2005 as a quasi-experiment, which to our knowledge, has not been
utilized to explore the education-fertility nexus before. We employ a fuzzy regression
discontinuity design (RDD) to identify exogenous increases in schooling and analyze the
downstream effects of schooling through an instrumental variable (IV) approach. To assess
women’s full teenage fertility, we follow the literature and measure education as well as
outcomes influenced by education several years after the policy was implemented. As such, we
use data from Burundi’s Demographic and Health Surveys (DHS) conducted in 2010/2011 and
2016/2017.
Our results show that Burundi’s FPE policy triggered large and statistically significant
increases in education, both for poor and for wealthy women. Being young enough to benefit
from FPE (i.e. being 13 years and younger at the time of implementation) increased girls’
education by 1.22 years compared to women too old to benefit from the policy (14 years and
older). This corresponds to a 34% increase compared to the mean years of schooling for
untreated cohorts. Given that education was strongly increased for both poor and wealthy
women, we are able to investigate this particularly interesting source of heterogeneity in the
outcomes influenced by education, i.e. employ instrumental variables (IV) regressions for poor
and wealthy subgroups separately. These IV-estimates provide evidence that policy-induced
education mainly affects the poor: One additional year of schooling reduces their desired
number of children by 6.4% relative to untreated cohorts and lowers the likelihood of having
their first child during teenage years by 6.9 percentage-points. Estimates of these outcomes for
women from wealthy households are considerably lower, at 2.2% and 1.2 percentage-points,
respectively, and importantly, statistically indistinguishable from zero.
2 In settings where increases in schooling were apparent for both poor and wealthy subgroups (see Adu-Boahen
and Yamauchi 2018; Makate and Makate 2018), results are not indicative of a heterogeneity in the downstream
factors influenced by schooling.
55
Regarding the mechanisms through which education is suggested to influence fertility,
literature has largely proposed two main pathways, often categorized by income and learning
(Michael 1973; McCrary and Royer 2011). The income mechanism rests on the seminal work
of Gary Becker (1960, 1965) and Jacob Mincer (1962, 1963) and emphasizes the increasing
opportunity costs of childbirth and rearing in response to an increase in women’s schooling.
Given the time intensity of these activities, comparative statics have suggested substitution
away from a higher quantity of children towards fewer, more qualitative children, i.e. children
that acquire more human capital (see Becker and Lewis 1973; Willis 1973; Lam and Duryea
1999). Concerning the second mechanism learning, education may influence fertility through
the dissemination of new information as well as the overall expansion of the capacity to access
and absorb information. As such, education may directly exert an effect on women’s
reproductive behavior through curricula (Bankole et al. 2007), both by raising the overall stock
of health knowledge and by facilitating its effective usage (Grossman 1972; Rosenzweig and
Schultz 1989). Education may also influence fertility decisions and health knowledge by
providing women with the skills necessary to attain knowledge about family planning outside
the premises of the school, for instance, through the usage of mass media outlets (Thomas et
al. 1991; Glewwe 1999). Moreover, schools are regarded as an important venue for the
socialization of students, which reinforces aspects of the learning channel and may affect both
channels alike through “assortative mating” (Caldwell 1976, 1980; Mare 1991; Bongaarts and
Watkins 1996; Behrman and Rosenzweig 2002). Besides these two distinct mechanisms, it has
been argued that schooling may reduce fertility simply by keeping women in school and thereby
shortening the available time for sexual activity, a so-called “incarceration effect” (Black et al.
2008; Berthelon and Kruger 2011). Much of the more recent literature explores these suggested
pathways by testing mediating factors such as women’s employment and remuneration on the
one hand, as well as literacy, access and usage of mass media outlets, knowledge, and
implementation of family planning, sexual activity and indicators of assortative mating
(partner’s characteristics) on the other. So far, the evidence is mixed regarding the empirical
support for each of these mechanisms (see Psaki et al. 2019).
The findings we present support both learning and income pathways alike: one additional
year of schooling raises women’s literacy rates by 6.7 percentage-points and increases the
likelihood of working outside of the household for cash by 5.7 percentage-points. Given that
both of these effects are apparent only for the poor women in our sample, these results agree
with the differential fertility responses we observe, substantiating the potential significance of
these two channels.
We conduct an array of robustness checks to validate our results and the heterogeneous
nature of them. We thereby test for potentially confounding differences between treated and
untreated groups going back the time of intervention, assess the influence of differing conflict
intensities of Burundi’s Civil War, and check the robustness of our results to placebo cut-offs,
56
differing bandwidths and functional forms (see details in Section 3.5). Given that our employed
wealth score measures household wealth at the time of the survey, we cannot preclude that
FPE may have induced women to increase household wealth levels via added schooling which
represents a potential endogeneity issue concerning our differential treatment effects. When
directly exploring whether women’s wealth score as well as other direct measures of income
and living standards are affected by education in our IV setting, we find no evidence for this
type of endogeneity. Given that over half of the women in our treated sample are still considered
dependents, it is not surprising that significant wealth changes have not (yet) materialized
through FPE-induced schooling. We also investigate data from the Multiple Indicator Cluster
Surveys (MICS) to test differences in outcomes of poor as well as wealthy treated and untreated
groups directly before and after Burundi’s free primary education policy in a difference-in-
differences framework. All of our results and interpretations remain robust, particularly their
heterogeneous nature.
The remainder of this paper is structured as follows: We introduce the background to our
identification and the empirical strategy in Section 3.2. The data employed is introduced in
3.3. Section 3.4 presents the empirical results and a discussion thereof. Section 3.5 provides our
robustness and validity checks. Section 3.6 offers concluding remarks.
3.2 Institutional Background and Identification Strategy
3.2.1. Institutional Background
The World Declaration On Education For All in 1990 and the Dakar Framework For Action in
2000 manifested the promise of developing countries to achieve universal primary education
(UPE) by the year 2015. Prompted by these commitments, many countries in sub-Saharan
Africa enacted policy changes toward free primary education at the turn of the century, mainly
by focusing on the elimination of user fees (Kattan 2006). User fees have been identified as a
crucial obstacle to school attendance in developing regions, particularly for the poor (Deininger
2003; Grogan 2008; Bold et al. 2013; Travaglianti 2017; Burlando and Bbaale 2022).
Subsequently, African countries have made tremendous strides in increasing educational
enrolment due to these interventions, improving the net enrolment rate (NER) from 53% to
79% between 1990 and 2018 (UIS 2022). However, many of them are still facing obstacles in
moving towards universal primary education: Large within-country differences in educational
attainment prevail and are outlined by factors such as individuals’ wealth, place of residence,
ethnicity, and gender (Levine et al. 2009; World Bank 2017).
Burundi is one of the poorest countries in the world, with an income per capita of $793
and over 65% of the population living below the national poverty line (World Bank 2021). The
country is densely populated, and most of its 11.2 million inhabitants (87%) live in rural areas
(Dunlop and King 2019). Burundi’s annual fertility rate and population growth stand at 5%
57
and 2.7%, respectively, well above the sub-Saharan African average of 4.5% and 2.5% (United
Nations 2022). Since its independence in 1962, Burundi has undergone several periods of
violence. The most noteworthy conflict was the genocidal massacre of 1993 which triggered the
subsequent civil war lasting, at least formally, until the year 2000, in which the Arusha Peace
and Reconciliation Agreement (APRA) was signed (Bundervoet et al. 2009; Travaglianti 2017).
Access to education was part of the struggle citizens fought for in the war and a core element
of APRA as well as the constitution adopted in 2005, when the direct fees of primary schooling
were removed (Travaglianti 2017; Dunlop and King 2019).3 The waiver of school fees was
accompanied by a surge in the construction of schools and an overall increase in budgetary
spending on education, especially at the primary level, which rose from 37% in 2001 to 54% in
2008 of the total educational budget (Travaglianti 2017). As a result, primary school enrolment
increased from 0.97 million in 2004 to 1.32 million in 2006, expanding the gross enrolment ratio
(GER) from 76% to 103% in just two years (UIS 2022). School participation continuously
increased thereafter, arriving at a maximum GER of 141% in 2011. Note that a shift of children
from wealthier families into private schools was not observed. At around 1.5% in the years
before the policy, the proportion of primary school children in private institutions has
historically been low and actually decreased after the introduction of FPE (UIS 2022).
3.2.2. Identification Strategy
To establish credibly causal effects of education, we draw upon the recent body of research and
consider Burundi’s free primary education policy (FPE) of 2005 as a natural experiment. We
identify exogenous (policy-induced) variation in schooling by employing a fuzzy regression
discontinuity design (RDD), exploiting the fact that FPE exogenously sorts women into
treatment or control, depending on women’s age at the time of implementation. To the best of
our knowledge, Burundi’s FPE of 2005 has not been used to study the education-fertility nexus
before. In Burundi, the full primary school cycle entails six years of schooling to be attained
during the official primary school ages of 7 to 12 (Ministere e la Presidence charge de la Bonne
Gouvernance et du Plan - MPBGP et al. 2017). However, overage enrolment is common. Data
from the 2005 Multiple Indicator Cluster Survey (MICS) shows a mean age of 8.2 for girls
entering first grade at the time of policy implementation (UNICEF 2005), which differs
somewhat between girls living in wealthier households (7.78 years) and girls from poorer ones
(8.46 years). Following the literature, we account for this de facto age of entry into primary
school in our estimations and consider a primary school leaving age of 13 rather than 12 (see
also Adu-Boahen and Yamauchi 2018; Keats 2018). As such, given that fees were removed from
the school year starting in September of 2005, women who were born in 1992 or later (aged 13
3 While the direct fees of schooling were removed, other costs of schooling such as uniforms, materials and
contributions to the school’s budget continued to persist, as was the case in many other countries implementing
FPE. Still, removing the direct schooling fees had a “psychological effect” (Travaglianti 2017; 112), deeming
school overall more affordable than before.
58
or younger at policy implementation), are considered young enough to benefit from free
schooling and are therefore sorted into the treatment group. Accordingly, women born in 1991
or earlier (aged 14 and older) were too old to benefit from the policy and sorted into the control
group. Hence, the RDD design in our setting compares women who were “just-treated” to
women who were just too old to benefit from FPE (see e.g. Behrman 2015a, 2015b; Makate
and Makate 2018; Keats 2018; Moussa and Omoeva 2020). Next to a late entry into school,
factors such as absenteeism and grade repetition are common in sub-Saharan Africa. In 2006,
there were approximately 24% overaged children enrolled in Burundi’s primary school grades
(UIS 2022). It is thereby plausible that women older than the de facto primary school age were
exposed to the policy change, by enrolling in lower primary school grades at later ages or by
continuing with their remaining primary school years (Behrman 2015b). On the other hand,
some girls younger than the de facto primary school ages have already left schooling behind
permanently, be it for marriage, for work, or similar duties, and will thereby not benefit from
the removal of tuition fees. Given such noncompliance, the estimation will be made through a
fuzzy regression discontinuity design.
Note that the effect estimated in a fuzzy RDD resembles an instrumental variable
estimate: It is defined as the ratio of the reduced form estimate of the discontinuity in the
outcome variable 𝑌𝑌𝑖𝑖 to the estimate of the discontinuity in the treatment, 𝑆𝑆𝑆𝑆ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑖𝑖, both
determined by the assignment variable 𝐵𝐵𝑖𝑖 (year of birth) and evaluated at the cut-off 𝑐𝑐 (Hahn
et al. 2001; Imbens and Lemieux 2008; Lee and Lemieux 2010). The estimation of our
downstream effects of schooling therefore consists of two stages: The first-stage, regressing the
treatment 𝑆𝑆𝑆𝑆ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑖𝑖 on the identifying instrument 𝑍𝑍𝑖𝑖, whereby 𝑆𝑆𝑆𝑆ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑖𝑖 is a continuous
measure for years of schooling and 𝑍𝑍𝑖𝑖 is a binary indicator 𝑍𝑍𝑖𝑖= 1[𝑌𝑌𝑌𝑌𝑌𝑌𝑌𝑌 𝑜𝑜𝑜𝑜 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵ℎ≥𝑐𝑐(1992)].
The second-stage, regressing the outcome variable 𝑌𝑌𝑖𝑖 on the instrumented endogenous
variable, 𝑆𝑆𝑆𝑆ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑖𝑖
, estimated in the first stage.
The regression discontinuity literature favors a non-parametric, or local, estimation with
lower-order polynomials over a global specification with higher-order polynomials (Gelman and
Imbens 2019). Thus, an appropriate window of data (hereafter bandwidth) has to be selected
for both sides of the cut-off. We choose the bandwidth that minimizes the modified “leave-one-
out” cross-validation function (CV) introduced by Ludwig and Miller (2005) and countercheck
these results with the Imbens and Kalyanaraman (2012) proposed bandwidth selection
procedure. The value of the CV function (MSE) stabilizes at a bandwidth size of five birth year
cohorts on either side of the cut-off.4 Thus, all our estimates are based on a bandwidth size of
five, effectively comparing women born in 1987-1991 to women born in 1992-1996 (aged 9-13
and 14-18 at the time of the policy). Both our first and second-stage results are robust to
varying bandwidth sizes (see Section 3.5).
4 The CV confirms the bandwidth size of five within the subsamples of poor and wealthy women as well.
59
To fit an appropriate functional form to the bandwidth selected, we test six different
specifications of the regression equation. As such, we estimate linear, quadratic, and cubic
specifications, each with and without an interaction term between the treatment indicator 𝑍𝑍𝑖𝑖
and the assignment variable 𝐵𝐵𝑖𝑖 (allowing the slope to vary before and after the cut-off). We
select a linear functional form with a varying slope before and after the cut-off as our baseline
specification, given that it minimizes the Akaike information criterion (AIC) among the tested
specifications. Both our first- and second-stage results are robust to the choice of the functional
form (see Section 3.5). The main results presented in this paper are estimated through the
following two-stage regression equation:
𝑆𝑆𝑆𝑆ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑖𝑖= 𝛼𝛼0 + 𝛼𝛼1𝑍𝑍𝑖𝑖+ 𝛼𝛼2(𝐵𝐵𝑖𝑖−1992) + 𝛼𝛼3𝑍𝑍𝑖𝑖× (𝐵𝐵𝑖𝑖−1992) + 𝛼𝛼𝑘𝑘𝑋𝑋𝑖𝑖+ 𝑒𝑒𝑖𝑖
(3.1)n
𝑌𝑌𝑖𝑖= 𝛽𝛽0 + 𝛽𝛽1𝑅𝑅𝑅𝑅𝑆𝑆𝑆𝑆ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑖𝑖
+ 𝛽𝛽2(𝐵𝐵𝑖𝑖−1992) + 𝛽𝛽3𝑍𝑍𝑖𝑖× (𝐵𝐵𝑖𝑖−1992) + 𝛽𝛽𝑘𝑘𝑋𝑋𝑖𝑖 + 𝑢𝑢𝑖𝑖
(3.2)n
All variables are defined as above. 𝑋𝑋𝑘𝑘 is an additional vector of control variables and 𝑒𝑒𝑖𝑖, 𝑢𝑢𝑖𝑖 are
the idiosyncratic error terms. 𝛽𝛽1𝑅𝑅𝑅𝑅 is the causal effect of schooling and represents the local
average treatment effect (LATE) on compliers (Angrist et al. 1996; Lee and Lemieux 2010). As
is common in the literature employing RD designs, the assignment variable 𝐵𝐵𝑖𝑖 is centered at
the cut-off point 𝑐𝑐= 1992 such that the constituent terms of the interaction, namely 𝛼𝛼1 and
𝛽𝛽1𝑅𝑅𝑅𝑅 , can be directly interpreted as the effect at the discontinuity. Binary outcome variables
are estimated through a linear-probability model (e.g. Adu-Boahen and Yamauchi 2018; Keats
2018; Ozier 2018).5
3.3 Data and Descriptive Statistics
The data we use come from the individual recodes of the 2010/2011 and the 2016/2017 Burundi
Demographic and Health Surveys (DHS). DHS are cross-sectional, household-based surveys
collecting a broad array of information on topics such as demographics, education, employment
and occupation, as well as fertility and family planning (Croft et al. 2018). The main
respondents are women of reproductive age (15-49).
As a result of the empirical specification (bandwidth), the sample is restricted to women
born five years before and after the cut-off year 1992, resulting in our main sample of 7,714
women aged 20-30 at the time of the survey (born between 1987-1996). We exclude women
younger than 20 years of age to assess women’s full teenage fertility and to analyze outcomes
for women who have mostly completed schooling (see Ferré 2009; Keats 2018).6 We thereby
measure outcomes of women several years after the policy was conducted. Following the
5 Our estimates on binary outcome variables are robust to the use of a Probit model.
6 The DHS Household recodes show that only 24% of 20-year-old women are still attending school. Of them,
88% are enrolled in secondary school.
60
recommendation by the DHS, we weight all analyses with the included survey weights, making
the outputs representative at the national level, and account for the pooling of datasets by
normalizing the sum of weights across surveys. Table 3.1 reports summary statistics for the
main sample as well as for the poor and wealthy subgroups separately.
To explore heterogeneities, we employ the DHS wealth index score, which is constructed
by assessing households’ dwelling and living arrangements, i.e. construction materials,
sanitation facilities, as well as household durables and assets such as radio and TV, bicycles,
cars, etc., and by placing them on a relative scale of wealth within the sample (Rutstein and
Johnson 2004). We consider respondents from households scoring on and below the median
wealth score as “poor”, and correspondingly, respondents with a denoted household wealth
level above the median wealth score as “wealthy”. Summary Statistics (Table 3.1) reveal that
the poor subgroup is significantly deprived in terms of access to basic utilities and services:
Less than 1% of poor households have access to electricity or a flush toilet, only 17% have
drinking water piped into their dwelling, and almost all of the households in the poorer
subgroup (99%) are dependent on wood as a fuel source. The increased effort in the provision
of many of these basic goods can figure largely in household members’ time-use, particularly
in females’ on whom these tasks are often levied (Ilahi 2000; Levine et al. 2009). Wealthy
households are endowed much more favorably in these basic goods and utilities, which is why
an investigation into the differential effects of women’s schooling along household wealth can
be highly informative. Note that the DHS wealth index captures respondents’ wealth at the
time of the survey. We discuss and address potential caveats pertaining to this wealth grouping,
alternatives, and robustness checks in Section 3.5.
3.3.1. Dependent Variables
To assess adolescent fertility outcomes, we employ women’s reported age at first birth and
construct a dichotomous indicator of having had a first birth before the age of 20. Similarly,
we investigate the incidence of teenage marriage by estimating the probability of having been
married before the age of 20.
To explore the pathways through which education is suggested to influence fertility, we
first investigate respondents’ levels of literacy as well as their desired number of children. These
variables serve as proxies for investigating a potential learning pathway. Note that additional
years of education must not necessarily lead to improved learning outcomes, especially given
that the abolition of school fees in Africa has been associated with the overcrowding of schools
and compromised learning environments for children (Deininger 2003; Lucas and Mbiti 2012;
Bold et al. 2013). The measure on women’s literacy thereby also serves as a direct assessment
of the quality and effectiveness of FPE-induced schooling. To investigate a potential income
pathway, we test whether women have work other than domestic housework, e.g. are working
in the family business/farm, are selling items, are (self)-employed, etc., currently or anytime
61
Table 3.1: Summary Statistics
Control Treatment
Control Treatment
Control Treatment
Number of Observations
4476
3238
2302
1556
2174
1682
Age at Survey
24.19
21.96
24.16
21.93
24.23
21.99
[3.05]
[1.46]
[3.03]
[1.47]
[3.08]
[1.45]
Education (completed schooling years)
3.55
5.60
1.99
3.98
5.56
7.45
[4.18]
[4.30]
[2.79]
[3.60]
[4.77]
[4.29]
Never received schooling {0,1}
0.44
0.23
0.57
0.32
0.26
0.13
Residence: Rural {0,1}
0.87
0.85
0.99
0.99
0.73
0.69
Number of Siblings
6.20
5.91
6.18
5.95
6.22
5.86
[2.47]
[2.42]
[2.45]
[2.33]
[2.51]
[2.53]
Religion
Catholic {0,1}
0.59
0.55
0.62
0.61
0.54
0.49
Protestant {0,1}
0.34
0.36
0.31
0.31
0.38
0.42
Muslim {0,1}
0.03
0.04
0.01
0.02
0.05
0.05
Wealth Quintile (1-5)
3.01
3.09
1.94
1.97
4.38
4.38
Dwelling
Electricity {0,1}
0.10
0.12
0.00
0.00
0.22
0.26
Wood Fuel {0,1}
0.85
0.81
0.99
0.98
0.66
0.61
Piped Water {0,1}
0.32
0.37
0.17
0.17
0.53
0.60
Flush Toilet {0,1}
0.04
0.06
0.00
0.00
0.10
0.13
Fertility
Age at First Birth
19.67
18.98
19.59
18.99
19.80
18.97
[2.55]
[2.06]
[2.43]
[1.98]
[2.72]
[2.20]
Has given Birth before Age 20 {0,1}
0.34
0.32
0.37
0.39
0.31
0.25
Reproductive Behavior
Age First Marriage
18.78
18.17
18.60
18.03
19.07
18.42
[2.72]
[2.19]
[2.52]
[2.09]
[2.98]
[2.34]
Married before Age 20 {0,1}
0.48
0.39
0.54
0.49
0.41
0.28
Age at First Sex
13.63
11.55
14.28
12.88
12.85
10.04
[8.62]
[8.74]
[8.08]
[8.21]
[9.16]
[9.07]
Has had Sex before Age 20 {0,1}
0.55
0.51
0.59
0.58
0.50
0.42
Pathway: Learning
Literacy: Able to read sentence {0,1}
0.56
0.71
0.43
0.60
0.73
0.84
Desired Number of Children
4.01
3.67
4.05
3.63
3.96
3.72
[1.42]
[1.33]
[1.43]
[1.30]
[1.41]
[1.37]
Pathway: Income
Worked Last Year {0,1}
0.86
0.81
0.92
0.90
0.78
0.71
Works outside of the HH. for Cash/Money {0,1}
0.08
0.09
0.08
0.10
0.07
0.08
Sample
Full Sample
Poor
Wealthy
Notes: The sample consists of the 2010/2011 and 2016/2017 Burundi DHS Female Surveys, using a bandwidth
size of five birth year cohorts on either side of the cut-off. Respondents are restricted to age 20 and older at the
time of the survey. The “Control” group is comprised of individuals born between 1987-1991, the “Treatment”
group is comprised of individuals born between 1992-1996. The “poor” and “wealthy” subsamples include women
whose households score on and below as well as above the median wealth level, respectively. The statistics are
weighted using the DHS sample weights and account for the pooling of survey rounds.
62
within the last year. Additionally, we specifically test whether respondents’ work is employed
work and if it is remunerated, i.e. whether they are working outside of the household for
cash/money. We investigate additional factors associated with the proposed mechanisms. This
includes respondents’ reproductive behavior, measured by women’s age at first sexual
intercourse as well as their knowledge and usage of contraceptive methods (condoms). We also
investigate women’s engagement with mass media outlets such as newspapers-, radio- and tv-
usage and explore further aspects of assortative mating, i.e. test husbands’ characteristics such
as their age, education or fertility preferences.
3.3.2. Independent Variables
Our main explanatory variable of interest is women’s years of schooling. DHS surveys report
the number of completed schooling years, which is a continuous measurement of educational
attainment. The variable is constructed by asking respondents: “What is the highest level of
school you attended?” and subsequently, “what is the highest (standard/form/year) you
completed at that level?” (The DHS Program 2015; 2).7
We include a set of control variables that were plausibly fixed before the FPE policy
implementation. We thereby add dummy variables capturing women’s religious affiliation, as it
may influence women’s behavior through doctrines and values stipulating certain roles of
women within the family (Makate and Makate 2018). We also add a control variable indicating
the respondent’s number of siblings. A larger number of siblings could negatively influence girls
in their ability to receive schooling (Ewemooje et al. 2020). It could also have a non-negligible
impact on girls’ own desired family size. To account for regional discrepancies in schooling as
well as fertility patterns, we include province fixed effects for Burundi’s eighteen provinces8 and
also add a dummy variable indicating whether the respondent resides in rural or urban
dwellings. Lastly, we add survey fixed effects to account for differences in women’s educational
attainment or fertility outcomes, which might vary systematically between the two periods of
the survey rounds.
Given the usual difficulty in choosing between a sufficient number of controls and
overfitting, we also investigate the robustness of our first- and second-stage results to the Post-
Double-Selection (PDS) algorithm developed in Belloni et al. (2014). We thereby include
further potential control variables from the DHS as well as other sources and allow for higher
order terms as well as the interaction among them. In total, we consider 896 controls. Our
results are robust to PDS’ selected controls, in fact, our coefficients are more precisely
estimated, if anything. Results are provided in the Appendix, Table B.1 through Table B.4.
7 For example, if the respondent’s highest level of schooling is secondary education and she completed 2 years
at that level, the variable is constructed by adding all primary school grades (6 in Burundi) + 2 years of
secondary school = 8 years of total schooling.
8 We use the geo-coded sample clusters of the 2010/2011 survey round to allocate households to Burundi’s 18
provinces to “match” the province information provided in the 2016/2017 survey round.
63
3.4 Results
3.4.1. The Effect of FPE on Schooling: First-Stage RDD Evidence
Figure 3.1 and Figure 3.2 provide visual evidence that Burundi’s FPE affected women’s
educational attainment. They depict long-term trends in education for women by birth year
using the pooled sample (Figure 3.1) and the sub-samples split by household wealth (Figure
3.2).
Figure 3.1: Women’s Schooling by Birth Year Cohort
Figure 3.1 illustrates that women who benefited from free primary schooling, i.e. women
born in or after 1992, show for an immediate and substantial increase in educational outcomes
compared to the cohorts born prior. And given that girls born in later years had increasingly
more (primary) schooling years left to attend without having to pay fees, educational
attainment is increasing in birth year cohorts after the cut-off.
Figure 3.2 suggests that both poor and wealthy women were able to benefit from the
introduction of FPE. However, the observation of increasing educational attainment for women
born after the cut-off is dissimilar and can only be observed for the poor subgroup. Taking
account of the very low levels of schooling of poor women born before the cut-off, it seems
Notes: The plot depicts yearly averages of highest schooling grades achieved by individuals born
within the respective birth cohort indicated on the x-axis. The “Local Linear Fit” fits the slope
estimated from a regression of y on x using the (local) data window indicated by the different
color hue. Data come from the 2010/2011 and the 2016/2017 Burundi DHS Female Surveys.
Respondents are restricted to age 20 and older. All estimates are weighted using the DHS sample
survey weights.
64
likely that mainly girls from currently poor households were constrained by the direct cost of
schooling prior to FPE. A reason for the larger initial discontinuity, but non-increasing trend
line, for the wealthy women after the introduction of the policy might be that FPE precisely
induced the women from wealthy households who had achieved some primary education but
were constrained to complete the full six-year cycle, to finish primary schooling. Note that we
also observe a steep incline in educational attainment for cohorts born two decades earlier (in
the 1970s), followed by a sharp drop-off for women born in and after 1980. The increase, as
Figure 3.2: Women’s Schooling by Birth Year Cohort, separated by Household Wealth
well as the decline in schooling for these cohorts, has been attributed to the expansion of the
primary education sector in the 1980s, the successes of which were quickly eradicated for
cohorts subject to the violent conflict during Burundi’s Civil War (Obura 2008; Verwimp and
Van Bavel 2014). Subsequent birth cohorts’ educational attainment slowly recovered to pre-
war levels (Travaglianti 2017) and ultimately surpassed them in one instance when tuition fees
were removed in 2005. Given our identification strategy, these pre-policy developments do not
affect our results as we draw on data from birth cohorts born in 1987 to 1992 for our estimation
(see Subsection 3.2.2, and in particular, Section 3.5 for robustness and validity tests).
Table 3.2 presents first-stage discontinuity as specified in regression equation (3.1). The
Notes: The plot depicts yearly averages of highest schooling grades achieved by individuals born within the
respective birth cohort indicated on the x-axis. The “Local Linear Fit” fits the slope estimated from a
regression of y on x using the (local) data window indicated by the different color hue. Estimates for the
“poor” and “wealthy” are generated from subsamples split at the median household wealth. Data come from
the 2010/2011 and the 2016/2017 Burundi DHS Female Surveys. Respondents are restricted to age 20 and
older. All estimates are weighted using the DHS sample survey weights.
65
results are presented for the full sample (columns 1 and 2), as well as the poor (columns 3 and
4) and wealthy subsamples (columns 5 and 6), separately.9 Odd columns show results of a
parsimonious specification, only including survey fixed effects. For the discussion of results, we
refer to the more stringent specification, including all controls and fixed effects presented in
the even columns. The discontinuous increase in schooling generated by Burundi’s FPE policy
is presented in row one. The coefficient estimate represents the increase in the years of education
Table 3.2: First-Stage Estimates, Years of Schooling
(1)
(2)
(3)
(4)
(5)
(6)
1[Year of Birth ≥ 1992]
1.344*** 1.223***
0.849*** 0.908***
1.356*** 1.280***
(0.255)
(0.223)
(0.228)
(0.220)
(0.397)
(0.367)
Mean of Control Group
Interaction: 1 × Year of Birth
0.135
0.164**
0.378*** 0.339***
-0.179
-0.099
(0.090)
(0.077)
(0.081)
(0.078)
(0.141)
(0.130)
Year of Birth
0.052
0.038
0.059
0.064
0.116
0.070
(0.054)
(0.047)
(0.046)
(0.044)
(0.090)
(0.085)
Effect on the second cohort treated by FPE
1[YoB ≥ 1992] + 1 × YoB + YoB
1.426***
1.312***
1.250***
(0.197)
(0.187)
(0.338)
Observations
7,713
7,709
3,857
3,856
3,856
3,853
Adjusted R-squared
0.054
0.263
0.099
0.183
0.041
0.187
Controls
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
Regression Sample
Full Sample
Poor
Wealthy
[3.55]
[1.99]
[5.56]
9 The regression results may be visually inspected in Figure B.1 through Figure B.3 in the Appendix (without
controls and fixed effects).
Notes: The results in each column are produced by a separate regression. The sample used for the
estimations consists of the 2010/2011 and 2016/2017 Burundi’s DHS Female Surveys, using a bandwidth
size of five birth year cohorts on either side of the cutoff. Respondents are restricted to age 20 and older
at the time of the survey. Columns (1) and (2) are estimated from our full sample. Columns (3) and (4)
are estimated using the subsample of poor women (below and including the median wealth score), and
columns (5) and (6) are estimated using the subsample of wealthy women (scoring above the median
wealth score). Odd columns include survey fixed effects. Even columns include region (province) fixed
effects as well as controls on religious affiliation, number of siblings and urban/rural status. The variable
“Year of Birth” has been re-centered such that the coefficient in row one can be directly interpreted as
the discontinuous difference at the cutoff year 1992. “Mean of Control Group” indicates the mean of the
dependent variable for women in untreated birth year cohorts (the control group). Estimates are weighted
using the DHS sample weights and account for the pooling of survey rounds. The standard errors reported
are clustered at the survey-cluster level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively.
66
for women who were 13 years old (born in 1992) at the time of the policy compared to women
who were 14 years old (born in 1991). The estimate is economically large and statistically
significant at the 1% level throughout all columns. Being just young enough to benefit from
free schooling increased women’s educational attainment between 0.908 (specification 4) and
1.280 years (specification 6). In comparison to related findings, girls “just-treated” by tuition-
free primary schooling gained about 0.61 years of schooling in Ethiopia (Behrman 2015b)
between 0.31 to 0.46 years in Malawi (Behrman 2015a, 2015b) and between 0.64 to 1.24 years
in Uganda (Keats 2018; Makate and Makate 2018). Our slightly larger estimate may be
explained by the low levels of education prior to the FPE. Interestingly, Burundian women
from the wealthy subgroup had a larger increase in schooling at the cut-off than the poor.
However, as Figure 3.2 shows, poor women born after 1992 continuously increased their
educational attainment, which is captured by the large positive effect of the interaction term
“1[Year of Birth ≥ 1992] × Year of Birth”. Hence, when comparing the effect of the second
treated cohort after the cut-off (born in 1993), the poor already show for larger increases in
education compared to the wealthy, (0.908+0.339+0.064=) 1.312 years to (1.280-
0.099+0.070=) 1.250 years, respectively. Note that the mean schooling years of the control
cohorts differ largely between the wealthy and the poor. Poor women not exposed to the policy
had an average of 1.99 years of schooling, compared to 5.56 years for the wealthy. As such,
while FPE added about the same number of schooling years for poor and rich women, it more
than doubled the amount of schooling for the poor in relative terms. See Subsection 3.4.3 for
a more thorough discussion of this observation. The coefficient estimates are robust to the
inclusion of all of our controls and fixed effects, as well as controls selected by the PDS
algorithm (see Table B.1 in the Appendix).
3.4.2. The Effects of Schooling: Second-Stage Evidence
The first-stage estimates provide evidence for substantial policy-induced increases in schooling.
Thus, they present strong instruments for estimating the effect of additional education on our
dependent variables in an IV setting. Given that both poor and wealthy women exhibit large
and significant discontinuities in schooling, we have the opportunity to additionally test for
differential treatment effects of added education for poor and wealthy women separately.
Literacy and Desired Fertility. Table 3.3 panel A investigates the effect of increased
schooling due to the FPE policy on women’s literacy. The first-stage discontinuity delivers a
strong instrument as shown by the F-Statistic (29.9). Results show that one additional year of
schooling increased women’s literacy by 5.0 percentage-points, which marks an increase of 8.8%
compared to the mean of the control group. This is in line with findings from related studies
which consistently show positive impacts of education on women’s literacy, e.g. Keats (2018)
or Behrman (2015b). However, there is relevant heterogeneity in our estimated effect of
schooling as shown in columns (3) through (6). Poor women increased their rate of literacy
67
through policy-induced schooling by 15.6% compared to (a statistically insignificant) increase
of 3.7% by wealthy women. Importantly, note that untreated cohorts of wealthy women had
higher levels of literacy overall, 73% compared to 43% for the poor.
These results suggest that for the poor, free access to schooling was successful in providing
them with basic skills, presumably acquired at lower levels of schooling. It also hints towards
the effectiveness of FPE-induced schooling in general and goes against some of the suggestions
of depreciating school quality due to issues such as overcrowding (e.g. Deininger 2003; Lucas
and Mbiti 2012; Bold, Kimenyi, and Sandefur 2013). The results are also consistent with the
aggregate evidence that switching into private schools was not observed in Burundi in the years
following the policy implementation (UIS 2022).
Table 3.3: Second-Stage Results, Literacy and Desired Fertility
(1)
(2)
(3)
(4)
(5)
(6)
Panel A: Literacy {0,1}
Schooling
0.054*** 0.050***
0.067**
0.067**
0.033
0.027
(0.015)
(0.017)
(0.033)
(0.030)
(0.022) (0.024)
Mean of Control Group
F-Statistic
27.7
29.9
13.8
17.1
11.4
11.9
Observations
7,708
7,704
3,856
3,855
3,852
3,849
Panel B: Desired Number of Children
Schooling
-0.144** -0.151**
-0.307** -0.261**
-0.093
-0.088
(0.060)
(0.063)
(0.136)
(0.118)
(0.086) (0.086)
Mean of Control Group
F-Statistic
25.9
28.3
12.5
15.8
11.4
11.9
Observations
7,520
7,516
3,764
3,763
3,756
3,753
Controls
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
Regression Sample
Full Sample
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
[4.01]
[4.05]
[3.96]
[0.56]
[0.43]
[0.73]
Burundi has high rates of fertility, with an average of 5 births per woman (United Nations
2022). Panel B of Table 3.3 shows that a large portion of these high rates of fertility is reflected
in the number of children a woman desires to bear throughout her lifetime, with an average of
Notes: The results in each panel and column are produced by a separate regression. The IV estimate and
key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or after
the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the
reported Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights
and account for the pooling of survey rounds. The standard errors reported are clustered at the survey-
cluster level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes
below Table 3.2.
68
4 children per woman in our full sample. Our second-stage estimates suggest that education
negatively influences women’s desired lifetime fertility, as shown in panel B of Table 3.3. Our
results are comparable to recent evidence from other countries: Behrman (2015b), Keats (2018),
and Masuda and Yamauchi (Masuda and Yamauchi 2020) report decreases in the number of
desired children between -0.11 and -0.34, which are well in line with our estimated coefficient
of -0.151. However, different from the extant literature, our results show that the effect is driven
mainly by women from the lower socio-economic strata of the population (columns 3 to 6).
Compared to women in the control group, treated women decreased their fertility preference
by 3.7%, the poorer subgroup by 6.4%, and the wealthy by (a statistically insignificant) 2.2%.
An explanation for this is that education exerts its negative impact on women’s desired fertility
through increased levels of literacy, which were raised only for the poor. This finding is
supportive evidence of the pathway learning since literacy acts as a proxy for girls’ learning
(capacity). Overall, these results suggest an education-driven convergence of fertility preference
between the poor and the wealthy.
Fertility and Reproductive Behavior. Panel A and B of Table 3.4 investigate the effect of
education on women’s fertility outcomes. A one-year increase in schooling postpones women’s
age at first birth by almost half a year (panel A, column 2) and reduces the likelihood of
teenage childbearing by 3.4 percentage-points (panel B, column 2).10 Once more, the effect of
added education seems to be driven by women from the poorer half of the population, which
is shown when comparing the estimates of column (4) with those in column (6) both in panel
A and in panel B, exploring women’s age at first birth and the probability of having had a first
birth during teenage years.11 One additional year of education for the poor decreases the
probability of having had a first birth before the age of 20 by 6.9 percentage-points, compared
to a statistically insignificant 1.2 percentage-points for women from wealthier households.
Marriage and cohabitation are arguably one of the most direct channels through which teenage
childbearing is promoted (Presler-Marshall and Jones 2012). Our results show that one year of
added schooling decreases the likelihood of being married before the age of 20 by 6.1 percentage-
points, which corresponds to a 12.7% decrease in comparison to the control group (panel C,
column 2). As such, the evidence from Burundi corresponds to that of studies on other
developing countries assessing the link between education and fertility, which consistently
report significant reductions in the likelihood of teenage marriage alongside decreased teenage
childbearing (e.g. Keats 2018; Masuda and Yamauchi 2020).
10 We further test the probability of having had a first birth at additional age increments (from 16 to 23) in
Table B.5 in the Appendix and show that the effect of decreased teenage childbirth was apparent at the ages
19 and 20. This might be interpreted as evidence against an “incarceration effect”, given that women were
most likely out of school at these age ranges.
11 Note that the IV-estimate for wealthy women in panel A of Table 3.4 suffers from weak instrument bias,
with an F-Statistic of 2.1 (Staiger and Stock 1997). This is due to a sample reduction given that not all women
in our sample, especially the wealthy, have born children at the time of the survey. Estimates from panel B
are arguably a more reliable indicator of altered fertility behavior via increased schooling.
69
Table 3.4: Second-Stage Results, Fertility and Reproductive Behavior
(1)
(2)
(3)
(4)
(5)
(6)
Panel A: Age at first Birth
Schooling
0.440**
0.468**
0.435
0.449*
0.640*
0.757
(0.179)
(0.204)
(0.274)
(0.263)
(0.381)
(0.564)
Mean Ctrl. Group
F-Statistic
14.9
11.9
9.1
10.3
3.9
2.1
Observations
4,918
4,917
2,830
2,830
2,088
2,087
Panel B: First Birth before Age 20 {0,1}
Schooling
-0.037**
-0.034*
-0.056
-0.069*
-0.026
-0.012
(0.018)
(0.019)
(0.041)
(0.039)
(0.024)
(0.025)
Mean of Control Group
F-Statistic
27.9
30.0
13.8
17.0
11.7
12.2
Observations
7,713
7,709
3,857
3,856
3,856
3,853
Panel C: Married before Age 20 {0,1}
Schooling
-0.063*** -0.061***
-0.078** -0.090**
-0.059** -0.045*
(0.018)
(0.018)
(0.039)
(0.037)
(0.025)
(0.025)
Mean of Control Group
F-Statistic
27.9
30.0
13.8
17.0
11.7
12.2
Observations
7,713
7,709
3,857
3,856
3,856
3,853
Panel D: Sex before Age 20 {0,1}
Schooling
-0.044**
-0.042**
-0.047
-0.064*
-0.046*
-0.031
(0.018)
(0.019)
(0.037)
(0.035)
(0.028)
(0.029)
Mean of Control Group
F-Statistic
28.5
29.9
14.8
17.9
11.2
11.4
Observations
7,623
7,619
3,807
3,806
3,816
3,813
Controls
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
Instrumented by 1[YoB ≥ 1992]
[0.55]
[0.59]
[0.50]
Full
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
Regression Sample
Instrumented by 1[YoB ≥ 1992]
[0.48]
[0.54]
[0.41]
[19.67]
[19.59]
[19.80]
Instrumented by 1[YoB ≥ 1992]
[0.34]
[0.37]
[0.31]
As before, we explore the effects at further age increments in Table B.6 showing that the
negative effect of schooling on marriage is only significant at one age increment. Importantly,
the occurrence of teenage marriage was decreased for both the poor (panel C, column 4) and
wealthy women (panel C, column 6), by 16.7% and 11%, respectively. Given that wealthy
Notes: The results in each panel and column are produced by a separate regression. The IV estimate
and key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or
after the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the
reported Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights
and account for the pooling of survey rounds. The standard errors reported are clustered at the survey-
cluster level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes
below Table 3.2.
70
women delayed their first marriage but not their first birth, our results suggest that a reduction
in teenage marriage may not automatically lead to lower levels of (desired) fertility.
To further explore the effect of education on reproductive behavior, we investigate
women’s age at first sexual intercourse. Panel D shows that an additional year of education
makes women 7.6% less likely to report having had their first reported sexual intercourse before
the age of 20. Exploring estimates for all age increments shows that sexual exposure was
reduced also before the age of 18 (Table B.7). Given that all of our fertility-indicating variables
(including sexual activity, birth, and marriage) were significantly reduced in later teenage years,
i.e. years in which women were most likely already out of school, reduced fertility through
“incarceration” seems unlikely. Once more, the effects of increased education are heterogeneous
and only statistically significant for poorer women. Unaltered sexual exposure for the wealthy
could explain the gap between delayed marriage but unreduced fertility described above.
Income. Recent research reports inconclusive evidence concerning the support for an income
mechanism (Ali and Gurmu 2018; Makate and Makate 2018; Masuda and Yamauchi 2020;
Moussa and Omoeva 2020; Zenebe Gebre 2020). Only Chicoine (2021), Grépin and Bharadwaj
(2015), and Keats (2018) and find increases in (employed) work opportunities, a shift to higher-
skilled employment (contrary to agricultural work or similar), and higher remuneration.
Table 3.5 shows that policy-induced education did not increase the likelihood of having
any type of work or occupation different than domestic housework (panel A). However, added
education did seem to help women specifically into remunerated employment opportunities,
that is, significantly more women are now working outside of the household for cash/money
(panel B). The now common heterogeneity regarding poor and rich women emerges again: each
year of policy-induced schooling increased poor women’s remunerated employment
opportunities by 5.7 percentage-points, compared to a statistically insignificant 0.8 percentage-
point change for the wealthy. These findings lend support to an income pathway: More educated
(poor) women show for increased opportunity costs of childbearing, given by an increase in
remunerated employment opportunities. Higher opportunity costs are reflected in (poor)
women’s updated fertility preferences, which are adjusted accordingly (downwards).
Learning. The reported increase in women’s literacy and the accompanying decrease in their
desired number of children emphasize the potential significance of women’s learning on
subsequent fertility outcomes. This link is especially relevant for poorer women. To further
explore the workings of this channel, we first test women’s knowledge and usage of family
planning (condoms) in Table B.8 panel A, and panel B in the Appendix. While we do not find
an effect of increased awareness of modern contraceptives (panel A)12, education increases the
likelihood of having used a condom with the last sexual partner (panel B). Again, this effect is
12 General knowledge on condoms as a contraceptive method is already high, 94% for untreated cohorts, such
as any potential effect must be small.
71
Table 3.5: Second-Stage Results, Income
(1)
(2)
(3)
(4)
(5)
(6)
Panel A: Worked last Year {0,1}
Schooling
-0.003
-0.002
0.017
0.005
0.000
-0.002
(0.015)
(0.013)
(0.024)
(0.020)
(0.025) (0.022)
Mean of Control Group
F-Statistic
27.9
30.0
13.8
17.0
11.7
12.2
Observations
7,713
7,709
3,857
3,856
3,856
3,853
Panel B: Works outside of the Household for Cash/Money {0,1}
Schooling
0.023**
0.026**
0.058*
0.057**
0.007
0.008
(0.011)
(0.013)
(0.031)
(0.027)
(0.014) (0.015)
Mean of Control Group
F-Statistic
27.9
30.0
13.8
17.0
11.7
12.2
Observations
7,713
7,709
3,857
3,856
3,856
3,853
Controls
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
[0.86]
[0.92]
[0.78]
Instrumented by 1[YoB ≥ 1992]
[0.08]
[0.08]
[0.07]
Regression Sample
Full Sample
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
Table 3.6
likely driven by poor women. To assess the potential channels through which knowledge on
modern family norms or contraceptive methods is dispersed, we test for women’s engagement
with mass media outlets such as newspapers, radio, and television in Table B.9. We do not find
any robust indication of that more education also increased exposure to information via these
sources. Inasmuch as assortative mating, and thereby altered husband-wife relationships
mediate fertility outcomes, we test for the effect of education on husband’s age, education as
well as the difference in husbands’ and wives’ desired fertility. We find no statistically significant
effects for such augmentation to take place (Table B.10).
3.4.3. Further Evidence on the Heterogeneity
A salient feature distinguishing poor and wealthy subgroups within our main sample is their
educational exposure altogether and, correspondingly, the levels of schooling at which FPE-
induced educational gains were realized. This subsection explores this difference more
profoundly. Specifically, we investigate the grade levels at which the gains of education were
Notes: The results in each panel and column are produced by a separate regression. The IV estimate and
key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or after
the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the
reported Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights
and account for the pooling of survey rounds. The standard errors reported are clustered at the survey-
cluster level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes
below Table 3.2.
72
realized through FPE for our full sample, as well as for the poor and wealthy subgroups,
separately. Given that our IV-estimates resemble local average treatment effects (LATE), this
exercise informs us about the sub-population of “compliers” driving our results (see Acemoglu
and Angrist 2001; Angrist and Pischke 2009). Figure 3.3 depicts the shifts in grade levels
achieved due to FPE visually, plotting the difference in the conditional probability (on the y-
axis) of having completed at least a given school grade (on the x-axis) for women that were
just young enough to benefit from the policy change (13 years and younger), compared to
women who too old to benefit from the removal of fees (14 years and older). The differences
Figure 3.3: FPE-Induced Shifts in Grade Levels
depicted are conditional CDF changes, estimated from regressions including our usual
covariates and including survey weights.13 Note that we do not report significance levels for the
full sample, as they are significant at the 1% level (10% level) up until year 12 (year 13).
13 Note that the changes of the three different samples are normalized by their respective first-stages as outlined
in Angrist and Pischke (2009), giving us the contribution (weight) of the respective schooling level change
towards the average causal response over all educational levels.
Notes: The plot depicts differences in the conditional probability (on the y-axis) of having completed at least a
given school grade (on the x-axis) for women that were just young enough to benefit from the policy change (13
years and younger), compared to women who too old to benefit from the removal of fees (14 years and older). The
differences depicted are conditional CDF changes, estimated from regressions of having at least the respective school
level on our treatment indicator and our usual covariates. Note that we do not report significance levels for the full
sample, as they are significant at the 1% level (10% level) up until year 12 (year 13). Data come from the 2010/2011
and the 2016/2017 Burundi DHS Female Surveys. Respondents are restricted to age 20 and older. Estimates are
weighted using the DHS sample weights and account for the pooling of survey rounds.
73
The figure shows that for the poor, schooling increases induced by FPE were highest at
the (early) primary school level, increases for the wealthy were highest (and significant) at later
primary- to secondary schooling years: Being born in or after 1992 increased poor women’s
probability of having attained at least some primary school years (between 1-3 years) between
10.8 and 16.3 percentage-points, contrary to (statistically insignificant) increases between 2 and
5.3 percentage-points for the wealthy for these earlier school years. Increases for wealthier
women catch up to the level of poor women’s increases while moving rightwards along the x-
axis and exceed them at secondary school grades. This indicates that removing primary school
fees induced wealthier women to also transition into secondary school (e.g. see also Keats 2018).
Overall, the exercise shows that treatment groups across the poor and wealthy
subsamples differ systematically in their levels of added schooling induced by their exposure to
FPE. Burundian women from the wealthier strata attained higher grade levels independent of
the policy. For poor women, however, FPE mainly increased the probability to obtain lower
grade levels of schooling, which proves sufficient to induce behavioral changes.14 Reconciling
these insights with our main findings suggests that next to women’s household wealth, schooling
attained at lower grade levels may matter more in influencing women’s outcomes regarding
literacy, fertility and work than later schooling years, at least judging by our specific setting
for women exposed to FPE in Burundi.15
3.5 Robustness Checks and Validity Tests
The assumptions behind our IV-estimates require that women born (just) before, and women
born (just) after the cut-off do not systematically vary in characteristics influencing their
outcomes concerning fertility, literacy and work other than by their differential exposure to
schooling. We included control variables as well as fixed effects to address this issue in the first
instance and check these findings against a control selection algorithm (see PDS results in Table
B.1 through Table B.4 of the Appendix). However, given that we do not have information on
women at the time of policy implementation, there is a possibility of (further) unobserved
differences affecting our results.
We investigate the robustness and validity of our results in four distinct ways. (a) We
assimilate women’s circumstances at the time of the intervention using Burundi’s MICS survey
from 2005 and investigate the continuous nature of several influencing characteristics of
treatment and control cohorts, thereby measuring the potential influence of confounding
factors. (b) We re-assess our baseline results, specifically their heterogeneous nature of them,
14 This goes against findings prior to the year 2000 in which a handful of developing countries (including
Burundi) show a curvilinear relationship between schooling and fertility, meaning that the early years of
education actually increased women’s levels of childbearing . Note, however, that these findings were attributed
to a country’s very (early) status along the fertility transition for one (Martin 1995) and secondly, to
education’s differential impact on mediating factors such as contraceptive usage and breastfeeding (Jain 1981).
15 Our IV-estimates represent the local average treatment effect (LATE), which is specific to the setting and
sample at hand (Angrist et al. 1996).
74
again by moving closer to the intervention period, using the MICS surveys of 2000 as well as
2005 in a difference-in-differences framework. (c) We further validate our results by assessing
the impact of Burundi’s Civil War as a potential source of bias. (d) We analyze the robustness
of our estimates to varying bandwidths as well as functional forms and check our results against
placebo cut-offs.
Accounting for potential Confounders. DHS surveys do not include retrospective
information on respondents. To nonetheless explore the potentially discontinuous nature of
important covariates at the time of policy implementation, we utilize the Burundi Multiple
Indicator Cluster Survey (MICS) conducted in 2005 and identify household members born in
our treatment and control cohorts 1992-1996 and 1987-1991, respectively (UNICEF 2005).16
We test for the assumption of unconfoundedness by estimating equation (3.1), replacing
women’s schooling with various (household) characteristics of these earlier sampled individuals
as dependent variables. Although not identical to the respondents in the main sample,
individuals born within these years and surveyed in 2005 represent plausible proxy-respondents
corresponding to our treatment and control cohorts situated at the time of policy
implementation. To assess the potential influence of parents’ socio-economic status and their
child investments, we also test for discontinuities in characteristics of “potential mothers” (see
Keats 2018) by identifying women in the 2005 MICS who have had births between the years
1987-1996. All of the results are presented in Table B.11 in the Appendix.
The estimates presented show statistically insignificant differences such that there are no
apparent discontinuities between proxy treatment and proxy control cohorts at the time of
policy implementation. It, therefore, seems unlikely that there were systematic differences
influencing the eventual outcomes of sample respondents today, which are conflated with their
differential access to schooling. Our results hold when dropping observations to include only
female household members. They also hold when testing for a differential discontinuous effect
among poor and wealthy subgroups separately.
Wealth Grouping. A potential concern regarding the grouping of women by wealth in our
analysis is that women increased their (household) wealth due to policy-induced schooling. To
directly test for this possibility, we replace our outcome variable in equation (3.2) with women’s
DHS wealth score, as well as several other indicators of wealth such as household’s access to
electricity, piped water, a flush toilet and several consumer durables such as a tv, radio and
bicycle. The results are presented in Table B.12 and show no robust evidence for a direct
(causal) effect of FPE-induced schooling on women’s household wealth. Moreover, more than
half of the women in our treated sample are still considered dependents, i.e. are neither the
head of the household (8%) nor the wife to the head (40%). Thus, household wealth is unlikely
to be affected by them and may be considered as given. Also note that if women’s additional
16 There are no DHS surveys available in Burundi between 1987 and 2010.
75
schooling did quickly allow them to move into the wealthier subgroup of the population, our
estimated coefficients would be biased downwards, representing conservative estimates of the
true effect of schooling on the poor.
We additionally test for potentially differing exposure as well as treatment between poor
and wealthy subgroups, again by moving closer to the time of policy intervention. We thereby
pool the cross-sectional Multiple Indicator Cluster Surveys (MICS) of 2000 and 2005 together
with the 2010/2011 DHS Survey and conduct a difference-in-difference-in-differences (triple-
difference) analysis effectively analyzing the difference of being wealthy or poor between treated
(policy-eligible) and untreated women before and after the policy (adopted from Keats 2018).
Specifically, we compare outcomes for women born in 1992-1996 and 1987-1991, i.e. aged 15-19
and 20-24 in the 2010/2011 Burundi DHS Female Survey (first difference). The difference in
outcomes between these two groups is then adjusted by a secular time-trend, given by the
difference between these two age cohorts before policy implementation, i.e. in 2000 and 2005
using the MICS Female Surveys (second difference). Comparing these differences both for
women from poor and from wealthy households (third difference) leads us to estimate the
following triple-difference specification:
𝑌𝑌𝑖𝑖,𝑟𝑟,𝑡𝑡= 𝛽𝛽0 + 𝛽𝛽1(𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖× 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡) + 𝛽𝛽2(𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖× 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡× 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖) + 𝛽𝛽3(𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖× 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖)
+ 𝛽𝛽4(𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡× 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖) + 𝛽𝛽5𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖+ 𝛽𝛽6Poor𝑖𝑖+ 𝛽𝛽𝑘𝑘𝑋𝑋𝑖𝑖+ 𝛾𝛾𝑡𝑡+ 𝛿𝛿𝑟𝑟+ 𝑢𝑢𝑖𝑖,𝑟𝑟,𝑡𝑡
(3.3)n
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 is a dummy for individuals aged 15-19, 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡 indicates respondents sampled in the
survey period 2010/2011 and 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖 switches on for women living in households below the
median wealth level. 𝛾𝛾𝑡𝑡 and 𝛿𝛿𝑟𝑟 represent survey and region fixed effects, respectively. We also
include controls for women’s age as well as the type of residence (urban/rural) and cluster
standard errors both at the survey-cluster and birth year level. Results for educational
attainment, literacy, fertility choices (concerning their last-born child) as well as household
wealth are depicted in Figure 3.4 which plots coefficient 𝛽𝛽1, as well as the linear combination
of 𝛽𝛽1 + 𝛽𝛽2 together with the corresponding 95% confidence intervals, allowing for a direct
difference-in-differences interpretation and comparison for both wealthy and poor subgroups.
The results show that educational attainment (measured in schooling levels) was
positively influenced for both wealthy and poor women, the poor showing higher absolute
increases in schooling, making it broadly consistent with our main results. The rates of literacy
were influenced analogously, and women report wanting to have their last-born child later in
life, the result only significant for the poor. Importantly, household wealth outcomes of treated
individuals were not significantly influenced by FPE. Figure B.4 in the Appendix provides
further evidence. In line with our main results, we do not find evidence of altered marriage or
fertility outcomes at these younger ages (15-19). Moreover, we perform a “placebo” analysis,
76
where we estimate the same triple-difference analysis, albeit measuring these changes from 2000
(pre) to 2005 (post), a non-intervention period.17 These results show statistically
indistinguishable differences for all tested outcomes between the two groups over time, both
for wealthy and poor women, as should be expected for a “placebo” test.
Figure 3.4: Difference-in-Difference-in-Differences
As a further test for the robustness of our heterogeneous effects, in particular, their
sensitivity to differing wealth groupings, we re-categorize individuals in our main analysis
“poor” when living in households where the main flooring consists of sand, dung or dirt (see
Keats 2018). Results are given in Table B.13, and again support our interpretations. Note that
poor women tend to reside in rural areas (see Table 3.1), which is why we control for women’s
17 A question on fertility preference was not asked in the 2000 MICS, which is why there is no placebo-estimate
on this outcome.
Notes: The “Point Estimates” represent difference-in-differences coefficients, including a 95% confidence interval,
for the respective dependent variable indicated on the y-axis. Data come from the 2000 and 2005 Multiple Cluster
Indicator Surveys (MICS) as well as the 2010/2011 Burundi DHS survey, using the women’s recodes. The
coefficients are estimated through “triple-difference” estimation (3.3), comparing individuals aged 15-19 to
individuals aged 20-24 in 2000, 2005 and 2010 for both “wealthy” and for “poor” households (splitting at the
median household wealth level). The estimates are weighted using MICS/DHS sample weights and account for
the pooling of sample rounds. Standard errors are clustered at the birth-year and enumeration-area level.
77
urban/rural status in all estimations. Importantly, our results also hold when conducting the
(wealthy vs. poor) subsample analysis among rural regions only.18
Burundian Civil War. The massacres of 1993 and the subsequent civil war was the most
violent period in Burundi’s history, with an estimated total of 300,000 deaths (UNFPA 2002).
Conflicts were most intense from 1993 to 1998, the years thereafter outlined by significantly
lower conflict intensities (Mercier et al. 2020).
The massacres and the civil war in Burundi have been linked to negative effects on
households’ wealth, children’s health, and children’s likelihood of completing primary school
(Bundervoet et al. 2009; Verwimp and Van Bavel 2014; Mercier et al. 2020).19 Depending on
the investigated outcome, these studies define individuals affected by the war born between
1981-1998, which includes both our treatment and our control cohorts born between 1987 and
1996 such that there is no differential effect to be expected of the war itself between individuals
sorted into treatment and control. And although conflict intensities have varied, given the
timeline and the spatial dispersal of conflict, effects are unlikely to systematically change for
women born just after our chosen cut-off year in 1992. In other words, being 13 in 2005 should
carry no special property for women but our identification in their differential access to
schooling given by the timing of Burundi’s FPE. Even if other peace-building policies by the
newly formed government coincide with FPE’s implementation year of 2005, it is unlikely for
these measures to exert their effect in a sharp, discontinuous fashion starting with girls 13 years
or younger only, i.e. such peace-building policies would affect all age groups from this point in
time onward.
If there were conflict-induced changes affecting women’s eventual fertility, literacy or
work outcomes differently across treatment and control, it is sensible to think that the
discontinuous nature of these outcomes should be apparent at “cut-off” years other than 1992,
too. To investigate this possibility, we re-estimate equation (3.1), replacing women’s schooling
with all of our main outcome variables, and move our specified window size of five years around
the corresponding placebo cut-off years starting from 1981 up until 1995. The results are plotted
in Figure B.5 through Figure B.12. Across all outcome variables tested, there are no
discontinuities apparent in other years but at our specified cut-off date, as we would expect if
all effects were induced by FPE. As has also been shown in panel a) in Table B.11, there are
no significant discontinuities in children’s circumstances at the time of the policy
implementation for our defined treatment cohort. Specifically, treated cohorts do not show a
18 We are unable to conduct (poor vs. wealthy) subsample estimations in urban regions, as there are too few
(poor) individuals living in urban environments altogether.
19 Note that the negative effects on schooling were significantly smaller for women (Verwimp and Van Bavel
2014). Research on the impacts of similar civil wars, for example, in neighboring Rwanda has confirmed the
negative impact on schooling (and the lesser negative impact on women) at least in the short- to medium
term (Akresh and de Walque 2008; Bundervoet and Fransen 2018; La Mattina 2018) as well as negative effects
on fertility outcomes (Kraehnert et al. 2019).
78
higher likelihood of having a living mother or a living father, which acts as a proxy for the
war’s intensity (Bundervoet 2009), nor having better maternal care or improved health inputs
as a child.
Our findings are also robust to a varying bandwidth (see below), specifically to a smaller
one. If conflict affected our treatment and control cohorts differently, estimates calculated from
women born only a couple of birth years apart would likely cause biased effects to vanish.
However, all of our effects remain significant at low bandwidth sizes, especially in the full
sample. Hence, for our analysis of the impact of schooling induced by FPE, it is plausible to
assume that treatment and control groups were similarly affected by the civil war, or other
influencing factors following the war, like peace-building policies, such that we are confident in
consistently identifying the isolated effect of education with our estimation strategy.
To nonetheless test for the potential influence of differing war exposure, we remove
provinces with above median conflict intensities from our sample and re-estimate our main
outcome variables (Table B.14). We use data from Bundervoet (2009), who calculates conflict
intensity as the estimated percentage of a parent (mother or father) being killed in the
massacres of 1993 using historical records from Chrétien and Mukuri (2002).20 We subsequently
drop provinces from our sample with above median conflict intensities, i.e. where above 6.6%
of mothers or fathers are estimated to have been killed in the massacres. To compare, the mean
conflict intensity across provinces is 9%, the upper quartile of provinces had over 16% of
parental losses, and individuals in the highest affected province had an occurrence of over 22%
parental deaths. Our interpretations remain largely unchanged and robust when removing these
conflict-intense provinces.21
Varying Bandwidth and Functional Form. To investigate the sensibility of the estimated
first-stage discontinuity as well as our second-stage results to a varying bandwidth, we re-
estimate our main results using window sizes ranging from 2 to 10 birth year cohorts. Figure
B.13 and Figure B.14 plot the estimated first-stage discontinuities for each bandwidth size
together with the 95% confidence interval, each for the full sample, and separated along
household wealth. We find significant increases in schooling from a bandwidth size of as little
as two (three for the subgroup estimates) birth year cohorts on each side of the cut-off. The
estimates stabilize from a bandwidth size of 5 onwards. Concerning the second-stage estimates,
Figure B.15 through Figure B.22 provide evidence for a robust effect of education on fertility
and associated factors regarding bandwidth size, and especially, concerning the heterogeneous
nature of our main results. Estimates for the poorer subgroup are generally larger and more
20 This may not fully account for provincial intensities of subsequent civil war years. Thus, we also explored
the results by leaving out the most affected provinces of the first years of the civil war as defined by
Bundervoet, Verwimp, and Akresh (2009). The results hold when excluding these regions.
21 We do not report separate effects for poor and wealthy subsamples, given that the instrument strength for
both subgroups decreased significantly caused by the stark drop of the sample size.
79
statistically significant at all data window sizes. This is especially true at smaller bandwidth
sizes, which presumably, present less potential bias in our estimates given that longer-lived
time trends are less apparent when comparing women born only a couple of years apart.
Next to the bandwidth choice, we test for sensitivity due to changing functional form as
a further investigation into the robustness of our results. We adjust the functional form of
equations (3.1) and (3.2) by adding higher-order polynomials of the running variable (year of
birth), alternating the interaction of them with the treatment indicator (1[YOB≥1992]). We
stick to our selected bandwidth size of five birth year cohorts on each side of the cut-off. The
results are presented in Table B.15 and Table B.16 of the Appendix for first- and second-stage
estimates, respectively. The estimate of the discontinuity remains significant throughout all
common RDD specifications, apart from specification (f). The size of the significant,
discontinuous effect varies from 1.170 to 1.498 years of education for the full sample. These
first-stage estimates are carried over to the second-stage in Table B.16, providing an interesting
insight into the heterogeneous nature of our results. Throughout the specifications, the main
estimates remain broadly robust in size and in significance. With increasing terms added to the
equation, the IV-estimate tends to drop in precision (note also the dwindling F-Statistic as can
be inferred from the estimates in Table B.16), losing statistical significance for most estimates
from specification (d) and onwards. It is noteworthy that estimates generated from the sample
of poor women are more robust than the ones estimated from the wealthier half of the sample.
Up to the quadratic specification (c), five of the seven usually significant coefficients remain
(three of them at the 5% level) for the poor subgroup of women, whereas none of the estimates
of the wealthy subgroup are significant.
3.6 Conclusion
This article investigates the effect of female education on fertility by exploiting Burundi’s free
primary education policy (FPE) of 2005 as a natural experiment. We identify exogenous
increases in schooling through a fuzzy regression discontinuity design (RDD). Subsequently, we
conduct instrumental variable estimations examining the causal effect of schooling on
adolescent fertility outcomes as well as on commonly associated mechanisms.
Our results show that Burundi’s FPE policy increased educational outcomes for women
regardless of their socio-economic standing. Being young enough to benefit from free primary
schooling (age 13 or younger) increased women’s educational attainment by a substantial 1.22
years. Employing policy-induced variation in individual education as an instrument, we find
that one year of additional education delayed women’s first birth by half a year, and overall,
decreased their probability of having their first childbirth before the age of 20. Investigating
the mechanisms through which education is theorized to reduce women’s fertility, we find
supportive evidence for both learning and income pathways. Additional schooling led to higher
literacy levels, decreased desired family sizes, and improved paid labor market opportunities.
80
Numerous robustness and sensitivity tests support our findings and interpretations.
Our estimates thereby confirm and complement the established negative causal effect of
schooling on teenage fertility documented in the literature and provide further support for
closely related mechanisms. What specifically distinguishes our results from the ones found in
recent literature is the marked heterogeneity in the downstream effects of policy-induced
schooling. While both poor and wealthy women gain additional years of schooling due to
Burundi’s FPE, the outcomes affected via policy-induced education are mainly driven by
behavioral changes of the poor. Effects of additional schooling may therefore be expected to
materialize differently depending on the socio-economic background of women, which
complements the existing knowledge base. This insight holds relevant lessons when envisaging
future educational interventions.
81
3.7 References
Acemoglu, Daron, and Joshua D. Angrist. 2001. ‘How Large Are Human-Capital Externalities?
Evidence from Compulsory-Schooling Laws’. In NBER Macroeconomics Annual 2000,
edited by Ben Bernanke, Kenneth Rogoff, and National Bureau of Economic Research.
Cambridge, Mass.: MIT Press.
Adu-Boahen, Emmanuel, and Chikako Yamauchi. 2018. ‘The Effect of Female Education on
Adolescent Fertility and Early Marriage: Evidence from Free Compulsory Universal
Basic Education in Ghana’. Journal of African Economies 27 (2): 227–48.
Ainsworth, Martha, Kathleen Beegle, and Andrew Nyamete. 1996. ‘The Impact of Women’s
Schooling on Fertility and Contraceptive Use: A Study of Fourteen Sub-Saharan
African Countries’. The World Bank Economic Review 10 (1): 85–122.
Akresh, Richard, and Damien de Walque. 2008. ‘Armed Conflict and Schooling: Evidence from
the 1994 Rwandan Genocide’. World Bank Policy Research Working Paper 4606, 38.
Ali, Fatma Romeh M., and Shiferaw Gurmu. 2018. ‘The Impact of Female Education on
Fertility: A Natural Experiment from Egypt’. Review of Economics of the Household
16 (3): 681–712.
Angrist, Joshua D., Guido W. Imbens, and Donald B. Rubin. 1996. ‘Identification of Causal
Effects Using Instrumental Variables’. Journal of the American Statistical Association
91 (434): 444–55.
Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An
Empiricist’s Companion. Princeton, NJ: Princeton University Press.
Bankole, Akinrinola, Fatima H. Ahmed, Stella Neema, Christine Ouedraogo, and Sidon
Konyani. 2007. ‘Knowledge of Correct Condom Use and Consistency of Use among
Adolescents in Four Countries in Sub-Saharan Africa’. African Journal of Reproductive
Health 11 (3): 197–220.
Becker, Gary S. 1960. ‘An Economic Analysis of Fertility’. In Demographic and Economic
Change in Developed Countries, edited by Universities-National Bureau, 209–40. New
York: National Bureau of Economic Research, distributed by Columbia University
Press.
———. 1965. ‘A Theory of the Allocation of Time’. The Economic Journal 75 (299): 493–517.
Becker, Gary S., and H. Gregg Lewis. 1973. ‘On the Interaction between the Quantity and
Quality of Children’. Journal of Political Economy 81 (2, Part 2): S279–88.
Behrman, Jere R., and Mark R. Rosenzweig. 2002. ‘Does Increasing Women’s Schooling Raise
the Schooling of the Next Generation?’ American Economic Review 92 (1): 323–34.
Behrman, Julia A. 2015a. ‘The Effect of Increased Primary Schooling on Adult Women’s HIV
Status in Malawi and Uganda: Universal Primary Education as a Natural Experiment’.
Social Science & Medicine 127 (February): 108–15.
———. 2015b. ‘Does Schooling Affect Women’s Desired Fertility? Evidence From Malawi,
Uganda, and Ethiopia’. Demography 52 (3): 787–809.
Belloni, Alexandre, Victor Chernozhukov, and Christian Hansen. 2014. ‘Inference on
Treatment Effects after Selection among High-Dimensional Controls’. The Review of
Economic Studies 81 (2): 608–50.
Berthelon, Matias E., and Diana I. Kruger. 2011. ‘Risky Behavior among Youth: Incapacitation
Effects of School on Adolescent Motherhood and Crime in Chile’. Journal of Public
Economics 95 (1–2): 41–53.
Black, Sandra E., Paul J. Devereux, and Kjell G. Salvanes. 2008. ‘Staying in the Classroom
and out of the Maternity Ward? The Effect of Compulsory Schooling Laws on Teenage
Births*’. The Economic Journal 118 (530): 1025–54.
82
Bold, Tessa, Mwangi S. Kimenyi, and Justin Sandefur. 2013. ‘Public and Private Provision of
Education in Kenya’. Journal of African Economies 22 (suppl 2): ii39–56.
Bongaarts, John, and Susan Cotts Watkins. 1996. ‘Social Interactions and Contemporary
Fertility Transitions’. Population and Development Review 22 (4): 639–82.
Bundervoet, Tom. 2009. ‘Livestock, Land and Political Power: The 1993 Killings in Burundi’.
Journal of Peace Research 46 (3): 357–76.
Bundervoet, Tom, and Sonja Fransen. 2018. ‘The Educational Impact of Shocks in Utero:
Evidence from Rwanda’. Economics & Human Biology 29: 88–101.
Bundervoet, Tom, Philip Verwimp, and Richard Akresh. 2009. ‘Health and Civil War in Rural
Burundi’. Journal of Human Resources 44 (2): 536–63.
Burlando, Alfredo, and Edward Bbaale. 2022. ‘Fertility Responses to Schooling Costs: Evidence
from Uganda’s Universal Primary Education Policy’. Economic Development and
Cultural Change 70 (3): 1017–39.
Caldwell, John C. 1976. ‘Toward A Restatement of Demographic Transition Theory’.
Population and Development Review 2 (3/4): 321–66.
———. 1980. ‘Mass Education as a Determinant of the Timing of Fertility Decline’. Population
and Development Review 6 (2): 225–55.
Chicoine, Luke. 2021. ‘Free Primary Education, Fertility, and Women’s Access to the Labor
Market: Evidence from Ethiopia’. The World Bank Economic Review 35 (2): 480–98.
Chrétien, Jean-Pierre, and Melchior Mukuri, eds. 2002. Burundi, La Fracture Identitaire:
Logiques de Violence et Certitudes Ethniques, 1993-1996. Hommes et Sociétés. Paris:
Karthala.
Cochrane, Susan Hill. 1979. Fertility and Education: What Do We Really Know? World Bank
Staff Occasional Papers 26. Baltimore: Published for the World Bank [by] Johns
Hopkins University Press.
Croft, Trevor N., Aileen M.J. Marshall, and Courtney K. Allen. 2018. ‘Guide to DHS
Statistics’. Rockville, Maryland, USA: ICF International.
Deininger, Klaus. 2003. ‘Does Cost of Schooling Affect Enrollment by the Poor? Universal
Primary Education in Uganda’. Economics of Education Review 22 (3): 291–305.
Dunlop, Emily, and Elisabeth King. 2019. ‘Education at the Intersection of Conflict and Peace:
The Inclusion and Framing of Education Provisions in African Peace Agreements from
1975-2017’. Compare: A Journal of Comparative and International Education 0 (0): 1–
21.
Evans, David K., and Amina Mendez Acosta. 2021. ‘Education in Africa: What Are We
Learning?’ Journal of African Economies 30 (1): 13–54.
Ewemooje, Olusegun S., Elizabeth Biney, and Acheampong Yaw Amoateng. 2020.
‘Determinants of Fertility Intentions among Women of Reproductive Age in South
Africa: Evidence from the 2016 Demographic and Health Survey’. Journal of Population
Research 37 (June): 265–89.
Ferré, Céline. 2009. ‘Age At First Child: Does Education Delay Fertility Timing? The Case Of
Kenya’. In Policy Research Working Paper. WPS 4833. The World Bank.
Gelman, Andrew, and Guido Imbens. 2019. ‘Why High-Order Polynomials Should Not Be Used
in Regression Discontinuity Designs’. Journal of Business & Economic Statistics 37 (3):
447–56.
Glewwe, Paul. 1999. ‘Why Does Mother’s Schooling Raise Child Health in Developing
Countries? Evidence from Morocco’. The Journal of Human Resources 34 (1): 124–59.
Grant, Monica J. 2015. ‘The Demographic Promise of Expanded Female Education: Trends in
the Age at First Birth in Malawi’. Population and Development Review 41 (3): 409–38.
Grépin, Karen A., and Prashant Bharadwaj. 2015. ‘Maternal Education and Child Mortality
in Zimbabwe’. Journal of Health Economics 44 (December): 97–117.
83
Grogan, Louise. 2008. ‘Universal Primary Education and School Entry in Uganda’. Journal of
African Economies 18 (2): 183–211.
Grossman, Michael. 1972. ‘On the Concept of Health Capital and the Demand for Health’.
Journal of Political Economy 80 (2): 223–55.
Hahn, Jinyong, Petra Todd, and Wilbert Klaauw. 2001. ‘Identification and Estimation of
Treatment Effects with a Regression-Discontinuity Design’. Econometrica 69 (1): 201–
9.
Ilahi, Nadeem. 2000. ‘The Intra-Household Allocation of Time and Tasks: What Have We
Learnt from the Empirical Literature?’ In Policy Research Report on Gender and
Development, 48. Working Paper Series No. 13. Washington, D.C.: The World Bank.
Imbens, Guido W., and Karthik Kalyanaraman. 2012. ‘Optimal Bandwidth Choice for the
Regression Discontinuity Estimator’. The Review of Economic Studies 79 (3): 933–59.
Imbens, Guido W., and Thomas Lemieux. 2008. ‘Regression Discontinuity Designs: A Guide
to Practice’. Journal of Econometrics 142 (2): 615–35.
Jain, Anrudh K. 1981. ‘The Effect of Female Education on Fertility: A Simple Explanation’.
Demography 18 (4): 577–95.
Kattan, Raja Bentaouet. 2006. ‘Implementation of Free Basic Education Policy’. 47186. The
World Bank.
Keats, Anthony. 2018. ‘Women’s Schooling, Fertility, and Child Health Outcomes: Evidence
from Uganda’s Free Primary Education Program’. Journal of Development Economics
135: 142–59.
Kraehnert, Kati, Tilman Brück, Michele Di Maio, and Roberto Nisticò. 2019. ‘The Effects of
Conflict on Fertility: Evidence From the Genocide in Rwanda’. Demography 56 (3):
935–68.
La Mattina, Giulia. 2018. ‘How Persistent Is the Effect of Conflict on Primary Education?
Long-Run Evidence from the Rwandan Genocide’. Economics Letters 163 (February):
32–35.
Lam, David, and Suzanne Duryea. 1999. ‘Effects of Schooling on Fertility, Labor Supply, and
Investments in Children, with Evidence from Brazil’. The Journal of Human Resources
34 (1): 160.
Lee, David S., and Thomas Lemieux. 2010. ‘Regression Discontinuity Designs in Economics’.
Journal of Economic Literature 48 (2): 281–355.
Levine, Ruth, Cynthia, Lloyd, Margaret Green, and Caren Grown. 2009. Girls Count: A Global
Investment & Action Agenda. Washington, D.C.: Center for Global Development.
Lloyd, Cynthia B, and Juliet Young. 2009. ‘New Lessons: The Power of Educating Adolescent
Girls’. Girls Count. New York: Population Council: Coalition for Adolescent Girls.
Lucas, Adrienne M., and Isaac M. Mbiti. 2012. ‘Access, Sorting, and Achievement: The Short-
Run Effects of Free Primary Education in Kenya’. American Economic Journal:
Applied Economics 4 (4): 226–53.
Ludwig, Jens, and Douglas L. Miller. 2005. ‘Does Head Start Improve Children’s Life Chances?
Evidence from a Regression Discontinuity Design’. In NBER Working Paper Series
11702. Cambridge, MA: National Bureau of Economic Research.
Makate, Marshall, and Clifton Makate. 2016. ‘The Causal Effect of Increased Primary
Schooling on Child Mortality in Malawi: Universal Primary Education as a Natural
Experiment’. Social Science & Medicine 168 (November): 72–83.
———. 2018. ‘Education and Teenage Childbirth in Uganda: Understanding the Links from
the Uganda Demographic and Health Survey’. International Journal of Social
Economics 45 (5): 746–64.
Mare, Robert D. 1991. ‘Five Decades of Educational Assortative Mating’. American
Sociological Review 56 (1): 15–32.
84
Martin, Teresa Castro. 1995. ‘Women’s Education and Fertility: Results from 26 Demographic
and Health Surveys’. Studies in Family Planning 26 (4): 187–202.
Martínez, Elin, and Agnes Odhiambo. 2018. Leave No Girl behind in Africa - Discrimination
in Education against Pregnant Girls and Adolescent Mothers. New York, NY: Human
Rights Watch.
Masuda, Kazuya, and Chikako Yamauchi. 2020. ‘How Does Female Education Reduce
Adolescent Pregnancy and Improve Child Health?: Evidence from Uganda’s Universal
Primary Education for Fully Treated Cohorts’. The Journal of Development Studies 56
(1): 63–86.
McCrary, Justin, and Heather Royer. 2011. ‘The Effect of Female Education on Fertility and
Infant Health: Evidence from School Entry Policies Using Exact Date of Birth’.
American Economic Review 101 (1): 158–95.
Mercier, Marion, Rama L. Ngenzebuke, and Philip Verwimp. 2020. ‘Violence Exposure and
Poverty: Evidence from the Burundi Civil War’. Journal of Comparative Economics 48
(4): 822–40.
Michael, Robert T. 1973. ‘Education and the Derived Demand for Children’. Journal of
Political Economy 81 (2): 128–64.
Mincer, Jacob. 1962. ‘Labor Force Participation of Married Women: A Study of Labor Supply’,
45.
———. 1963. ‘Market Prices, Opportunity Costs, and Income Effects’. In Measurement in
Economics: Studies in Mathematical Economics and Econometrics in Memory of
Yehuda Grunfield, 67–82. Stanford, CA: Carl F. Christ, Stanford University Press.
Ministere e la Presidence charge de la Bonne Gouvernance et du Plan - MPBGP, Ministere de
la Sante Publique et de la Lutte contre le Sida - MSPLS, Institut de Statistiques et
d’etudes economiques du Burundi - ISTEEBU, and ICF. 2017. Burundi Troisieme
Enquete Demographique et de Sante 2016-2017. Bujumbura, Burundi: MPBGP,
MSPLS, ISTEEBU, and ICF.
Moussa, Wael, and Carina Omoeva. 2020. ‘The Long-Term Effects of Universal Primary
Education: Evidence from Ethiopia, Malawi, and Uganda’. Comparative Education
Review 64 (2): 179–206.
Obura, Anna P. 2008. Staying Power: Struggling to Reconstruct Education in Burundi since
1993. Paris: International Institute for Educational Planning.
Osili, Una Okonkwo, and Bridget Terry Long. 2008. ‘Does Female Schooling Reduce Fertility?
Evidence from Nigeria’. Journal of Development Economics 87 (1): 57–75.
Ozier, Owen. 2018. ‘The Impact of Secondary Schooling in Kenya: A Regression Discontinuity
Analysis’. Journal of Human Resources 53 (1): 157–88.
Presler-Marshall, Elizabeth, and Nicola Jones. 2012. ‘Charting the Future: Empowering Girls
to Prevent Early Pregnancy’. London: Overseas Development Institute.
Psaki, Stephanie R., Erica K. Chuang, Andrea J. Melnikas, David B. Wilson, and Barbara S.
Mensch. 2019. ‘Causal Effects of Education on Sexual and Reproductive Health in Low
and Middle-Income Countries: A Systematic Review and Meta-Analysis’. SSM -
Population Health 8.
Rosenzweig, Mark R., and T. Paul Schultz. 1989. ‘Schooling, Information and Nonmarket
Productivity: Contraceptive Use and Its Effectiveness’. International Economic Review
30 (2): 457–77.
Rutstein, Shea O., and Kiersten Johnson. 2004. ‘The DHS Wealth Index’. In DHS Comparative
Reports No. 6. Calverton, Maryland, USA: Macro International.
Staiger, Douglas, and James H. Stock. 1997. ‘Instrumental Variables Regression with Weak
Instruments’. Econometrica 65 (3): 557–86.
The DHS Program. 2015. ‘Demographic and Health Surveys Model Questionnaire - Woman’.
85
Thomas, Duncan, John Strauss, and Maria-Helena Henriques. 1991. ‘How Does Mother’s
Education Affect Child Height?’ The Journal of Human Resources 26 (2): 183–211.
Travaglianti, Manuela. 2017. ‘How Abolishing School Fees Increased Support for the
Incumbent in Burundi’. African Affairs 116 (462): 101–24.
UIS. 2022. ‘UIS Statistics’. UNESCO Institute for Statistics. 2022.
UNFPA, United Nations Population Fund. 2002. ‘Situation De´mographique et Sociale Au
Burundi. Re´sultats de l’Enqueˆte Sociode´mographique et de Sante´ de La
Reproduction’. Burundi: De´partement de la Population.
UNICEF. 2005. ‘Multiple Indicator Cluster Surveys (MICS)’.
United Nations, ed. 1995. Report of the International Conference on Population and
Development: Cairo, 5-13 September 1994. New York: United Nations.
———. 2001. ‘Road Map towards the Implementation of the United Nations Millennium
Declaration: Report of the Secretary-General.’ New York: United Nations Department
of Public Information.
———. 2022. ‘World Population Prospects: The 2022 Revision’. United Nations, Department
of Economic and Social Affairs, Population Division.
Verwimp, Philip, and Jan Van Bavel. 2014. ‘Schooling, Violent Conflict, and Gender in
Burundi’. The World Bank Economic Review 28 (2): 384–411.
Wild, Frederik, and David Stadelmann. 2024. ‘Heterogeneous Effects of Women’s Schooling on
Fertility, Literacy and Work: Evidence from Burundi’s Free Primary Education Policy’.
Journal of African Economies 33 (1): 67–91.
Willis, Robert J. 1973. ‘A New Approach to the Economic Theory of Fertility Behavior’.
Journal of Political Economy 81 (2, Part 2): S14–64.
World Bank. 2017. World Development Report 2018: Learning to Realize Education’s Promise.
The World Bank.
———. 2021. World Development Indicators. Washington, D.C.: The World Bank.
Zenebe Gebre, Tihtina. 2020. ‘Free Primary Education, Timing of Fertility, and Total
Fertility’. The World Bank Economic Review 34 (3): 730–48.
86
87
Appendix B
Heterogeneous Effects of Women’s Schooling on
Fertility, Literacy and Work: Evidence from
Burundi’s Free Primary Education Policy
88
Table B.1: RD Estimates, Years of Schooling (Controls following Post-Double-Selection)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
1[Year of Birth ≥ 1992]
1.344*** 1.223*** 1.266***
0.849*** 0.908*** 0.893***
1.356*** 1.280*** 1.357***
(0.255)
(0.223)
(0.211)
(0.228)
(0.220)
(0.210)
(0.397)
(0.367)
(0.362)
Mean of Control Group
Interaction: 1 × Year of Birth
0.135
0.164**
0.081
0.378*** 0.339*** 0.196***
-0.179
-0.099
-0.123
(0.090)
(0.077)
(0.073)
(0.081)
(0.078)
(0.073)
(0.141)
(0.130)
(0.121)
Year of Birth
0.052
0.038
-0.530**
0.059
0.064
0.098
0.116
0.070
-0.882**
(0.054)
(0.047)
(0.226)
(0.046)
(0.044)
(0.228)
(0.090)
(0.085)
(0.360)
Effect on the second cohort treated by FPE
1[YoB ≥ 1992] + 1 × YoB + YoB
1.426***
1.312***
1.250***
(0.197)
(0.187)
(0.338)
Observations
7,713
7,709
7,263
3,857
3,856
3,783
3,856
3,853
3,480
Adj. R-squared
0.054
0.263
-
0.099
0.183
-
0.041
0.187
-
Controls (Post Double Selection)
NO
YES
(28)
NO
YES
(25)
NO
YES
(23)
Province Fixed Effects
NO
YES
PDS
NO
YES
PDS
NO
YES
PDS
Survey Fixed Effects
YES
YES
PDS
YES
YES
PDS
YES
YES
PDS
Regression Sample
Full
Poor
Wealthy
[3.55]
[1.99]
[5.56]
Notes: The results in each column are produced by a separate regression. Columns (3), (6) and (9) report results using
the Post Double Selection (PDS) algorithm developed in Belloni et al. (2014), considering a total of 896 potential
controls. The included number of PDS’ selected controls are given in parenthesis at the bottom of the table. Estimates
are weighted using the DHS sample weights and account for the pooling of survey rounds. The standard errors reported
are clustered at the survey-cluster level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
See full notes below Table 3.2.
89
Table B.2: Second-Stage Results, Literacy and Desired Fertility (Post-Double-Selection)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Panel A: Literacy {0,1}
Schooling
0.054*** 0.050*** 0.059***
0.067** 0.067**
0.072**
0.033
0.027
0.037
(0.015)
(0.017)
(0.016)
(0.033)
(0.030)
(0.029)
(0.022) (0.024)
(0.023)
Mean of Control Group
F-Statistic
27.7
29.9
35.5
13.8
17.1
18.7
11.4
11.9
13.9
Observations
7,708
7,704
7,258
3,856
3,855
3,782
3,852
3,849
3,476
Panel B: Desired Number of Children
Schooling
-0.144** -0.151** -0.153**
-0.307** -0.261** -0.281**
-0.093
-0.088
-0.079
(0.060)
(0.063)
(0.061)
(0.136)
(0.118)
(0.119)
(0.086) (0.086)
(0.079)
Mean of Control Group
F-Statistic
25.9
28.3
34.9
12.5
15.8
17.5
11.4
11.9
14.7
Observations
7,520
7,516
7,087
3,764
3,763
3,691
3,756
3,753
3,396
Controls (Post Double Selection)
NO
YES
(34,36)
NO
YES
(27,27)
NO
YES
(23,26)
Province Fixed Effects
NO
YES
PDS
NO
YES
PDS
NO
YES
PDS
Survey Fixed Effects
YES
YES
PDS
YES
YES
PDS
YES
YES
PDS
Regression Sample
Full
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
[4.01]
[4.05]
[3.96]
[0.56]
[0.43]
[0.73]
Notes: The results in each panel and column are produced by a separate regression. Columns (3), (6) and (9) report
results using the Post Double Selection (PDS) algorithm developed in Belloni et al. (2014), considering a total of 896
potential controls. The included number of PDS’ selected controls are given in parenthesis at the bottom of the table.
The IV-estimate and key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or
after the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported
Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for the pooling
of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents significance
at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
90
Table B.3: Second-Stage Results, Fertility and Reproductive Behavior (Post-Double-
Selection)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Panel A: Age at First Birth
Schooling
0.440**
0.468**
0.511***
0.435
0.449*
0.516**
0.640*
0.757
0.686
(0.179)
(0.204)
(0.198)
(0.274)
(0.263)
(0.245)
(0.381)
(0.564)
(0.465)
Mean of Control Group
F-Statistic
14.9
11.9
14.1
9.1
10.3
12.6
3.9
2.1
2.0
Observations
4,918
4,917
4,769
2,830
2,830
2,800
2,088
2,087
2,087
Panel B: First Birth before Age 20 {0,1}
Schooling
-0.037**
-0.034*
-0.049***
-0.056
-0.069*
-0.086**
-0.026
-0.012
-0.020
(0.018)
(0.019)
(0.019)
(0.041)
(0.039)
(0.040)
(0.024)
(0.025)
(0.024)
Mean of Control Group
F-Statistic
27.9
30.0
36.7
13.8
17.0
18.5
11.7
12.2
14.1
Observations
7,713
7,709
7,263
3,857
3,856
3,783
3,856
3,853
3,480
Panel C: Married before Age 20 {0,1}
Schooling
-0.063*** -0.061*** -0.081***
-0.078** -0.090** -0.113***
-0.059** -0.045*
-0.064**
(0.018)
(0.018)
(0.020)
(0.039)
(0.037)
(0.040)
(0.025)
(0.025)
(0.025)
Mean of Control Group
F-Statistic
27.9
30.0
35.5
13.8
17.0
18.4
11.7
12.2
13.7
Observations
7,713
7,709
7,263
3,857
3,856
3,783
3,856
3,853
3,480
Panel D: Sex before Age 20 {0,1}
Schooling
-0.044**
-0.042** -0.063***
-0.047
-0.064*
-0.082**
-0.046*
-0.031
-0.052**
(0.018)
(0.019)
(0.019)
(0.037)
(0.035)
(0.037)
(0.028)
(0.029)
(0.026)
Mean of Control Group
F-Statistic
28.5
29.9
36.1
14.8
17.9
19.1
11.2
11.4
13.6
Observations
7,623
7,619
7,180
3,807
3,806
3,733
3,816
3,813
3,447
Controls (Post Double Selection)
NO
YES
(24,47,43,46)
NO
YES
(19,33,35,37)
NO
YES
(47,28,28,23)
Province Fixed Effects
NO
YES
PDS
NO
YES
PDS
NO
YES
PDS
Survey Fixed Effects
YES
YES
PDS
YES
YES
PDS
YES
YES
PDS
Instrumented by 1[YoB ≥ 1992]
[0.48]
[0.54]
[0.41]
[0.34]
[0.37]
[0.31]
Regression Sample
Full
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
[19.67]
[19.59]
[19.80]
Instrumented by 1[YoB ≥ 1992]
[0.55]
[0.59]
[0.50]
Notes: The results in each panel and column are produced by a separate regression. Columns (3), (6) and (9) report
results using the Post Double Selection (PDS) algorithm developed in Belloni et al. (2014), considering a total of 896
potential controls. The included number of PDS’ selected controls are given in parenthesis at the bottom of the table.
The IV-estimate and key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or
after the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported
Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for the pooling
of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents significance
at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
91
Table B.4: Second-Stage Results, Income (Controls following Post-Double-Selection)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Panel A: Worked Last Year {0,1}
Schooling
-0.003
-0.002
0.003
0.017
0.005
0.002
0.000
-0.002
0.004
(0.015)
(0.013)
(0.014)
(0.024)
(0.020)
(0.020)
(0.025) (0.022)
(0.021)
Mean of Control Group
F-Statistic
27.9
30.0
34.5
13.8
17.0
18.8
11.7
12.2
13.5
Observations
7,713
7,709
7,263
3,857
3,856
3,783
3,856
3,853
3,480
Panel B: Works outside of the Household for Cash/Money {0,1}
Schooling
0.023**
0.026**
0.024**
0.058*
0.057**
0.052**
0.007
0.008
0.003
(0.011)
(0.013)
(0.012)
(0.031)
(0.027)
(0.025)
(0.014) (0.015)
(0.014)
Mean of Control Group
F-Statistic
27.9
30.0
35.1
13.8
17.0
18.7
11.7
12.2
13.3
Observations
7,713
7,709
7,263
3,857
3,856
3,783
3,856
3,853
3,480
Controls
NO
YES
(44,39)
NO
YES
(33,29)
NO
YES
(28,29)
Province Fixed Effects
NO
YES
PDS
NO
YES
PDS
NO
YES
PDS
Survey Fixed Effects
YES
YES
PDS
YES
YES
PDS
YES
YES
PDS
Regression Sample
Full
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
[0.86]
[0.92]
[0.78]
Instrumented by 1[YoB ≥ 1992]
[0.08]
[0.08]
[0.07]
Notes: The results in each panel and column are produced by a separate regression. Columns (3), (6) and (9) report
results using the Post Double Selection (PDS) algorithm developed in Belloni et al. (2014), considering a total of 896
potential controls. The included number of PDS’ selected controls are given in parenthesis at the bottom of the table. The
IV-estimate and key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or after
the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported Kleibergen-
Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for the pooling of survey
rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents significance at the 1,
5 and 10 percent level, respectively. See full notes below Table 3.2.
92
Table B.5: Second-Stage Results, Birth before Age {0,1}
Age 16 Age 17 Age 18 Age 19 Age 20 Age 21 Age 22 Age 23
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Panel A: Full
Mean of Control Group
[0.03]
[0.07]
[0.12]
[0.22]
[0.34]
[0.48]
[0.60]
[0.68]
Schooling
0.002
0.012
0.009
-0.031* -0.034*
-0.033
-0.033
-0.018
(0.007) (0.010) (0.014) (0.017) (0.019) (0.029) (0.027) (0.029)
F-Statistic
30.0
30.0
30.0
30.0
30.0
12.6
13.8
11.3
Observations
7,709
7,709
7,709
7,709
7,709
6,412
5,553
4,576
Panel B: Poor
Mean of Control Group
[0.03]
[0.07]
[0.13]
[0.23]
[0.37]
[0.52]
[0.65]
[0.74]
Schooling
-0.003
0.019
0.016
-0.054
-0.069*
0.001
-0.041
-0.025
(0.013) (0.021) (0.029) (0.034) (0.039) (0.066) (0.045) (0.042)
F-Statistic
17.0
17.0
17.0
17.0
17.0
4.9
9.8
8.7
Observations
3,856
3,856
3,856
3,856
3,856
3,217
2,792
2,325
Panel C: Wealthy
Mean of Control Group
[0.03]
[0.07]
[0.11]
[0.19]
[0.31]
[0.44]
[0.52]
[0.60]
Schooling
0.008
0.010
0.006
-0.021
-0.012
-0.062
-0.037
-0.008
(0.009) (0.014) (0.018) (0.023) (0.025) (0.039) (0.043) (0.058)
F-Statistic
12.2
12.2
12.2
12.2
12.2
5.9
4.5
3.1
Observations
3,853
3,853
3,853
3,853
3,853
3,195
2,761
2,251
Controls
YES
YES
YES
YES
YES
YES
YES
YES
Province Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
First Birth before Age ... {0,1}
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
Notes: The results in each panel and column panel are produced by a separate regression. The IV-estimate
at each age increment tests the outcome that women had a first birth before the indicated age while restricting
the sample to respondents aged at least that of the investigated age. The IV-estimate and key explanatory
variable “Schooling” is instrumented by the binary indicator of being born in or after the cutoff, “1[Year of
Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported Kleibergen-Paap rk
Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for the pooling of survey
rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents
significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
93
Table B.6: Second-Stage Results, Married before Age {0,1}
Age 16 Age 17 Age 18 Age 19
Age 20
Age 21 Age 22 Age 23
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Panel A: Full
Mean of Control Group
[0.07]
[0.13]
[0.22]
[0.34]
[0.48]
[0.58]
[0.66]
[0.71]
Schooling
-0.012
-0.007
-0.007
-0.032
-0.061***
-0.044
-0.032
-0.017
(0.011) (0.013) (0.017) (0.019)
(0.018)
(0.027) (0.025) (0.028)
F-Statistic
30.0
30.0
30.0
30.0
30.0
12.6
13.8
11.3
Observations
7,709
7,709
7,709
7,709
7,709
6,412
5,553
4,576
Panel B: Poor
Mean of Control Group
[0.08]
[0.14]
[0.24]
[0.38]
[0.54]
[0.65]
[0.74]
[0.79]
Schooling
-0.021
0.009
-0.008
-0.059
-0.090**
-0.020
-0.024
-0.009
(0.022) (0.028) (0.033) (0.037)
(0.037)
(0.064) (0.039) (0.041)
F-Statistic
17.0
17.0
17.0
17.0
17.0
4.9
9.8
8.7
Observations
3,856
3,856
3,856
3,856
3,856
3,217
2,792
2,325
Panel C: Wealthy
Mean of Control Group
[0.06]
[0.11]
[0.19]
[0.30]
[0.41]
[0.49]
[0.55]
[0.61]
Schooling
-0.006
-0.023
-0.004
-0.011
-0.045*
-0.066*
-0.042
-0.020
(0.013) (0.017) (0.022) (0.027)
(0.025)
(0.038) (0.043) (0.050)
F-Statistic
12.2
12.2
12.2
12.2
12.2
5.9
4.5
3.1
Observations
3,853
3,853
3,853
3,853
3,853
3,195
2,761
2,251
Controls
YES
YES
YES
YES
YES
YES
YES
YES
Province Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Married before Age ... {0,1}
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
Notes: The results in each panel and column panel are produced by a separate regression. The IV-estimate
at each age increment tests the outcome that women had a first birth before the indicated age while restricting
the sample to respondents aged at least that of the investigated age. The IV-estimate and key explanatory
variable “Schooling” is instrumented by the binary indicator of being born in or after the cutoff, “1[Year of
Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported Kleibergen-Paap rk
Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for the pooling of survey
rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents
significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
94
Table B.7: Second-Stage Results, Sex before Age {0,1}
Age 16 Age 17
Age 18
Age 19
Age 20
Age 21 Age 22 Age 23
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Panel A: Full
Mean of Control Group
[0.09]
[0.15]
[0.26]
[0.42]
[0.55]
[0.66]
[0.73]
[0.78]
Schooling
-0.006
-0.002
-0.039**
0.000
-0.042**
-0.031
-0.016
-0.019
(0.011) (0.015)
(0.018)
(0.020)
(0.019)
(0.027) (0.024) (0.025)
F-Statistic
29.9
29.9
29.9
29.9
29.9
12.8
13.7
11.4
Observations
7,619
7,619
7,619
7,619
7,619
6,340
5,489
4,528
Panel B: Poor
Mean of Control Group
[0.09]
[0.16]
[0.27]
[0.43]
[0.59]
[0.71]
[0.78]
[0.83]
Schooling
-0.017
0.002
-0.076**
-0.018
-0.064*
-0.010
0.001
-0.004
(0.022) (0.028)
(0.035)
(0.036)
(0.035)
(0.054) (0.037) (0.035)
F-Statistic
17.9
17.9
17.9
17.9
17.9
5.4
10.1
8.8
Observations
3,806
3,806
3,806
3,806
3,806
3,176
2,756
2,298
Panel C: Wealthy
Mean of Control Group
[0.08]
[0.14]
[0.25]
[0.39]
[0.50]
[0.59]
[0.66]
[0.71]
Schooling
0.002
-0.006
-0.015
0.018
-0.031
-0.050
-0.039
-0.041
(0.016) (0.021)
(0.026)
(0.032)
(0.029)
(0.040) (0.042) (0.049)
F-Statistic
11.4
11.4
11.4
11.4
11.4
5.4
4.1
3.0
Observations
3,813
3,813
3,813
3,813
3,813
3,164
2,733
2,230
Controls
YES
YES
YES
YES
YES
YES
YES
YES
Province Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
First Sex before Age ... {0,1}
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
Notes: The results in each panel and column panel are produced by a separate regression. The IV-estimate
at each age increment tests the outcome that women had a first birth before the indicated age while
restricting the sample to respondents aged at least that of the investigated age. The IV-estimate and key
explanatory variable “Schooling” is instrumented by the binary indicator of being born in or after the cutoff,
“1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported Kleibergen-
Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for the pooling
of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents
significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
95
Table B.8: Second-Stage Results, Knowledge and Use of Condoms
(1)
(2)
(3)
(4)
(5)
(6)
Panel A: Knows Condom as contr. Method {0,1}
Schooling
0.010
0.010
0.011
0.012
0.010
0.011
(0.009)
(0.009)
(0.020)
(0.019)
(0.011) (0.011)
Mean of Control Group
F-Statistic
27.9
30.0
13.8
17.0
11.7
12.2
Observations
7,713
7,709
3,857
3,856
3,856
3,853
Panel B: Used Condom with last sex. Partner {0,1}
Schooling
0.040**
0.042*
0.033*
0.031*
0.070
0.090
(0.020)
(0.025)
(0.019)
(0.018)
(0.071) (0.126)
Mean of Control Group
F-Statistic
10.7
8.5
8.4
9.8
1.6
0.7
Observations
4,801
4,799
2,703
2,702
2,098
2,097
Controls
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
Full
Poor
Wealthy
Regression Sample
Instrumented by 1[YoB ≥ 1992]
[0.94]
[0.92]
[0.96]
Instrumented by 1[YoB ≥ 1992]
[0.02]
[0.01]
[0.04]
Notes: The results in each panel and column are produced by a separate regression. The IV-estimate and
key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or after the
cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported
Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for
the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **,
* represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
96
Table B.9: Second-Stage Results, Engagement with Mass Media
(1)
(2)
(3)
(4)
(5)
(6)
Panel A: Reads Newspaper at least once per Week {0,1}
Schooling
-0.011
-0.013
-0.003
-0.003
-0.025
-0.026
(0.008)
(0.008)
(0.011)
(0.011)
(0.015) (0.016)
Mean of Control Group
F-Statistic
27.9
30.0
13.9
17.1
11.7
12.2
Observations
7,712
7,708
3,856
3,855
3,856
3,853
Panel B: Listens to Radio at least once per Week {0,1}
Schooling
0.029*
0.025
0.015
0.014
0.023
0.019
(0.017)
(0.019)
(0.034)
(0.032)
(0.026) (0.027)
Mean of Control Group
F-Statistic
27.8
29.8
13.8
17.0
11.6
12.1
Observations
7,712
7,708
3,857
3,856
3,855
3,852
Panel C: Watches TV at least once per Week {0,1}
Schooling
0.014
0.010
0.007*
0.007*
0.013
0.012
(0.009)
(0.008)
(0.004)
(0.004)
(0.019) (0.017)
Mean of Control Group
F-Statistic
28.0
30.2
13.9
17.1
11.7
12.2
Observations
7,710
7,706
3,856
3,855
3,854
3,851
Controls
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
Full
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
Regression Sample
Instrumented by 1[YoB ≥ 1992]
[0.06]
[0.00]
[0.14]
[0.06]
[0.03]
[0.10]
Instrumented by 1[YoB ≥ 1992]
[0.42]
[0.30]
[0.58]
Notes: The results in each panel and column are produced by a separate regression. The IV-estimate and
key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or after the
cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported
Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for
the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **,
* represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
97
Table B.10: Second-Stage Results, Husband’s Characteristics
(1)
(2)
(3)
(4)
(5)
(6)
Panel A: Husband's Age
Schooling
-1.183
-1.411
-1.402
-1.177
-2.093
-4.269
(0.854)
(0.994)
(1.148)
(0.990)
(3.407) (10.813)
Mean of Control Group
F-Statistic
6.3
5.2
5.7
7.3
0.6
0.2
Observations
4,430
4,430
2,576
2,576
1,854
1,854
Panel B: Husband's Education
Schooling
0.593
0.591
0.115
0.164
1.575
2.322
(0.392)
(0.441)
(0.451)
(0.399)
(2.225)
(4.940)
Mean of Control Group
F-Statistic
5.9
5.1
5.7
7.2
0.4
0.2
Observations
4,485
4,485
2,606
2,606
1,879
1,879
Panel C: Husband's Desired Number of Children: Same? {0,1}
Schooling
0.065
0.078
0.078
0.071
0.102
0.134
(0.054)
(0.065)
(0.080)
(0.073)
(0.156)
(0.262)
Mean of Control Group
F-Statistic
9.7
7.6
7.5
8.7
1.2
0.5
Observations
3,899
3,899
2,240
2,240
1,659
1,659
Controls
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
[29.78]
[29.38]
[30.39]
Instrumented by 1[YoB ≥ 1992]
[3.59]
[2.49]
[5.30]
Instrumented by 1[YoB ≥ 1992]
[0.63]
[0.62]
[0.63]
Regression Sample
Full
Poor
Wealthy
Instrumented by 1[YoB ≥ 1992]
Notes: The results in each panel and column are produced by a separate regression. The IV-estimate and
key explanatory variable “Schooling” is instrumented by the binary indicator of being born in or after the
cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported
Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and account for
the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **,
* represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
98
Table B.11: Smoothness Test of Covariates, Multiple Indicator Cluster Survey (MICS)
Panel A: Household Characteristics
Attended
School
{0,1}
Wealth
Index
{1,5}
Urban
{0,1}
Household
Size
{1,14}
Mother
Alive
{0,1}
Father
Alive
{0,1}
Religion
(Head)
{1,6}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
-0.032
0.018
-0.016
-0.298
0.003
-0.068
0.038
(0.046)
(0.144)
(0.014)
(0.204)
(0.041)
(0.055)
(0.093)
Mean of Control Group
[0.79]
[3.01]
[0.07]
[5.61]
[0.82]
[0.65]
[1.40]
Observations
2,518
2,523
2,523
2,523
2,205
2,196
2,523
Adj. R-squared
(0.001)
0.002
0.001
0.012
0.023
0.036
0.00
Panel B: Mothers' & Children's Characteristics
Education
Level
{1,3}
Literacy
{0,1}
Ever
Married
{0,1}
Children
Born
{1,13}
Height at
Birth
{1,5}
Was
Breastfed
{0,1}
Received
Vitamin A
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
-0.048
-0.031
0.014
-0.351*
0.064
0.032
0.047
(0.088)
(0.064)
(0.044)
(0.185)
(0.159)
(0.029)
(0.088)
Mean of Control Group
[0.42]
[0.29]
[0.77]
[5.92]
[3.24]
[0.97]
[0.39]
Observations
1,124
1,792
1,913
1,913
857
875
874
Adj. R-squared
0.043
0.027
0.022
0.232
0.002
0.002
(0.00)
Multiple Indicator Cluster Survey (MICS)
1[Year of Birth ≥ 1992]
Last Born Child
1[Children's Year of Birth ≥ 1992]
Notes: The results in each panel and column are produced by a separate regression, using the Burundi Multiple Indicator
Cluster Survey (MICS) of 2005. The discontinuities estimated in panel A reflect changes in the characteristics of households
whose members are born in or after 1992, compared to households of members born in or before 1991. The discontinuities
estimated in panel B reflect changes in characteristics of "potential mothers", i.e. of women who have had their first birth
in or after 1992. Estimates are weighted using the MICS sample weights and account for the pooling of survey rounds.
The standard errors reported are clustered at the survey-cluster level. ***, **, * represents significance at the 1, 5 and 10
percent level, respectively. See full notes below Table 3.2.
99
Table B.12: Education and Women’s Household Wealth
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
Schooling
5755.257 3142.718
0.016
0.007
0.018
0.016
0.007
0.002
0.050*** 0.051**
0.006
0.000
0.020
0.024
(3913.7)
(2463.7)
(0.013)
(0.011)
(0.022)
(0.019)
(0.011)
(0.008)
(0.019)
(0.021)
(0.010)
(0.009)
(0.019) (0.020)
Mean of Control Group
F-Statistic
27.9
30.0
25.8
27.5
25.2
27.2
25.7
27.5
25.7
27.5
25.7
27.5
25.7
27.5
Observations
7,713
7,709
7,612
7,609
7,568
7,565
7,614
7,611
7,614
7,611
7,614
7,611
7,610
7,607
Controls
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
Province Fixed Effects
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
NO
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
[0.46]
[0.06]
[0.22]
Instrumented by 1[YoB ≥ 1992]
-[4444.59]
[0.10]
[0.32]
[0.04]
Dependent Variable
DHS
Wealth Score
Dwelling:
Electricity
{0,1}
Dwelling:
Piped Water
{0,1}
Dwelling:
Flush Toilet
{0,1}
Own:
Radio
{0,1}
Own:
TV
{0,1}
Own:
Bicycle
{0,1}
Notes: The results in each column are produced by a separate regression. The IV-estimate and key explanatory variable “Schooling” is instrumented by the binary indicator of being born
in or after the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using
the DHS sample weights and account for the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents significance at the 1, 5
and 10 percent level, respectively. See full notes below Table 3.2.
100
Table B.13: Second-Stage Results, Redefined Poor and Wealthy Subgroups
Panel A: Literacy, Preferences & Fertility
Poor
Wealthy
Poor
Wealthy
Poor
Wealthy
Poor
Wealthy
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Schooling
0.055**
0.036
-0.215**
-0.036
0.604**
0.232
-0.051*
-0.014
(0.024)
(0.023)
(0.091)
(0.104)
(0.276) (0.305)
(0.028)
(0.029)
Mean of Control Group
[0.56]
[0.86]
[4.01]
[3.72]
[19.67]
[20.49]
[0.34]
[0.22]
F-Statistic
20.9
5.9
20.6
5.5
8.6
2.3
21.0
5.8
Observations
5,801
1,718
5,651
1,682
4,040
775
5,804
1,719
Panel B: Marriage, Sex & Labour Market Outcomes
Poor
Wealthy
Poor
Wealthy
Poor
Wealthy
Poor
Wealthy
(1)
(2)
(1)
(2)
(3)
(4)
(5)
(6)
Schooling
-0.088***
-0.003
-0.064**
0.003
0.004
-0.035
0.045**
-0.005
(0.028)
(0.031)
(0.028)
(0.034)
(0.018) (0.029)
(0.018)
(0.020)
Mean of Control Group
[0.48]
[0.25]
[0.55]
[0.41]
[0.86]
[0.66]
[0.08]
[0.12]
F-Statistic
21.0
5.8
21.0
5.6
21.0
5.8
21.0
5.8
Observations
5,804
1,719
5,737
1,701
5,804
1,719
5,804
1,719
Controls
YES
YES
YES
YES
YES
YES
YES
YES
Province Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Dependent Variable
Literacy
{0,1}
Desired Number
of Childern
Age at
First Birth
First Birth
before Age 20
{0,1}
Instrumented by 1[YoB ≥ 1992]
Instrumented by 1[YoB ≥ 1992]
Worked
last Year
{0,1}
Sex before
Age 20
{0,1}
Married before
Age 20
{0,1}
Works outside
HH. For
Cash/Money
{0,1}
Notes: The results in each column are produced by a separate regression. The sample is comprised of provinces with below
median conflict intensities. The IV-estimate and key explanatory variable “Schooling” is instrumented by the binary
indicator of being born in or after the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is
given by the reported Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the DHS sample weights and
account for the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **,
* represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
101
Table B.14: Second-Stage Results, Dropping Provinces with above-median Conflict
Intensity
Literacy
{0,1}
Desired
Number of
Childern
Age at
first
Birth
First Birth
before Age
20
{0,1}
Married
before
Age 20
{0,1}
Sex
before
Age 20
{0,1}
Worked
last
Year
{0,1}
Works
outside HH.
for Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Schooling
0.078***
-0.166**
0.448**
-0.043**
-0.062***
-0.038*
-0.016
0.019
(0.017)
(0.076)
(0.206)
(0.020)
(0.022)
(0.021)
(0.016)
(0.013)
Mean of Control Group
[0.54]
[4.27]
[19.46]
[0.37]
[0.51]
[0.56]
[0.85]
[0.07]
F-Statistic
22.6
20.8
9.8
22.6
22.6
22.8
22.6
22.6
Observations
3,823
3,692
2,471
3,824
3,824
3,772
3,824
3,824
Controls
YES
YES
YES
YES
YES
YES
YES
YES
Province Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Dependent Variable
Instrumented by 1[YoB ≥ 1992]
Notes: The results in each column are produced by a separate regression. The estimate in row one “1[Year of Birth ≥
1992]” represents the discontinuous increase in schooling by women “just-treated” by the policy (13 years old), compared
to women just too old to benefit from free schooling (14 years old). Estimates are weighted using the DHS sample weights
and account for the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***,
**, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
102
Table B.15: First-stage Discontinuity in Schooling, Varying Functional Form
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
1[Year of Birth ≥ 1992]
1.234***
0.956*** 1.284***
1.223***
0.908*** 1.280***
1.170***
0.772*** 1.348***
1.498***
0.843**
1.787***
1.228***
0.683**
1.526***
0.413
-0.510
0.569
(0.223)
(0.220)
(0.366)
(0.223)
(0.220)
(0.367)
(0.228)
(0.230)
(0.365)
(0.365)
(0.355)
(0.619)
(0.296)
(0.296)
(0.477)
(0.731)
(0.717)
(1.283)
Interaction: 1 × Year of Birth
0.164**
0.339***
-0.099
0.428**
0.193
0.647**
-1.385
-1.832*
-1.698
(0.077)
(0.078)
(0.130)
(0.169)
(0.141)
(0.286)
(0.959)
(1.000)
(1.596)
Interaction: 1 × (Year of Birth)2
0.965***
0.922***
0.706
-1.385
-0.294
-0.182
(0.302)
(0.328)
(0.484)
(0.959)
(0.449)
(0.696)
Interaction: 1 × (Year of Birth)3
-0.116
-0.170***
-0.147*
(0.408)
(0.063)
(0.085)
Year of Birth
0.105***
0.200***
0.028
0.038
0.064
0.070
0.127***
0.257***
0.004
-0.370
-0.081
-0.466
0.103
0.295***
-0.069
1.252
1.919**
1.301
(0.037)
(0.038)
(0.058)
(0.047)
(0.044)
(0.085)
(0.040)
(0.043)
(0.060)
(0.241)
(0.236)
(0.415)
(0.095)
(0.097)
(0.148)
(0.874)
(0.881)
(1.523)
Year of Birth^2
0.011
0.028***
-0.013
-0.066*
-0.024
-0.086
0.013
0.026***
-0.008
0.535*
0.718**
0.568
(0.007)
(0.007)
(0.013)
(0.039)
(0.038)
(0.066)
(0.009)
(0.010)
(0.015)
(0.315)
(0.322)
(0.549)
Year of Birth^3
0.001
-0.002
0.003
0.066*
0.081**
0.072
(0.004)
(0.004)
(0.006)
(0.034)
(0.036)
(0.060)
Intercept
6.497***
3.238*** 7.135***
6.288***
2.804*** 7.258***
6.440***
3.099*** 7.197***
5.753***
2.601*** 6.559***
-3.873*** 3.172*** 7.051***
7.023***
4.172***
7.933***
(0.436)
(0.788)
(0.596)
(0.442)
(0.779)
(0.622)
(0.436)
(0.780)
(0.601)
(0.518)
(0.854)
(0.812)
(0.284)
(0.811)
(0.659)
(0.830)
(0.986)
(1.403)
Observations
7709
3856
3853
7709
3856
3853
7709
3856
3853
7709
3856
3853
7709
3856
3853
7709
3856
3853
Adj. R-squaed
0.262
0.179
0.187
0.263
0.183
0.187
0.263
0.182
0.187
0.264
0.184
0.187
0.263
0.182
0.187
0.264
0.185
0.187
Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Province Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Functional Form
(d)
Quadratic Interaction
(e)
Cubic
(f)
Cubic Interaction
(a)
Linear
(b)
Linear Interaction
(c)
Quadratic
Notes: The results in each column are produced by a separate regression. The estimate in row one “1[Year of Birth ≥ 1992]” represents the discontinuous increase in schooling by women “just-treated”
by the policy (13 years old), compared to women just too old to benefit from free schooling (14 years old). The standard errors reported are clustered at the survey-cluster level. Estimates are
weighted using the DHS sample weights and account for the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents significance at the
1, 5 and 10 percent level, respectively. See full notes below Table 3.2.
103
Table B.16: Second-Stage Results: Varying Functional Form
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
Full
Poor
Wealthy
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
Literacy {0,1}
Schooling
0.051***
0.070**
0.027
0.050***
0.067**
0.027
0.043**
0.057
0.025
0.066***
0.081
0.041
0.052**
0.052
0.039
0.043
0.143
0.043
(0.017)
(0.028)
(0.024)
(0.017)
(0.030)
(0.024)
(0.018)
(0.035)
(0.022)
(0.023)
(0.055)
(0.030)
(0.023)
(0.058)
(0.027)
(0.203)
(0.204)
(0.216)
F-Statistic
[30.4]
[19.0]
[12.0]
[29.9]
[17.1]
[11.9]
[26.2]
[11.3]
[13.4]
[16.7]
[5.6]
[8.0]
[17.0]
[5.3]
[10.0]
[0.3]
[0.5]
[0.1]
Desired Number of Children
Schooling
-0.150**
-0.248**
-0.088
-0.151**
-0.261**
-0.088
-0.153**
-0.306**
-0.076
-0.035
-0.091
-0.027
-0.063
-0.177
-0.025
-1.233
0.759
-0.657
(0.063)
(0.112)
(0.086)
(0.063)
(0.118)
(0.086)
(0.066)
(0.148)
(0.079)
(0.079)
(0.205)
(0.093)
(0.075)
(0.207)
(0.084)
(2.850)
(1.049)
(1.416)
F-Statistic
[28.9]
[17.5]
[11.9]
[28.3]
[15.8]
[11.9]
[24.4]
[10.5]
[12.8]
[17.4]
[5.7]
[9.1]
[17.8]
[5.2]
[11.2]
[0.2]
[0.8]
[0.2]
Age at First Birth
Schooling
0.401**
0.367
0.710
0.468**
0.449*
0.757
0.632**
0.676**
0.837
0.225
0.433
0.115
0.324
0.566
0.238
1.942
8.224
11.278
(0.191)
(0.245)
(0.554)
(0.204)
(0.263)
(0.564)
(0.256)
(0.342)
(0.607)
(0.215)
(0.421)
(0.319)
(0.205)
(0.415)
(0.273)
(3.569)
(139.187) (160.028)
F-Statistic
[12.9]
[11.6]
[2.1]
[11.9]
[10.3]
[2.1]
[9.5]
[7.6]
[2.0]
[10.3]
[5.0]
[3.0]
[11.0]
[5.2]
[4.0]
[0.3]
[0.0]
[0.0]
First Birth before Age 20 {0,1}
Schooling
-0.033*
-0.066*
-0.013
-0.034*
-0.069*
-0.012
-0.038**
-0.083*
-0.016
-0.038
-0.122
0.003
-0.037
-0.125
0.003
-0.251
0.167
-0.176
(0.019)
(0.037)
(0.025)
(0.019)
(0.039)
(0.025)
(0.020)
(0.047)
(0.023)
(0.029)
(0.079)
(0.035)
(0.027)
(0.080)
(0.032)
(0.454)
(0.370)
(0.419)
F-Statistic
[30.5]
[19.0]
[12.3]
[30.0]
[17.0]
[12.2]
[26.3]
[11.3]
[13.6]
[16.9]
[5.6]
[8.3]
[17.2]
[5.3]
[10.2]
[0.3]
[0.5]
[0.2]
Married before Age 20 {0,1}
Schooling
-0.062*** -0.088** -0.044*
-0.061*** -0.090**
-0.045*
-0.059*** -0.099**
-0.037
-0.082*** -0.180**
-0.039
-0.074*** -0.179**
-0.027
-0.226
-0.030
-0.364
(0.018)
(0.035)
(0.025)
(0.018)
(0.037)
(0.025)
(0.019)
(0.044)
(0.023)
(0.030)
(0.090)
(0.034)
(0.027)
(0.087)
(0.031)
(0.399)
(0.239)
(0.763)
F-Statistic
[30.5]
[19.0]
[12.3]
[30.0]
[17.0]
[12.2]
[26.3]
[11.3]
[13.6]
[16.9]
[5.6]
[8.3]
[17.2]
[5.3]
[10.2]
[0.3]
[0.5]
[0.2]
Sex before Age 20 {0,1}
Schooling
-0.042**
-0.063*
-0.031
-0.042**
-0.064*
-0.031
-0.039**
-0.067
-0.025
-0.042
-0.110
-0.008
-0.043
-0.112
-0.011
-0.115
0.083
-0.037
(0.019)
(0.033)
(0.028)
(0.019)
(0.035)
(0.029)
(0.020)
(0.041)
(0.027)
(0.029)
(0.077)
(0.037)
(0.028)
(0.076)
(0.035)
(0.250)
(0.276)
(0.223)
F-Statistic
[30.3]
[19.8]
[11.6]
[29.9]
[17.9]
[11.4]
[26.5]
[12.0]
[13.0]
[17.3]
[5.8]
[8.2]
[17.4]
[5.5]
[9.8]
[0.4]
[0.5]
[0.3]
Worked Last Year {0,1}
Schooling
-0.002
0.003
-0.002
-0.002
0.005
-0.002
0.000
0.015
-0.004
0.003
0.005
0.005
0.009
0.010
0.013
-0.183
0.080
-0.209
(0.013)
(0.019)
(0.022)
(0.013)
(0.020)
(0.022)
(0.015)
(0.026)
(0.021)
(0.021)
(0.039)
(0.034)
(0.021)
(0.040)
(0.031)
(0.320)
(0.181)
(0.444)
F-Statistic
[30.5]
[19.0]
[12.3]
[30.0]
[17.0]
[12.2]
[26.3]
[11.3]
[13.6]
[16.9]
[5.6]
[8.3]
[17.2]
[5.3]
[10.2]
[0.3]
[0.5]
[0.2]
Instr. by 1[YoB ≥ 1992]
Functional Form
Instr. by 1[YoB ≥ 1992]
Instr. by 1[YoB ≥ 1992]
Instr. by 1[YoB ≥ 1992]
Instr. by 1[YoB ≥ 1992]
Instr. by 1[YoB ≥ 1992]
(d)
Quadratic Interaction
(e)
Cubic
(f)
Cubic Interaction
(a)
Linear
(b)
Linear Interaction
(c)
Quadratic
Instr. by 1[YoB ≥ 1992]
Notes: Table continued on next page.
104
Works outside of the Household for Cash/Money {0,1}
Schooling
0.026**
0.055**
0.008
0.026**
0.057**
0.008
0.024*
0.061*
0.006
0.058***
0.124*
0.038
0.063***
0.140*
0.036
0.054
-0.021
0.087
(0.012)
(0.026)
(0.015)
(0.013)
(0.027)
(0.015)
(0.013)
(0.033)
(0.015)
(0.023)
(0.073)
(0.024)
(0.023)
(0.078)
(0.024)
(0.166)
(0.145)
(0.245)
F-Statistic
[30.5]
[19.0]
[12.3]
[30.0]
[17.0]
[12.2]
[26.3]
[11.3]
[13.6]
[16.9]
[5.6]
[8.3]
[17.2]
[5.3]
[10.2]
[0.3]
[0.5]
[0.2]
Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Province Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Survey Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Instr. by 1[YoB ≥ 1992]
Notes: The results in each panel and in each column are produced by a separate regression. The IV-estimate and key explanatory variable “Schooling” is instrumented by the binary indicator of being
born in or after the cutoff, “1[Year of Birth (YoB) ≥ 1992]”. The strength of the excluded instrument is given by the reported Kleibergen-Paap rk Wald F-Statistic. Estimates are weighted using the
DHS sample weights and account for the pooling of survey rounds. The standard errors reported are clustered at the survey-cluster level. ***, **, * represents significance at the 1, 5 and 10 percent
level, respectively. See full notes below Table 3.2.
105
Figure B.1 Regression Discontinuity Estimate, Full Sample
Figure B.2: Regression Discontinuity Estimate, Poor Subgroup
Figure B.3: Regression Discontinuity Estimate, Wealthy Subgroup
Notes: The figure visually depicts the Regression Discontinuity estimate 𝛼𝛼2 (slope before 1992) and the
combined effect 𝛼𝛼2 + 𝛼𝛼3 (slope after 1992) produced from regression (3.1). See full notes in Table 3.2.
Notes: Identically produced as Figure B.1 using the subsample of “poor” women.
Notes: Identically produced as Figure B.1 using the subsample of “wealthy” women.
106
Figure B.4 Difference-in-Difference-in-Differences
Notes: The “Point Estimates” represent difference-in-differences coefficients, including a 95% confidence interval,
for the respective dependent variable indicated on the y-axis. Data come from the 2000 and 2005 Multiple Cluster
Indicator Surveys (MICS) as well as the 2010/2011 Burundi DHS survey, using the women’s recodes. The
coefficients are estimated through “triple-difference” estimation (3.3), comparing individuals aged 15-19 to
individuals aged 20-24 in 2000, 2005 and 2010 for both “wealthy” and for “poor” households (splitting at the
median household wealth level). The estimates are weighted using MICS/DHS sample weights and account for
the pooling of sample rounds. Standard errors are clustered at the birth-year and enumeration-area level. See
subsection 3.5 for a detailed discussion.
107
Figure B.5: Placebo Discontinuity, Literacy {0,1}
Figure B.6: Placebo Discontinuity, Desired Number of Children
Figure B.7: Placebo Discontinuity, Age at First Birth
Notes: The figure visually depicts separate reduced form Regression Discontinuity estimate 𝛼𝛼1 produced
from regression (3.1). for varying (placebo) intervention years. See Section 3.5 for a detailed discussion.
Notes: Identically produced as Figure B.5 but replacing the outcome variable.
Notes: Identically produced as Figure B.5 but replacing the outcome variable.
108
Figure B.8: Placebo Discontinuity, First Birth before Age 20 {0,1}
Figure B.9: Placebo Discontinuity, Married before Age 20 {0,1}
Figure B.10: Placebo Discontinuity, Sex before Age 20 {0,1}
Notes: Identically produced as Figure B.5 but replacing the outcome variable.
Notes: Identically produced as Figure B.5 but replacing the outcome variable.
Notes: Identically produced as Figure B.5 but replacing the outcome variable.
109
Figure B.11: Placebo Discontinuity, Worked the Last Year {0,1}
Figure B.12: Placebo Discontinuity, Works outside of the Household for Cash/Money {0,1}
Figure B.13: First-Stage Discontinuity by Bandwidth
Notes: Identically produced as Figure B.5 but replacing the outcome variable.
Notes: Identically produced as Figure B.5 but replacing the outcome variable.
Notes: The figure plots separate first-stage regression estimates produced via equation
(3.1) for various sample bandwidths. See Section 3.5 for a more detailed discussion.
110
Figure B.14: First-Stage Discontinuity by Bandwidth, Separated by Wealth
Figure B.15: Second-Stage Estimate by Bandwidth, Literacy {0,1}
Figure B.16: Second-Stage Estimate by Bandwidth, Ideal Number of Children
Notes: Identically produced as Figure B.13 but using the “poor” and “wealthy” subsamples, respectively.
Notes: The figure plots separate second-stage regression estimates produced via equation (3.2) for various
sample bandwidths for both “Poor” and “Wealthy” subsamples. See Section 3.5 for a more detailed discussion.
Notes: Identically produced as Figure B.15 but replacing the outcome variable.
111
Figure B.17: Second-Stage Estimate by Bandwidth, Age at first Birth
Figure B.18: Second-Stage Estimate by Bandwidth, Birth before Age 20 {0,1}
Figure B.19: Second-Stage Estimate by Bandwidth, Married before Age 20 {0,1}
Notes: Identically produced as Figure B.15 but replacing the outcome variable.
Notes: Identically produced as Figure B.15 but replacing the outcome variable.
Notes: Identically produced as Figure B.15 but replacing the outcome variable.
112
Figure B.20: Second-Stage Estimate by Bandwidth, Sex before Age 20 {0,1}
Figure B.21: Second-Stage Estimate by Bandwidth, Worked Last Year {0,1}
Figure B.22: Second-Stage Estimate by Bandwidth, Works outside of HH. for Cash {0,1}
Notes: Identically produced as Figure B.15 but replacing the outcome variable.
Notes: Identically produced as Figure B.15 but replacing the outcome variable.
Notes: Identically produced as Figure B.15 but replacing the outcome variable.
113
114
Chapter 4
Regional Market Integration and Household
Welfare: Spatial Evidence from the East
African Community1
Abstract. The distributional consequences of trade liberalization in Africa are under-researched.
In this paper, I investigate the differential impact of the East African Community (EAC) on
household welfare using three distinct sets of longitudinal, geo-referenced household-level surveys
from the three founding members Kenya, Tanzania and Uganda. I thereby treat the re-establishment
of the EAC in 2001 – and the expansion to a customs union and common market in 2005 and 2010,
respectively – as a regional policy intervention having differential effects on individual households
governed by their geo-spatial location within the countries, a prediction I derive formally from a
canonical New Economic Geography (NEG) model. The model accommodates a distinct spatial
feature of the EAC, namely that all its members host economically dominating agglomerations that
are tucked away geographically from the regions bordering the respective community members. To
empirically test the hypotheses drawn from the model, I employ a reduced form, difference-in-
differences specification with treatment intensity given by households’ road distance to internal
EAC border crossings, effectively comparing household welfare between “remote” and “border”
households (first difference) before and after the intervention (second difference), while allowing for
a corresponding differential trend of outcomes over time in the pre-existing economic hubs. Results
reveal that households located closer to the internal EAC border did not experience relative welfare
effects compared to other regions following the re-establishment. Rather, the findings hint at
increased agglomeration in the “core”, as measured by an increase in consumption and population
density in the pre-existing economic hubs. I observe mixed evidence on extensive as well as intensive
labor market outcomes.
JEL Classification: F14, F15, R12, O15, O55
Keywords: FTA, East African Community, New Economic Geography, Difference-in-Differences
1 Acknowledgements. Funding from the DFG (project number 390713894) is gratefully acknowledged. I thank
the participants at the American Economic Association Annual Meeting 2024 (San Antonio, USA), at the
Centre for the Study of African Economies Conference 2024 (Oxford, UK), at the EPCS Annual Meeting 2024
(Vienna, Austria), at the African Studies the IPORA-EXC52 Workshop 2023 (Bordeaux, France), at the 2023
German Economic Association (VFS) Annual Meeting (Regensburg, Germany), at the German Development
Economics Conference 2023 (Dresden, Germany), at the 2023 BEST conference (Brisbane, Australia), at the
African Studies Association of Africa Bi-Annual Meeting 2022 (Cape Town, South Africa), at the LaTrobe-
Bayreuth Roundtable 2021 (Melbourne, Australia), at the EXC52 Cluster of Excellence Annual Conference
2023 and 2021 (Bayreuth, Germany), and at the Biannual Graduate Seminar of the faculty of economics at
the University of Bayreuth 2022 and 2021 for their insightful comments and remarks at various stages of the
project. I also thank Lara Bieske and Dimitrios Theodoridis for their extensive assistance in the gathering of
data. Last, I thank Axel Dreher for his specific comments on an early draft. All remaining errors are my own.
115
4.1 Introduction
“[…] the Community shall ensure: […] the
strengthening and consolidation of co-operation in agreed fields that
would lead to equitable economic development within the Partner
States and which would in turn, raise the standard of living and improve
the quality of life of their populations.”
Treaty of the East African Community (EAC 1999; 12)
Regional economic integration is widely regarded as welfare-enhancing and has been a
specifically popular policy intervention in developing economies (Schiff and Winters 2003).
Particularly in Africa, deepened cooperation and trade has been suggested as ways to alleviate
several barriers to development such as landlockedness, fragmented national markets as well as
poor transport- and communications infrastructure (United Nations Development Programme
2011; World Bank 2020). Research on the aggregate effects of trade and integration generally
supports such sentiments and point to largely positive (long-run) effects of trade liberalization
(see e.g. Frankel and Romer 1999; Feyrer 2019). However, donor agencies have been
emphasizing the potentially inequality-enhancing impact of trade within countries (e.g. World
Bank 2009), and there now exists a well-established literature which studies these distributional
concerns and provides evidence for them (for an overview see Pavcnik 2017). Heterogeneities
may form along factors such as the composition of labor markets (e.g. import-competing vs.
export-oriented), the income and consumption patterns of households, worker and capital
mobility, and the nature of the distortions affected, among others (Winters et al. 2004; Goldberg
and Pavcnik 2007; Winters and Martuscelli 2014). One aspect which has received particular
attention is the spatial consequence of trade liberalization, i.e. the question what happens to
countries’ internal economic geography in response to external trade liberalization (for an
overview see Brülhart 2011; Redding 2022).
Regarding developing economies, the evidence on such distributive effects mainly stem
from liberalization experiences in Asia or the Americas, with Mexico and India forming
prominent country-cases (for an overview see Engel et al. 2021; Barros and Martínez-Zarzoso
2022). In Africa, similar assessments have not been explored until recently and are split along
analyzing household-level evidence without specific regard for (intra-national) space (see Erten
et al. 2019; McCaig and McMillan 2020; Giovannetti et al. 2022)2, or for a spatially motivated
analysis, rely on the use of economic proxies such as light emitted by night (e.g. Cadot et al.
2015; Brülhart et al. 2017; Eberhard-Ruiz and Moradi 2019).
In this paper, using a distinct set of geo-referenced household-level surveys, I provide
novel evidence on the distributional effects of regional trade liberalization in Africa by
2 Here, exposure is typically defined at an administrative boundary and differentiated by the relative
composition of specific industries within these regions.
116
combining the spatial considerations of market integration with a household-level analysis. I
thereby treat the re-establishment of the East African Community (EAC) by Kenya, Tanzania
and Uganda in 2001 – and the expansion to a customs union and common market in 2005 and
2010, respectively – as a regional policy intervention having differential effects on individual
households governed by their geo-spatial location within the countries. I derive this prediction
from a canonical New Economic Geography (NEG) model which fits the distinct spatial layout
of the EAC. As such, and lending myself to previous variants of NEG, the model is put together
to accommodate the stark spatial heterogeneity present in all EAC countries, given by the
presence of economically dominating primate cities (Nairobi, Dar es Salaam and Kampala),
and further incorporates the fact that these pre-existing economic agglomerations are tucked
away geographically as seen by their relative positioning to their EAC counterparts. A robust
result from the simulations of the model is that as trade costs among member countries
decrease, the relative welfare in regions closer to the new markets, i.e. border regions, increase,
which corroborates previous theoretical as well as empirical results. Given that all three
countries have long hosted the preeminent economic agglomerations away from the borders,
regional market integration in the EAC is thereby predicted to act as a dispersion force and to
decrease previous spatial inequalities. These predictions are brought to an empirical test using
a distinct set of geo-referenced household surveys before and after trade liberalization.
Specifically, I employ a reduced-form difference-in-differences specification with treatment
intensity given by households’ road distance to internal EAC border crossings, effectively
comparing household welfare between “remote” and “border” households (first difference)
before and after the intervention (second difference), and also test for a similar differential
trend across these time periods for households living in “core” agglomerations.3
The empirical results show that households living closer to the internal EAC border did
not experience the expected relative welfare gains following the establishment of the EAC as
measured by an array of consumption measures as well as intensive and extensive labor market
outcomes. Rather, I observe that regional market accession within the EAC had a statistically
significant and economically relevant effect on households living in the preeminent interior
economic hubs. For instance, in comparison to pre-EAC periods, households surveyed in the
core increased their consumption of consumer durables by 12% relative to those of all other
regions. Further, they show a relevant decrease in the occurrence of lived poverty given by
deprivations in basic consumption such as food, water and medicine by 32%. I observe mixed
evidence on labor market outcomes from the cross-sectional data employed. However, the panel
dataset shows that individuals in core agglomerations had an 11% increase in the likelihood of
working in more skill-intensive occupations and also, show evidence of having salaried work.
Corresponding to these relative shifts in welfare, I also document a strong increase in population
density by up to one standard deviation in the years after these observed welfare effects, which
3 From now on, the predominant economic hubs are interchangeably referred to as “core” or “interior”.
117
provides evidence on increased agglomeration following the establishment of the EAC. The
results are robust to an array of empirical modifications, extensions as well as validity tests.
As such, these findings go against the general prediction of the present theoretical
simulations and also against the hypothesis prominently outlined in Krugman and Elizondo
(1996), which predicted a dispersion of the concentrated economic activity of developing
countries following liberalization. My results are also in contrast to other recent empirical
findings, which have regularly documented regions closer to the new market (potential) to profit
from the less costly access to them (for an overview see Brülhart 2011). Most notably, the
present results are at odds with Eberhard-Ruiz and Moradi (2019) who provide (remote
sensing) evidence that the EAC’s re-establishment led to a positive (one-off) effect on city
growth for cities closer to the internal EAC borders. While in disagreement with the general
prediction of the model and these previous contributions, the model allows a potential insight
into the reasons why such results may come about. In short, the presence of economically
dominating interiors in all partner countries potentially weakens the dispersion force of market
integration to the point where a regional pattern outlined by stark inequalities as existent prior
to the EAC remains a possible outcome even after regional trade liberalization. Such
explanations remain contentious, though, as the reduced-form empirics are ultimately not apt
to inform on specific parameter coefficients which would permit isolating a key factor driving
these observed effects.
The remainder of the paper is structured as follows. Section 4.2 introduces the relevant
empirical and theoretical literature. Section 4.3 provides the institutional background of
regional market integration in the EAC. Section 4.4 builds and simulates a New Economic
Geography model which incorporates the stylized facts of the EAC. The spatial corollaries of
regional integration drawn from the model motivate the empirical strategy introduced in
Section 4.5. Section 4.6 introduces the data. Section 4.7 provides the main empirical results,
robustness checks and validity tests, as well as several extensions to the results. Section 4.8
offers concluding remarks.
4.2 Related Literature
The paper relates to the body of research investigating the impact of trade on households and
welfare on the one hand, and the literature analyzing the spatial consequences of trade
liberalization on the other.
4.2.1. Trade and Household Welfare
Increased availability of detailed survey data has aided the growth of the literature assessing
the link between trade liberalization and household welfare (for an overview see Goldberg and
Pavcnik 2007; Winters and Martuscelli 2014; Pavcnik 2017; Barros and Martínez-Zarzoso 2022).
This body of research confirms the notion that trade does not unequivocally increase the welfare
118
of all households within a country, i.e. produces winners and losers. An analytical starting point
in thinking about these heterogeneous effects is given by a stylized production-consumption
schedule of households and may be encapsulated by ∆𝑊𝑊= (𝑞𝑞𝑖𝑖−𝑐𝑐𝑖𝑖)∆𝑝𝑝𝑖𝑖, whereby welfare
changes are explicitly moderated by (trade-induced) price changes (see Deaton 1997; Winters
et al. 2004). Depending on whether the household is a net consumer (𝑐𝑐𝑖𝑖) or producer (𝑞𝑞𝑖𝑖) of
product 𝑖𝑖, a given price change ∆𝑝𝑝𝑖𝑖 will either lead to net benefits or net losses. In his seminal
paper, Porto (2006) extends such partial equilibrium statics to a general equilibrium model of
trade, taking into account the simultaneous changes in prices of non-traded goods, and
subsequently, second-round effects resultant of altered factor- rewards and intensities in specific
industries. These dynamics are relevant in cases where specific sectors are facing increased
import-competition from international exporters or where export-oriented produces are
drawing increased demand from abroad. As such, one and the same trade policy may render
very different results across the population depending on the goods affected, households’
production and consumption schedule, and subsequent general equilibrium effects.
Porto’s approach has been subsequently extended and employed to study trade effects in
various contexts, including Mexico (Nicita 2009), Brazil (Borraz et al. 2013), India (Marchand
2012; 2019), Tunisia (Martínez-Zarzoso et al. 2016), as well as in six African countries (Nicita
et al. 2014). These studies typically employ changes in (non-)traded goods prices together with
their respective income- and consumption share reported in household surveys. To assess the
overall welfare impact, these changes are then compared across the (income or expenditure)
distribution to assess the pro-poor or pro-rich character of a trade-policy. Most of these studies
provide evidence of a pro-poor effect of trade, some of them showing mixed results, and Nicita
(2009) being the only exception in showing a clear “rich-only” impact of trade in Mexico.
A second branch of the literature on trade and household welfare has relied on “Bartik-
style” shift-share instruments to identify trade effects.4 Here, exposure is typically defined at
an aggregate level, such as at a particular administrative unit (e.g. districts or regions). The
intensity of trade on households living within a specific region is then differentiated by the pre-
liberalization concentration of industries and the respective tariff cuts (see Goldberg and
Pavcnik 2007; Winters and Martuscelli 2014). For instance, McCaig (2011) shows that the U.S.-
Vietnam Bilateral Trade Agreement accelerated poverty decline, as export growth due to tariff
removal was largest in the low-skilled labor-intensive apparel and clothing sectors. On the other
hand, Topalova (2010) provides evidence that India’s trade liberalization of 1991 actually
slowed poverty decline in the most affected regions, i.e. the ones intensive in agriculture, given
that such sectors faced increased import-competition. Related studies have looked at similar
issues in Brazil (Castilho et al. 2012), China (Emran and Hou 2013), India (Edmonds et al.
2010), Indonesia (Kis-Katos and Sparrow 2015), and Vietnam (Fukase 2013; Vo and Nguyen
2020). These studies are mixed in finding both decreases as well as increases in relative poverty
4 As introduced by Bartik (1991) as well as Blanchard and Katz (1992).
119
as measured by indicators such as consumption, employment, or wages. In Africa, the evidence
on liberalization experiences in this literature is almost universally negative. For instance,
drawing on South Africa’s trade liberalization of the 1990s, Erten et al. (2019) find decreased
formal as well as informal employment for more affected regions and no effects on wages for
those remaining employed. Relatedly, McCaig and McMillan (2020) find neither a contraction
nor an expansion of industries in neighboring Botswana, which was affected by the same
liberalization schedule.5 Rather, they report higher likelihoods of being employed informally
for more intensely affected regions. In the same vein, evidence from Ethiopia suggests increased
unemployment levels in regions more exposed to trade liberalization and import competition
in light of the Structural Adjustment Programs (SAP) of the early 1990s (Giovannetti et al.
2021). One exemption to these findings is Giovannetti et al. (2022) who provide evidence of a
negative effect of protective policies in Egypt shortly after the Spring Revolution. Interestingly,
they find neither positive nor negative results of trade liberalization in the preceding decades.
To my knowledge, there exists no study analyzing household-level welfare concerns of
trade liberalization in Africa from a spatial point of view.6 One exception to this is Cali (2014),
who assesses Uganda’s progressive liberalization policy with Kenya in the 1990s on wage
premia, i.e. changing returns to schooling.7 However, the variation across space is given at a
district level (comparable to GADM8 classification of 2), of which there are a total of 38 and
45 in the study across the two survey rounds, respectively. For comparison, in this paper,
households’ location is defined by latitude-longitude combinations comparable to GADM3 or
finer. As such, I draw from a minimum of 104 and a mean of 324 GPS locations per country
per round, or an average of 299, 326, 353 for Uganda, Tanzania and Kenya, respectively. As
such, analyzing the spatial response of household welfare to trade liberalization (with higher
precision) represents a relevant research gap I aim to fill. Motivating differential trade effects
across space requires an overview of the relevant theoretical and empirical findings in this
regard, which is provided in the next section.
4.2.2. Spatial Effects of Trade
The second strand of literature to which this paper relates investigates the spatial consequences
of trade liberalization. This growing body of research has its roots in New Economic Geography
(NEG) and has extended to an active field now better referred to as “quantitative spatial
economics” (for an overview see Redding and Rossi-Hansberg 2017; Brakman et al. 2019;
Redding 2022).
5 Botswana is a member of the South Africa Customs Union (SACU).
6 I was not able to identify a study exploiting the geo-referencing of survey locales to study these links on any
continent for that matter.
7 The analysis is motivated by a Hecksher-Ohlin type trade effects, thereby suggesting decreasing wage
inequality in a developing country who is labor abundant and human capital scarce.
8 As given by the Database of a Global Administrative Areas https://gadm.org/.
120
Krugman’s (1991) seminal paper was a crucial expansion on earlier conceptualizations of
spatial economic distribution, which mainly concentrated on allocations within cities, such as
the von-Thünen model (1826), or the relative size of cities (Henderson 1974, 1982). The
advantage of NEG in comparison to these earlier specifications lies in the fact that it can
explain the spatial distribution of cities against each other on a featureless plane such that
there are not simply “floating islands” (Brakman et al. 2019; 3). Krugman’s model is essentially
based on new-trade-theory (Krugman 1979, 1980) and combines monopolistic competition
(Dixit and Stiglitz 1977) with increasing returns to scale. Most importantly, trade costs factor
in between locations, regulating their spatial allocation against each other (Krugman 1991).
Hence, the endogenous allocation of activity ultimately boils down to producer- and consumer
maximization problems, who optimize over a given set of preferences and production
technology, while factoring in trade costs. Agglomeration is then a product of cost (forward)
and demand (backward) linkages which produce centripetal forces, while dispersion is a product
of increased competition, the costs of urban congestion, or immobile factors of production. For
instance, because firms operate under increasing returns to scale and incur transport costs,
they benefit from the increased demand in larger locations, i.e. move where demand is highest
(demand linkage).9 And given that consumers have a “love of variety” and will additionally
save on higher price tag for shipping, consumers prefer to locate close to (a large number) of
producers (cost linkage). However, while large regions offer firms high demand and consumers
lower prices, competition as well as costs of congestion (commuting, land rents) are increased
which decreases agglomeration tendencies. In the long-run, an equilibrium is given by the
balance of these forces, i.e. when the advantages and disadvantages of agglomeration or
dispersion, expressed in real wages, are net zero. In this scenario, there exists no incentive for
firms or workers to relocate.
This endogenization of the spatial allocation of economic activity has provided a
workhorse model and spurred subsequent extensions and applications to questions on how
spatial inequalities form and how they may be affected. Importantly, NEG allows the
comparative statics examination of what happens to the centrifugal and centripetal tensions in
response to changes in internal transportation costs or, importantly, external trade costs (for a
synthesis see Fujita et al. 2001).10 Concerning the latter, both theoretical and empirical results
vary in their prediction of whether liberalization increases or decreases spatial disparities within
countries (for an overview see Brülhart 2011). Krugman and Elizondo’s initial treatment (1996)
famously predicted the dissolving of the “giant Third World metropolis” of developing countries
9 Note that in large markets, the additional presence of a firm increases demand mechanically, and by being
able to pay higher wages, thereby further strengthening the backward linkage.
10 While the core model of NEG is known for its “bang-bang” property for changes in transport costs, i.e.
equilibria between complete spreading or agglomeration, subsequent adaptations have accommodated a wider
range of equilibria, using stronger centrifugal (dispersion) forces such as interregional labor immobility (e.g.
Krugman and Venables 1995), diminishing returns in the non-traded sector (e.g. Puga 1999), or housing (e.g.
Helpman 1998).
121
in response to external trade liberalization. The model extends the stylized two-region case to
a three-region-economy, with two regions situated in the home country, and one region (“the
rest of the world”) posing as the international market to which trade costs are successively
lowered (a 2+1 economy). Krugman and Elizondo (1996) sparked an array of refinements and
extensions to this basic setup. Interestingly, however, the prediction from these theoretical
advancements is far from uniform. While Behrens et al. (2003, 2007) confirm the original
prediction, several adaptations arrive at the contrary result, i.e. that increased trade
liberalization sparks intra-national agglomeration. For instance, in the same original 2+1 setup,
Paluzie (2001) as well as Brülhart et al. (2004) and Crozet and Koenig (2004) provide evidence
of increased agglomeration in response to external trade liberalization. Further studies have
extended the setup to 2+2 economies, confirming these predictions (Monfort and Nicolini 2000;
Monfort and van Ypersele 2003). The difference among all of these studies is how they chose
key elements from the “menu of building blocks” (Redding and Rossi-Hansberg 2017; 25), i.e.
how consumer preferences (CES or quasilinear) as well as dispersion forces (immobile workers
vs. congestion) are modeled.11 One particularly interesting adaptation of this literature is to
allow for heterogeneous intra-national space, i.e. regions (within-countries) to differ from one
another ex-ante in terms of their access to foreign markets. For instance, in Mansori (2003),
Brülhart et al. (2004), Crozet and Koenig (2004) and Behrens et al. (2006) they additionally
test what happens to the prediction if one region has better access to the international market
than the other, i.e. poses as a “border” or “gate” region. What this class of models show is that
in almost all instances, external trade liberalization leads to increased “draw” to the border,
i.e. to the region with the better foreign market access (Crozet and Koenig 2004b). However,
depending on the relative size and the export intensity of the home and foreign markets, this
draw to the border may be alleviated as the interior acts as a shield to foreign competition
(Brülhart et al. 2004). These effects may be further mediated by varying intra-national
transport costs which regulate the pass-through of changes in international trade costs towards
the interior as well as the symmetry of the foreign country (Behrens et al. 2006). These initial
refinements to asymmetric regions were first steps into what is now more richly embodied in
“quantitative spatial economics” whereby first-nature characteristics (e.g. local endowments
such as productivity, amenities or floor space) are paired with the “classical” second-nature
agglomeration and dispersion forces, which are produced by the endogenous relative position
of agents against each other (see for a distinction Redding 2022).
The empirical evidence reflects the ambiguity shown across these models. While evidence
from cross-country settings lean towards the convergence of economic activity in response to
trade liberalization, within-country evidence has shown increasing inequalities for various
settings (see for an overview Brülhart 2011). However, a rather robust result across the
11 The menu additionally consists of choices on building blocks such as production technology, trade costs,
externalities, labor mobility, as well as an endowment structure across regions, among others.
122
empirical literature is that regions with relatively better access to foreign markets, often border
regions or regions near the coast, generally stand to benefit comparatively more. This mirrors
the theoretical results of the class of models with heterogeneous intra-national space. Naturally,
whether this leads to convergence or divergence of economic activity within countries naturally
depends on the pre-liberalization diffusion of economic activity. For instance, convergence is
found to occur in settings where market access is higher in the historically economically weaker
border regions, as was the case in Austria (Brülhart et al. 2012) or Germany (Redding and
Sturm 2008).12 On the other hand, divergence is found somewhat more frequently, as
documented by the increasing activity to the already industrialized U.S.-Mexican border
following NAFTA (Hanson 1994, 1997), or in China, where trade has benefitted the already
more developed coastal areas (Kanbur and Zhang 2005). Next to singular country cases, a
growing field of literature employs large-scale evidence from satellite imagery, where lights
emitted by night serve as a proxy for economic activity to assess factors contributing to spatial
(within-country) inequality.13 So far, much of the evidence has a tendency for trade to increase
within-country inequality, and particularly so in developing regions (Ezcurra and Rodríguez-
Pose 2014; Hirte et al. 2020; Ezcurra and Del Villar 2021).14
Within-country evidence for Africa is scarce and is mostly conflated with the mentioned
large-scale studies of all world regions. Particular country-case investigations in Africa so far
have also exclusively relied on nighttime lights as a source of data across space. For instance,
Cadot et al. (2015) look at the influence of improved trade on the border shadow in sub-
Saharan Africa. Similarly, Brülhart et al. (2017) estimate this border shadow for Uganda and
Rwanda in specific. Analyzing trade facilitation measures such as “one-stop-border-posts”, both
papers show a tendency for trade to decrease previous spatial inequalities, i.e. benefit the
regions now better connected. Lastly, and most relevant to the present study, Eberhard-Ruiz
and Moradi (2019) analyze the impact of the East African Community on city growth within
Kenya, Tanzania and Uganda. Using remote-sensing data from before as well as after the re-
establishment of the EAC, the authors provide evidence for a short-whiled, relative increase in
observed nightlight growth for cities closer to internal borders.
As noted in the previous section, there exists no study for the African continent which
exploits the geo-referencing of household surveys to analyze (spatial) effects of trade. However,
evidently, there are distinct benefits in using household-level data to measure such distributive
effects. First, it allows to directly analyze consumption of households, which has been argued
to provide a better measure of overall welfare if one assumes consumption smoothing (Deaton
1997). In a similar vein, trade policies tend to alter prices in a non-uniform way which
12 Redding and Sturm (2008) show evidence for population movements west, increasing regional inequality in
Germany. The reason for my conclusion is that the population movement was induced by market loss of border
regions, rather than market gain, which would vice-versa lead to the opposite result.
13 See e.g. Gibson et al. (2020) on the various uses of night light data in economics.
14 One exception is Brülhart et al. (2019) who show that trade expansion reduces “border shadows”.
123
(differentially) affects both income and consumption, the effect of which is better captured by
consumption as an outcome of both (Goldberg and Pavcnik 2007). Third, the use of household
surveys may also allow us to analyze migration patterns and to test for a potential selection
into specific regions (see Subsection 4.7.2). And while household surveys also have some distinct
weaknesses such as lack of depth in variables, consistency, and comparability across time, or
(non-)response rates, they may also provide a countercheck on the results drawn from using
remote sensing data, e.g. lights emitted by night. Recent research on the quality of nightlight
data has cautioned of the quick application, particularly for developing contexts. Results have
suggested that precisely the areas relevant to development economists, i.e. low density, rural
(agricultural) areas, are due to non-negligible measurement issues (e.g. Bickenbach et al. 2016;
Gibson et al. 2020, 2021). Most relevant to our case, studies have shown that nightlight-to-
GDP elasticities may differ largely between rural and urban areas, which may lead to conflating
a systematic measurement error with policy impacts (Bluhm and McCord 2022).
4.3 Institutional Background
The East African Community (EAC) was originally founded by the Republics of Kenya,
Tanzania, and Uganda in 1967. Placed around Lake Victoria in East Africa, the three countries
share two common borders each and economic and political cooperation between the countries
has historical roots. In pre-independence periods, roughly from 1900-1960, they shared large
infrastructure outlays such as railways, telecommunication, postal service and a common
currency (Hazlewood 1979; EAC 1999). However, not soon after the first formal treaty towards
the establishment of an East African Community was signed in 1967, questions on sovereignty,
and particularly the “disproportionate sharing of benefits of the Community among the Partner
States” arose (EAC 1999; 1). While attempts at redistribution of benefits were made, it was
deemed insufficient by the member states and trade restrictions were levied between them even
while formally in union (Mugomba 1978; Hazlewood 1979). Next to a “lack of strong political
will” (EAC 1999; 1), these are often cited reasons for the ultimate demise of the original EAC
in 1977 when it was formally dissolved. However, the mutual interest in working together in a
union was kept alive in the decades thereafter, as seen by the gradual move towards the modern
EAC for instance by the establishment of the “Permanent Tripartite Commission for East
African Cooperation” in 1993 or the “East African Cooperation Development Strategy” in
1997, which focused on closer co-ordination in economic, political, fiscal, immigration,
infrastructural as well as social and cultural arenas (EAC 1999).
The institutional establishment of the modern day East African Community was initiated
with the treaty of 1999, which was ratified on July 7th of 2000, and the new EAC began to
operate as a free trade area on January 15th of 2001 (EAC 1999; Kaahwa 2003). Hence, it was
not before 2001 after which the substantial lowering of tariff rates by member states was
124
initiated.15 The EAC consistently moved towards deeper integration in the years thereafter,
with the protocol for a customs union operational from the 1st of January 2005 followed by a
transitional period to a common market on the 1st of July in 2010. While member states have
since ratified the move towards a monetary union in 2013, a common currency has not been
implemented as of yet. Figure 4.1 below depicts these developments quantitatively, by plotting
the simple (unweighted) average tariffs among the EAC founding members together with their
total merchandise and manufacturing trade in mUSD from 1995 to 2020 (UNCTAD 2022;
UNSD 2022).16
Figure 4.1: Tariffs and Trade in the East African Community (EAC)
The EAC has also expanded outwards to contiguous countries of the region, with the
accession of Burundi and Rwanda in 2007, South Sudan in 2016, the Democratic Republic of
Congo in 2022 and also Somalia most recently in 2023. However, the three founding members
15 For instance, Tanzania postponed many substantial tariff line removals to the budgetary year beginning
July 1st 2001 and for sugar even until July 1st of 2002, which was the 7th highest valued import in the years
between 1996 and 2001 of all 96 chapters in the H1 nomenclature (UNCTAD 2022). See also Eberhard-Ruiz
and Moradi (2019) for a more detailed account on the tariff structure around the implementation.
16 The numbers reported reflect current dollar values of the respective year. We use import values as there are
some gaps in the reporting of exports values.
Notes: The plot displays trade volumes and tariff rates among the three EAC founding members Kenya, Tanzania
and Uganda. The unweighted average tariff on the dotted line represents the “AHS simple average tariff” and comes
from the United Nations Conference on Trade and Development (UNCTAD) Trade Analysis Information System
(TRAINS). The values between 1995 and 2000 are irregularly reported and stem from a single country only, hence
why the average is taken. The trade volume depicted by the shaded area represents the total USD value of all
goods imported within the EAC member states (net of re-imports) and stems from the United Nations Commodity
Trade Statistics Database. “Manufactures Trade Value” is comprised of trade of merchandise classified under
Section 6 of the SITC Revision 3. Both are facilitated via the World Integrated Trade Solution (WITS) platform.
125
still account for the overwhelming majority of economic activity with over 70% of the EAC’s
total GDP in 2022. As such and given that I aim to evaluate effects over the entire timeline of
the modern EAC, this paper concentrates primarily on evidence drawn from the founding
members Kenya, Tanzania, and Uganda.17 Since the establishment of the EAC in 2001 the
three economies have grown economically by a total of 111%, 86% and 47%, respectively.
However, with a per capita GDP (PPP) of 2,624$ and 2,280$ two of the three countries,
Tanzania and Uganda, are still categorized as “low-income-countries”.18 Only Kenya has
graduated to a “lower-middle-income country” as of 2014, with a current GDP per capita of
4,882$. Importantly, population growth within these countries over the same time period
roughly doubled from 90 million in 2000 to 166m in 2022.19 Concerning the economic structure
of the countries, they are heavily reliant on agriculture and services. In Kenya, the service
sector makes up a total of 48% of the GDP, followed by agriculture (38%) and manufacturing
(9%). Services also dominate in Tanzania (40% of GDP), who hosts a large tourism sector,
with agriculture making up 32%. Manufacturing is not as important in Tanzania with a
contribution to total GDP of 6%. In Uganda, the respective figures total to 27%, 52% and 9%
(WTO 2019).
Concerning trade, merchandise exports display a relevant contribution to the economies’
GDP with shares of 25%, 28 %, and 29%, in Kenya, Tanzania and Uganda, respectively (World
Bank 2022). However, regarding the direction of trade, the large majority of merchandise is
still sourced and exported to markets outside of the continent, with extra-African export and
import shares of 40-50% and 80-90%, respectively. China, India, as well as markets in the
Middle East and the EU have been the predominant trading partners within the last decade
(WTO 2019).20 As such, the share of intra-community (“intra-EAC”) trade has been relatively
low, hovering around 10% of total trade since its establishment in 2000 (UNU-CRIS 2019).21
Some of the reasons for the relatively low volumes of (documented) intra-regional trade in the
EAC are outlined by the substitutability of goods produced, several non-tariff barriers of trade,
infrastructural shortcomings but also the importance of informal cross-border trade (WTO
2019). However, compared to the other eight officially recognized regional economic
communities (RECs) on the continent, the EAC has the second highest intra-regional trade
share, trailing only the Southern African Development Community (SADC) whose members’
intra-regional trade account for 20% of total trade.22 Further, there is a significant asymmetry
17 Also owing to data availability, which is particularly scant for Burundi and Rwanda in the decades preceding
and following the re-establishment of the EAC.
18 GDP figures are expressed in constant 2017 international (PPP) USD.
19 The population of the three countries grew from 31m to 54m in Kenya, 34m to 65m in Tanzania and from
24m to 47m in Uganda.
20 The main goods exported are primary products (mainly agriculture produce) which make up 60%, 61%,
and 43% of total export value across the three countries respectively, and those declared manufactures 28%,
18% and 18% (WTO 2019).
21 Between 10 to 20% when including the trade in services (IMF 2023).
22 The eight RECs have an average intra-regional trade share of 6% (UNU-CRIS 2019).
126
in the pattern of trade, as only 6% of imports are sourced within the EAC, but 20% of countries’
exports are directed to markets of the EAC (WTO 2019). The predominant type of goods
traded within the original EAC members are comprised of primary products such as mineral
fuels and oils, gemstones as well as cereals but also manufactured goods such as rolled iron,
steel and steel products, vehicles and electrical machinery, plastic goods, processed food and
beverages as well as pharmaceuticals (UNCTAD 2022).
One particularly pertinent aspect of the three economies is their spatial configuration,
which is outlined by exceedingly high levels of urban economic primacy.23 While most of the
population is dependent on agriculture and lives in rural environments (70% in 2022, down
from 80% in 2000), the large majority of the countries’ economic activity is concentrated in the
three geographically limited primate cities Nairobi, Dar Es Salaam, and Kampala, respectively
(World Bank 2022). For instance, around the time of the EAC’s establishment in 2001, Dar es
Salaam hosted only 7% of the country’s population but 51% of formal employment of the
private sector and contributed to over 57% of the total wage bill (Tanzania National Bureau of
Statistics 2004, 2006).24 Considering that the administrative region of Dar es Salaam makes up
a mere 0.16% of Tanzania’s total land area, this describes a large intra-national (spatial)
discrepancy in economic activity. To compare, the next largest contributors to the wage bill in
2001 were Kilimanjaro, Arusha and Dodoma with 6%, 5% and 5%, respectively, and land shares
of 1.5% and 4% and 5%. This pattern has continued to persist and is particularly pronounced
in the high value-added manufacturing sector. For instance, in 2008, Dar es Salaam hosted 55%
of manufacturing establishments (30% of the manufacturing labor force) while contributing
over 51% of the country’s total value added (Tanzania National Bureau of Statistics 2010). In
the latest available survey of 2016, Dar es Salaam still contributed to over 41% to total value
added, albeit hosting a smaller share of 27% of all manufacturing establishments, and 32% of
the manufacturing workforce, which is however, well over twice the amount the next largest
region Morogoro.25 The structure of the EAC partner countries evinces the same spatial
pattern. Concerning Kenya, Nairobi accounted for 46% of (formal) wage employment in 2001
and for 51% of the total wage bill among main towns (Kenya Central Bureau of Statistics 2003;
Kenya National Bureau of Statistics 2011).26 In 2009, almost a decade later, these figures were
virtually unaltered. Together with the second largest industrial hub, Mombasa, these figures
increase to over 63% and 69% in 2009 for the employment and wage bill, respectively. Again,
23 Note the important difference from the “primacy” based purely on urban population shares most commonly
discussed.
24 For instance, around the time of the EAC’s establishment in 2002, Dar es Salaam hosted only 7% of the
country’s mainland population but contributed to over 40% of the total wage bill and hosted 57% of total
employment in the private sector (Tanzania National Bureau of Statistics 2006, 2007).
25 Dar es Salaam is also a hub for large firms, hosting over 33% of all firms sized over 100 employees and 13
out of the 44 firms over 500 employees. The second largest region Pwani, which encloses Dar es Salaam
geographically, hosts a mere 7% of such (Tanzania National Bureau of Statistics 2018).
26 Earnings in informal sector and rural small scale agriculture as well as pastoralists activities are excluded
(Kenya National Bureau of Statistics 2011; 236).
127
to compare, Nairobi makes up only 0.12% of the total land area and 7 (8%) of the population
as per the census of 1999 (2009) (Kenya Central Bureau of Statistics 2001; Kenya National
Bureau of Statistics 2011). Concerning the industrial structure, the main sectors clustered in
Nairobi are manufacturing, construction, and financial services, and in 2009 Nairobi hosted
49% of all manufacturing employment and 51% of the total manufacturing wage bill (Kenya
National Bureau of Statistics 2011, 2013).27 And lastly, in Uganda, Kampala hosted 45% of all
formal businesses establishments in 2001 and 2006, followed by Mbarara and Wakiso as the
second largest industrial cities with a share of 5% each (Uganda Bureau of Statistics 2003,
2007). If one includes the “Central” region of Uganda which encloses Kampala geographically,
the figure increases to 63% in 2001 and 65% in 2006.28 Similar to Nairobi and Dar Es Salaam,
Kampala contains the majority of the high value-added manufacturing sector with 42% of all
firms operating in Kampala and 61% together with the central region in 2006. As such, the
Kamapala region contributed to 47% of value added in 2006 and over 77% when including the
central district (Uganda Bureau of Statistics 2006).29 Similar to Nairobi and Dar es Salaam,
Kampala only makes up 0.09% of the total land area and 5% (4%) of the population in 2002
(2014) (Uganda Bureau of Statistics 2016).30
Note that high rates of urbanization are a not a particularly surprising characteristic of
developing countries and a research topic which has received increased attention as of late
(Glaeser 2014; Jedwab and Vollrath 2015; Gollin et al. 2016). However, for the present study,
next to its obvious connection to NEG models as introduced in Subsection 4.2.2, the urban
primacy as described in the EAC is particularly relevant in at least two further aspects. For
one, all three urban centers (Nairobi in Kenya, Dar es Salaam in Tanzania and Kampala in
Uganda) are geographically tucked away as viewed from the common borders connecting the
respective EAC partner state(s).31 In the data, the average road distance to internal EAC
border crossings for the three cities is 395km, 922km, and 269km, respectively.32 The connection
via road is particularly relevant for intra-EAC trade, as over 95% of the regional trade in the
area is transported via the road network, and only 5% via rail (see Figure C.8 and Nathan
Associates (2011)). Second, based purely on urban population shares, Kenya (30%) Tanzania
(36%) and Uganda (26%) show for relatively low levels of urban population primacy, compared
to most other (low & middle income) countries (World Bank 2022).33 Paired with the skewed
27 The respective figures for construction and financial services are 75% and 64% (Kenya National Bureau of
Statistics 2011).
28 Establishments with 5 employees or more. If one includes informal businesses, Kampala has contained 30%
and 29% of all businesses in 2001 and 2011 and 60% and 59% when including the central region, respectively
(Uganda Bureau of Statistics 2003, 2012).
29 Kampala also hosts the majority of large firms with 40% of firms with 100 employees or more in 2006
(Uganda Bureau of Statistics 2007). The central region also had the largest increase in manufacturing
businesses, with a 40% increase between 2001 to 2006.
30 Together with the Central region, this increases to 20%.
31 See Figure 4.2 of Subsection 4.4.2 for a stylized depiction of this spatial structure.
32 The minimum distances to the nearest EAC border crossings are 152km, 389km and 185km, respectively.
33 The share of urban population in high-income countries is 80%, the average of low- and middle-income
countries is 52%. Latin America
128
economic activity towards primate cities of the EAC, this is somewhat of a deviation from the
“urbanization without growth” hypothesis (Fay and Opal 2000). Lastly, recent research has
suggested natural resource-dependency to cause the emergence of “consumption” rather than
“production” cities (Gollin et al. 2016). Interestingly, in the last two decades, with the exception
of Uganda, Kenya and Tanzania have ranked below the 10% cutoff of natural resource rents to
GDP suggested in the paper (World Bank 2022). Hence, studying the household effects of trade
liberalization where the countries opening up to trade are outlined by the described spatial
structure provides a promising case which may inform the literature on the causes and
consequences of urbanization in developing contexts.
The key takeaway of the section is that the three countries joining in on a common trade
union are all outlined by stark regional economic inequalities, given by the presence of economic
hubs positioned away from the common borders of the community. To which extent this prior
spatial configuration has bearings on the effects of regional market integration is formalized
and discussed in the following section, in which I build a quantitative spatial equilibrium that
accommodates these stylized facts.
4.4 A Four Region Economy
To lay the theoretical groundwork on which to analyze the exposure of regional market
integration in the EAC across space, this section develops a canonical, four-region spatial
equilibrium model, which combines aspects from the models discussed in Subsection 4.2.2. The
model is built on Krugman’s (1991) core fundamentals while adding an external economy as
introduced by Krugman and Elizondo (1996). However, rather than the 2+1 cases in which the
external economy acts as one region (e.g. Paluzie 2001), I extend the foreign economy to two
regions (2+2) as in Monfort and Nicolini (2000) and Zeng and Zhao (2010). Finally, the model
is rendered unique as I tweak the structure of intra-national transport costs borrowing from
the 2+1 models of Crozet and Koenig (2004) and Brülhart et al. (2012) such that regions
within both countries are outlined by differential access to foreign markets, i.e. the model
encapsulates heterogeneous intra-national space in both economies. The model thereby allows
to additionally analyze the potential implications of foreign economic (in)equality on the
domestic distribution of activity in the context of increasing regional integration.34 As
anticipated in the previous section, this accommodates the stylized facts spatial features in how
trade may play out among the EAC member countries. Note that I refrain from
34 To my knowledge, (Behrens et al. 2006) were the first to introduce such heterogeneous space in a > 3 region
model and to analyze the potential effect of the foreign economic distribution on domestic outcomes. One
point of departure to the present model is that they do not explicitly model both countries as being “gated”,
i.e. only one of the countries outlined by heterogeneous intra-national space. And second, they make us of a
nonhomothetic preferences structure with a quasilinear utility introduced by (Ottaviano et al. 2002). Given
an implied decreasing expenditure share on manufactures, this has been deemed potentially inappropriate in
developing contexts (Fujita and Mori 2005) and I elect the standard utility function of the Cobb-Douglas /
CES type of the original models.
129
computationally more involved multi-region approaches, as the 2+2 case with heterogeneous
intra-national space encapsulates the stylized facts of the EAC and keeps the model tractable.
As such, we set-out with a four-region world economy consisting of 𝑅𝑅 locations denoted
by 𝑟𝑟= {1, 2, 3, 4}, where regions 1 and 2 are assigned to be in the “home” (from now
interchangeably referred to as “domestic”) country and regions 3 and 4 in the “foreign country”.
As will be specified in Subsection 4.4.2, regions 1 and 4 pose as the “core” (from now on
interchangeably referred to as “interior”) economic hubs while regions 2 and 3 are placed at
the border to formalize the stylized facts introduced above. Note that for most of the analysis
conducted across this section, I refer to effects on regions 1 and 2 of the domestic country.
However, by symmetry between the two countries, the results readily translate into the view
from the “foreign” country also, i.e. for regions 3 and 4.
Moving on with the model, there are two sectors in the economy, manufacturing, and
agriculture. The latter sector is characterized by perfect competition and produces the
homogenous agricultural good “food”, 𝐹𝐹, under constant returns to scale using the immobile,
inelastically supplied input “farmers”. The modern manufacturing sector is characterized by
monopolistic competition and thereby produces a variety of differentiated goods,
“manufactures”, using the input factor “workers” or “labor”. Farmers and workers within each
country are drawn from a total population mass 𝐿𝐿 of which 𝐿𝐿𝑀𝑀(𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎) = 𝛿𝛿· 𝐿𝐿 are engaged
in manufacturing and the rest 𝐿𝐿𝐹𝐹(𝑜𝑜𝑜𝑜𝑜𝑜) = (1 −𝛿𝛿)𝐿𝐿 in agriculture, hence 0 < 𝛿𝛿< 1.
Manufacturing workers are mobile between regions but not across sectors or countries, i.e. only
mobile between regions 1 and 2 or 3 and 4, respectively. As such, the total manufacturing
workforce within countries is fixed, but workers allocate themselves endogenously across regions
over time in response to real wage differentials. The respective shares of manufacturing of each
region are given by 𝜆𝜆𝑟𝑟, which satisfies ∑
𝜆𝜆𝑟𝑟= 1
𝑅𝑅
𝑖𝑖=𝑟𝑟
. We make the simplification that (𝜆𝜆1 +
𝜆𝜆2) = (𝜆𝜆3 + 𝜆𝜆4), such that the total manufacturing workforce of the two countries is equal,
albeit with the potential to be unequally distributed within. The distribution of the immobile
agricultural farmers is exogenously fixed and spread evenly across all regions such that their
respective shares across regions are given by 𝜙𝜙1 = 𝜙𝜙2 = 𝜙𝜙3 = 𝜙𝜙4 = 0.25. To ease notification,
we set the total population mass of the economy 𝐿𝐿 to 𝐿𝐿= 2 and assume countries to be of
equal size, i.e. 𝐿𝐿𝐹𝐹(𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜) = 𝐿𝐿𝐻𝐻(𝑜𝑜𝑜𝑜𝑜𝑜) = 1. As we will see later, this allows us to express the share
of the manufacturing workforce for each region in a country by a 𝜆𝜆, which is bound between
zero and one. This facilitates the interpretation of 𝜆𝜆 as a measure of the relative economic
disparity within a country and further eases interpretation down the line.
4.4.1. Consumer Preferences and Behavior
As in classical NEG models, a consumer decides how to spend her income 𝑌𝑌 with a preference
assumed to be of Cobb-Douglas type. In fact, all consumers have a preference representation
of Cobb-Douglas which combines a utility derived from the consumption of the agricultural
130
good, 𝐹𝐹, as well as a Dixit-Stiglitz (Dixit and Stiglitz 1977) Constant-Elasticity-of-Substitution
(CES) sub-utility for manufactures, 𝑀𝑀:
𝑈𝑈= 𝐹𝐹1−𝛿𝛿· 𝑀𝑀𝛿𝛿
(4.1)n
𝑀𝑀= °± 𝑐𝑐𝑖𝑖
𝜌𝜌
𝑛𝑛
𝑖𝑖=1
³
1𝜌𝜌
(4.2)n
𝑤𝑤𝑤𝑤𝑤𝑤ℎ 0 < 𝛿𝛿< 1 𝑎𝑎𝑎𝑎𝑎𝑎 0 < 𝜌𝜌< 1
Whereby 𝛿𝛿 denotes the share of income spent on consumption of the manufacturing variety
such that the share of income not spent on manufactures (1 −𝛿𝛿) is spent on the consumption
of food. Note that 𝛿𝛿 also represents the share of the population engaged in manufacturing. This
represents a normalization and is without consequence in the model as the shares are both
exogenously set parameters (Brakman et al. 2020). 𝑐𝑐𝑖𝑖 specifies the level of consumption of
manufacturing variety 𝑖𝑖 of a total of 𝑛𝑛 varieties, among which the consumer chooses with
elasticity 𝜌𝜌. 𝜌𝜌 is chosen to be constrained between 0 and 1 such that varieties are substitutable
but not perfect substitutes. 𝜌𝜌 is set to 𝜀𝜀=
1
1−𝜌𝜌 such that epsilon represents the elasticity of
substitution. From (4.2) it is immediate that 𝑀𝑀 is increasing more strongly in 𝑛𝑛 than in 𝑐𝑐
which reflects the well-known “love of variety” property, the strength of which is regulated by
𝜀𝜀. The consumer problem is then given by maximizing utility 𝑈𝑈 subject to the budget
constraint:
𝑌𝑌= 𝑝𝑝𝐹𝐹𝐹𝐹+ ± 𝑝𝑝𝑖𝑖𝑐𝑐𝑖𝑖
𝑛𝑛
𝑖𝑖=1
(4.3)n
Solving the consumer problem thereby involves first finding an optimal allocation of income 𝑌𝑌
on 𝐹𝐹 and 𝑀𝑀, and then, maximizing the sub-utility derived from consumption of the composite
index 𝑀𝑀 subject to the budget constraint for such manufacturing varieties which follows from
the first optimization problem. Hence, our first optimization problem is given by:
max 𝑈𝑈= 𝐹𝐹1−𝛿𝛿· 𝑀𝑀𝛿𝛿
𝑠𝑠. 𝑡𝑡. 𝑌𝑌= 𝑝𝑝𝐹𝐹𝐹𝐹+ ± 𝑝𝑝𝑖𝑖𝑐𝑐𝑖𝑖
𝑛𝑛
𝑖𝑖=1
Some algebra leads to the well-known result that consumers spend share 𝛿𝛿 of income 𝑌𝑌 on
manufactures, and (1 −𝛿𝛿)𝑌𝑌 on food:
131
𝑝𝑝𝐹𝐹𝐹𝐹= (1 −𝛿𝛿)𝑌𝑌
(4.4)n
± 𝑝𝑝𝑖𝑖𝑐𝑐𝑖𝑖
𝑛𝑛
𝑖𝑖=1
= 𝛿𝛿𝛿𝛿
(4.5)n
The next step involves finding the optimal spending among manufacturing varieties 𝑛𝑛, which
is encapsulated by the following optimization problem:
max 𝑀𝑀= °± 𝑐𝑐𝑖𝑖
𝜌𝜌
𝑛𝑛
𝑖𝑖=1
³
1𝜌𝜌
𝑠𝑠. 𝑡𝑡. ± 𝑝𝑝𝑖𝑖𝑐𝑐𝑖𝑖
𝑛𝑛
𝑖𝑖=1
= 𝛿𝛿𝛿𝛿
Taking the ratio of first order conditions for a pair of varieties, the maximization problem
yields the equality of marginal rates of substitution to price ratios:
𝑐𝑐𝑖𝑖
𝑝𝑝−1
𝑐𝑐𝑗𝑗
𝑝𝑝−1 = 𝑝𝑝𝑖𝑖
𝑝𝑝𝑗𝑗
or 𝑐𝑐𝑖𝑖= 𝑝𝑝𝑖𝑖
−𝜀𝜀· 𝑝𝑝𝑗𝑗
𝜀𝜀𝑐𝑐𝑗𝑗
(4.6)n
Once we substitute this result into the budget constraint for manufactures (4.5) we get:
± 𝑝𝑝𝑖𝑖𝑐𝑐𝑖𝑖
𝑛𝑛
𝑖𝑖=1
= ± 𝑝𝑝𝑖𝑖· º𝑝𝑝𝑖𝑖
−𝜀𝜀· 𝑝𝑝𝑗𝑗
𝜀𝜀𝑐𝑐𝑗𝑗»
𝑛𝑛
𝑖𝑖=1
= 𝑝𝑝𝑗𝑗
𝜀𝜀𝑐𝑐𝑗𝑗· ± 𝑝𝑝𝑖𝑖
1−𝜀𝜀
𝑛𝑛
𝑖𝑖=1
= 𝑐𝑐𝑗𝑗= 𝑝𝑝𝑗𝑗
−𝜀𝜀· 𝐼𝐼𝜀𝜀−1 · 𝛿𝛿𝛿𝛿
(4.7)n
𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢 𝐼𝐼≡°± 𝑝𝑝𝑖𝑖
1−𝜀𝜀
𝑛𝑛
𝑖𝑖=1
³
1
1−𝜀𝜀
(4.8)n
Hence, the expenditure needed to attain 𝑀𝑀 is:
𝑀𝑀= ¾± 𝑐𝑐𝑗𝑗
𝜌𝜌
𝑛𝑛
𝑗𝑗=1
¿
1𝜌𝜌
= ¾±(𝑝𝑝𝑗𝑗
−𝜀𝜀· 𝐼𝐼𝜀𝜀−1 · 𝛿𝛿𝛿𝛿)𝑝𝑝
𝑛𝑛
𝑗𝑗=1
¿
1𝜌𝜌
= 𝐼𝐼𝜀𝜀−1 · 𝛿𝛿𝛿𝛿¾±À𝑝𝑝𝑗𝑗
−𝜀𝜀𝜀𝜀Á
𝑛𝑛
𝑗𝑗=1
¿
1𝜌𝜌
,
𝑀𝑀= 𝐼𝐼𝜀𝜀−1 · 𝛿𝛿𝛿𝛿· 𝐼𝐼−𝜀𝜀
(4.9)n
132
Where we made use of that −𝜀𝜀𝜀𝜀= 1 −𝜀𝜀, and 1
𝑝𝑝= −𝜀𝜀
1−𝜀𝜀 , given that 𝜀𝜀=
1
1−𝑝𝑝. Given that 𝐼𝐼
multiplied by the quantity composite manufacturing consumption 𝑀𝑀 is equal to expenditure
𝛿𝛿𝛿𝛿, 𝐼𝐼 is also known as the price index, which measures the minimum cost of purchasing
manufacturing goods bundle 𝑀𝑀. Consumer demand functions are thereby:
𝐹𝐹= (1 −𝛿𝛿)𝑌𝑌
𝑝𝑝𝐹𝐹
(4.10)n
𝑀𝑀= 𝛿𝛿𝛿𝛿
𝐼𝐼
(4.11)n
Plugging these utility-maximizing consumption levels of F and M into (4.1) leads to the indirect
utility function:
𝑈𝑈= 𝛿𝛿𝛿𝛿(1 −𝛿𝛿)1−𝛿𝛿· 𝑌𝑌· 𝐼𝐼−𝛿𝛿(𝑝𝑝𝐹𝐹)−(1−𝛿𝛿)
(4.12)n
Hence, the maximum attainable welfare is a function of the income 𝑌𝑌 weighted by the cost of
living as given by price indices 𝐼𝐼 and 𝑝𝑝𝐹𝐹 together with their relative consumption shares δ and
1 −δ (Fujita et al. 2001).
4.4.2. Transport Costs and Heterogeneous intra-national Space
All manufacturing varieties can be consumed in each home or foreign location. However,
evidently, a variety locally consumed but not produced needs to be imported, which entails
transport costs. As is standard in NEG models, these transport costs are encapsulated by the
Samuelson-Von Thünen iceberg-type, which envisions only a fraction of the goods to arrive at
a destination, i.e. goods “melting” in transit (von Thünen 1826; Samuelson 1952). Thereby, a
producer located in region 1 has to dispatch an additional amount together with the demanded
amount, summing to 𝑇𝑇, for 1 𝑇𝑇
⁄ to arrive at the destination. For instance, if 20% of the
dispatched goods regularly melt away en-route between regions 𝑖𝑖 and 𝑗𝑗, iceberg transport costs
are given by 𝑇𝑇𝑖𝑖𝑖𝑖= 1.25. In other words, for one-unit of a good produced in region 𝑖𝑖 to arrive
at region 𝑗𝑗, suppliers located in region 1 have to dispatch 1.25 units of the good. Note at this
point that we assume food to be transported costlessly across all national and international
regions, which represents a standard simplification of NEG models.
133
As anticipated above, the present model is outlined by heterogeneous intra-national
space, which will be operationalized by a specific transport cost structure. The reason for this
adjustment is added realism on the one hand, but more importantly, because the spatial layout
of the EAC as anticipated in Section 4.3 lends itself naturally to this modification. Note again
that all three countries are outlined by economically dominating urban centers (Nairobi in
Kenya, Dar es Salaam in Tanzania and Kampala in Uganda) which are geographically tucked
away from the common borders connecting the respective EAC partner state(s) (see Figure 4.2
below for a stylized depiction).35 To operationalize this specific spatial layout of the EAC in
the model, I assume that among the two regions within each country, one of the regions has
better access to the foreign market, i.e. is a “border” or “gated” region (Crozet and Koenig
2004b; Behrens et al. 2006). As such, shipping goods from a non-border region to a foreign
location means transiting through this region, i.e. higher trade costs.36 This effectively places
the four regions on a line with regions 1 and 4 at the end of the spectrum and regions 2 and 3
connecting the two home and two foreign countries.
As expected, regions 1 and 4 represent the economic hubs of the countries, i.e. Nairobi,
Dar es Salaam and Kampala, and are denoted as “interior” or “core” regions. As in Brülhart
et al. (2012), I formalize this transport structure by accumulating all transport costs which
accrue throughout the transit, i.e. simply multiply all types iceberg transport costs 𝑇𝑇 which lie
between the origin and the destination region. For instance, for region 1, which is an interior
region, sending (importing) goods to (from) regions 2, 3, and 4 entails total iceberg transport
costs of 𝑇𝑇12 = 𝑇𝑇12, 𝑇𝑇13 = 𝑇𝑇12 · 𝑇𝑇23, and 𝑇𝑇14 = 𝑇𝑇12 · 𝑇𝑇23 · 𝑇𝑇34. I additionally assume that intra-
national transport costs in the home and foreign country are identical and that transport costs
are symmetric, such that 𝑇𝑇12 = 𝑇𝑇34 = 𝑇𝑇𝐷𝐷(𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜), 𝑇𝑇23 = 𝑇𝑇32 = 𝑇𝑇𝐹𝐹(𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜) and 𝑇𝑇𝑖𝑖𝑖𝑖= 𝑇𝑇𝑗𝑗𝑗𝑗. Note at
this point, the distinction made between domestic transport costs, 𝑇𝑇𝐷𝐷 and foreign transport
costs 𝑇𝑇𝐹𝐹. Domestic transport costs represent the general cost of transporting goods across
distance (e.g. logistics and freight), while foreign transport costs additionally entail tariffs, and
further frictions such as processing costs (at borders), as well as non-tariff barriers or even
cultural differences (Behrens et al. 2007). Finally, note that transport costs are zero when
consuming a variety produced within the same region, i.e. 𝑇𝑇𝑖𝑖𝑖𝑖= 1, for all 𝑖𝑖= 𝑗𝑗. As such, the
transport costs of trading goods between the four sending regions of our model 𝑖𝑖= {1, 2, 3, 4}
and arrival regions 𝑗𝑗= {1, 2, 3, 4} can be summarized by the following five types of total trade
costs across regions.
35 In the data, the average road distance to EAC border crossings for the three cities is 395km, 922km, and
269km, respectively. These road distances are particularly relevant for intra-EAC trade, as over 95% of the
regional trade in the area is in fact transported via the road network, and only 5% via rail (see Figure C.8
and Nathan Associates 2011).
36 Note that this is assumed to hold both for the export and import of goods.
134
1 = 𝑇𝑇𝑖𝑖𝑖𝑖Å𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 𝑖𝑖 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 𝑖𝑖
𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 𝑗𝑗 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 𝑗𝑗
𝑇𝑇𝐷𝐷= 𝑇𝑇𝑖𝑖𝑖𝑖Å𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 1 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 2
𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 3 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 4
𝑇𝑇𝐹𝐹= 𝑇𝑇𝑖𝑖𝑖𝑖Å𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 2 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 3
𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 3 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 2
𝑇𝑇𝐷𝐷𝐷𝐷≡𝑇𝑇𝐷𝐷· 𝑇𝑇𝐹𝐹= 𝑇𝑇𝑖𝑖𝑖𝑖Å𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 1,2 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 3,4
𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 3,2 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 1,4
𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷≡𝑇𝑇𝐷𝐷· 𝑇𝑇𝐹𝐹· 𝑇𝑇𝐷𝐷= 𝑇𝑇𝑖𝑖𝑖𝑖Å𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 1 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 4
𝑓𝑓𝑓𝑓𝑓𝑓 𝑖𝑖= 4 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗= 1
Figure 4.2 depicts this spatial cost structure for the 2+2 model illustratively using Uganda and
Kenya as a stylized example. The dashed line in the countries depicts the main trade route
between the countries called the “northern corridor” (Nathan Associates 2011).37 The vertically
dotted line illustrates the border.
Figure 4.2: Transport Cost Structure in the four-region Economy
37 Note that the figure is not drawn up to scale and serves as a stylized model of the spatial trade structure,
only (see Figure C.7 of the Appendix for a more accurate depiction of the geography as well as a depiction of
the “central corridor” which connects the countries via Tanzania). Country outline maps are from vemaps.com.
Uganda
1
2
Kenya
4
3
2
3
1
4
135
As is depicted in Figure 4.2, regions 1 and 4 represent the economic hubs Kampala and
Nairobi, respectively, with 2 and 3 posing as the “border” regions.38 Note that Tanzania borders
the depicted countries to the south, respectively, and given the position of Dar es Salaam,
creates a similar spatial pattern.39 As such, this transport cost structure is assumed to be
symmetric and thereby extends to the two other trade pairs, Tanzania-Uganda and Tanzania-
Kenya, analogously. Granted that this a simplification of the spatial realities on the ground,
including varying absolute and relative distances, differing processing times etc., this transport
cost structure is nonetheless useful because it easily lets us operationalize the comparative
statics of a change in regional market integration and the subsequent effect on (pre-existing)
regional disparities in Subsection 4.4.5 by solely altering the costs of moving goods between
regions 2 and 3, i.e. by altering 𝑇𝑇𝐹𝐹.
Carrying on with the model, these transport costs imply that the delivered price is 𝑇𝑇𝑖𝑖𝑖𝑖
higher than the f.o.b. price.40 A standard assumption I follow is that all transport costs are
incurred by consumers such that the total cost of consuming one-unit of variety produced in 𝑖𝑖
in region 𝑗𝑗 increases to 𝑝𝑝𝑖𝑖= 𝑝𝑝𝑖𝑖𝑇𝑇𝑖𝑖𝑖𝑖. Note that given (4.7), the demand for a variety produced in
region 𝑖𝑖, consumed in location 𝑗𝑗 is now given by:
𝑐𝑐𝑗𝑗= (𝑝𝑝𝑖𝑖𝑇𝑇𝑖𝑖𝑖𝑖)−𝜀𝜀· 𝐼𝐼𝑗𝑗
𝜀𝜀−1 · 𝛿𝛿𝑌𝑌𝑗𝑗
(4.13)n
Note that this necessitates the simplifying assumption that one manufacturing variety is
produced at one location only, which follows from internal economies of scale, and also, that
all varieties 𝑛𝑛 produced in this respective location are produced using the same technology and,
therefore, price. The total price index 𝐼𝐼 of region 𝑗𝑗 is then given by:
𝐼𝐼𝑗𝑗= °± 𝑛𝑛𝑖𝑖· º𝑝𝑝𝑖𝑖𝑇𝑇𝑖𝑖𝑖𝑖»(1−𝜀𝜀)
𝑅𝑅
𝑖𝑖=1
³
1
1−𝜀𝜀
(4.14)n
To arrive at the total sales of a given variety 𝑖𝑖, we sum demand for this variety over all regions
𝑅𝑅 using (4.13) and note that the supply incurs shipping 𝑇𝑇𝑖𝑖𝑖𝑖 units of 𝑖𝑖. Hence, we arrive at:
𝑞𝑞𝑖𝑖= 𝛿𝛿± 𝑌𝑌𝑗𝑗· (𝑝𝑝𝑖𝑖𝑇𝑇𝑖𝑖𝑖𝑖
𝑅𝑅
𝑗𝑗=1
)−𝜀𝜀· 𝐼𝐼𝑗𝑗
𝜀𝜀−1 · 𝑇𝑇𝑖𝑖𝑖𝑖
(4.15)n
38 Malaba is the main border-crossing connecting these countries (Nathan Associates 2011).
39 See detailed maps of these routes, i.e. the “northern corridor” as well as the “central corridor” connecting
the larger region in Figure C.7 in the Appendix from Nathan Associates (2011).
40 The “mill” or “f.o.b.”, free on board, price, is the price charged at the “mill”, the production location, not
incurring shipping costs.
136
This encapsulates that total demand of a variety 𝑞𝑞𝑖𝑖 is decreasing in the price of the good 𝑝𝑝𝑖𝑖
and the transport cost incurred 𝑇𝑇𝑖𝑖𝑖𝑖 for the respective importing region. Demand is increasing
in income 𝑌𝑌𝑗𝑗 and price index 𝐼𝐼𝑗𝑗 of regions as well as in the share spent on manufactures 𝛿𝛿.
4.4.3. Producer Behavior
As defined previously, food is produced under constant returns to scale as well as under the
assumption of perfect competition. Given that we have just assumed food to be traded
costlessly across all regions, the price of food is equal everywhere, and so is the wage given that
farmers are paid their marginal product. We then set the technology coefficient of food
production to 1 such that 𝑤𝑤𝐹𝐹= 𝑝𝑝𝐹𝐹= 1 and the agricultural good acts as the numeraire
throughout the analysis. In the manufacturing sector, production technology is of increasing
returns to scale. It thereby involves a fixed cost of production 𝐹𝐹 and marginal costs per unit
𝑐𝑐. Given that labor is our only input factor, the production of a quantity 𝑞𝑞 of a variety 𝑖𝑖
produced in location 𝑖𝑖 is given by labor input requirement:
𝑙𝑙= 𝐹𝐹+ 𝑐𝑐𝑐𝑐
(4.16)n
and this is assumed to be the same technology for all varieties. Given increasing returns to
scale, consumer preference for variety, firms will choose to produce a variety, not produced by
any other firm such that a variety is produced only in one location by one firm.41 This has the
result that the number of available varieties is equal to the number of firms. The profit of a
specific firm producing at location 𝑖𝑖 with a given wage rate 𝑤𝑤𝑖𝑖, and an f.o.b. price 𝑝𝑝𝑖𝑖 is:
𝜋𝜋𝑖𝑖= 𝑝𝑝𝑖𝑖𝑞𝑞𝑖𝑖−𝑤𝑤𝑖𝑖(𝐹𝐹+ 𝑐𝑐𝑞𝑞𝑖𝑖)
(4.17)n
Making the simplification 𝑞𝑞= 𝐵𝐵𝑝𝑝𝑖𝑖
−𝜀𝜀 (see Brakman et al. 2020) and differentiating w.r.t. price
and setting equal to zero leads to the 𝑓𝑓. 𝑜𝑜. 𝑐𝑐.:
(1 −𝜀𝜀)𝐵𝐵𝑝𝑝𝑖𝑖
1−𝜀𝜀+ 𝜀𝜀𝑤𝑤𝑖𝑖· 𝑐𝑐𝑐𝑐𝑝𝑝𝑖𝑖
−𝜀𝜀−1 = 0
(4.18)n
Rearranging leads us to the well-known result that prices are a combination of f.o.b. price,
which are given by marginal costs 𝑤𝑤𝑖𝑖𝑐𝑐, and a mark-up, determined by the elasticity of
substitution 𝜀𝜀:
𝑝𝑝𝑖𝑖º1 −1
𝜀𝜀» = 𝑤𝑤𝑖𝑖𝑐𝑐, or
41 Where an additional assumption is that the number of varieties goes to infinity (Fujita et al. 2001).
137
𝑝𝑝𝑖𝑖= 𝑐𝑐𝑤𝑤𝑖𝑖
𝜌𝜌
(4.19)n
Given that we assume free entry and exit, profits are driven to zero. Using the new pricing rule
(4.19) in the profit function (4.17) and setting to zero leads:
𝜋𝜋𝑖𝑖= 𝑐𝑐𝑤𝑤𝑖𝑖
𝜀𝜀−1 (𝑞𝑞𝑖𝑖−𝐹𝐹(𝜀𝜀−1)
𝑐𝑐
)
(4.20)n
Hence, equilibrium output by any active firm 𝑖𝑖 is the constant:
𝑞𝑞∗= 𝐹𝐹
𝑐𝑐(𝜀𝜀−1)
(4.21)n
and the required labor input producing this amount is then given by plugging (4.21) into the
production technology used (4.16):
𝑙𝑙∗= 𝐹𝐹+ 𝑐𝑐È𝐹𝐹
𝑐𝑐(𝜀𝜀−1)É , 𝑜𝑜𝑜𝑜
𝑙𝑙∗= 𝐹𝐹𝐹𝐹
(4.22)n
Which carries the result that the number of varieties 𝑛𝑛 produced in a location 𝑖𝑖, and thereby
the number of manufacturing firms, is directly proportional to the manufacturing population
at this location, 𝜆𝜆𝑖𝑖𝛿𝛿𝛿𝛿:
𝑛𝑛𝑖𝑖= 𝜆𝜆𝑖𝑖𝛿𝛿𝛿𝛿
𝐹𝐹𝐹𝐹
(4.23)n
4.4.4. Short-run Equilibrium
In equilibrium, the output of firms must match demand by consumers. Using (4.15) we have:
𝑞𝑞𝑖𝑖
∗= 𝛿𝛿± 𝑌𝑌𝑗𝑗 · 𝑝𝑝𝑖𝑖
−𝜀𝜀· 𝑇𝑇𝑖𝑖𝑖𝑖
1−𝜀𝜀
𝑅𝑅
𝑗𝑗=1
· 𝐼𝐼𝑗𝑗
𝜀𝜀−1
(4.24)n
In other words, firms break even if the price they charge equals:
138
𝑝𝑝𝑖𝑖
𝜀𝜀= 𝛿𝛿
𝑞𝑞𝑖𝑖
∗± 𝑌𝑌𝑗𝑗 · 𝑇𝑇𝑖𝑖𝑖𝑖
1−𝜀𝜀
𝑅𝑅
𝑗𝑗=1
· 𝐼𝐼𝑗𝑗
𝜀𝜀−1
(4.25)n
Plugging in the pricing rule (4.19), leads to the well-known wage equation:
𝑤𝑤𝑖𝑖= Ê𝜀𝜀−1
𝜀𝜀𝜀𝜀Ë È 𝛿𝛿
𝑞𝑞𝑖𝑖
∗ ± 𝑌𝑌𝑗𝑗 · 𝑇𝑇𝑖𝑖𝑖𝑖
1−𝜀𝜀
𝑅𝑅
𝑗𝑗=1
· 𝐼𝐼𝑗𝑗
𝜀𝜀−1É
1𝜀𝜀
(4.26)n
To arrive at real wages, 𝜔𝜔, we simply have to divide nominal wages (4.26) by the cost of living,
which is a combination of the manufacturing price index of the region (4.14) and food prices:
𝜔𝜔𝑖𝑖= 𝑤𝑤𝑖𝑖· 𝐼𝐼−𝜀𝜀· (𝑝𝑝𝐹𝐹)−(1−𝛿𝛿)
(4.27)n
It is convenient to use some normalizations to simplify analysis (Fujita et al. 2001). Hence, we
redefined the marginal labor requirement is:
𝑐𝑐= 𝜀𝜀−1
𝜀𝜀
= 𝜌𝜌
(4.28)n
Then, (4.19) turns to:
𝑝𝑝𝑖𝑖= 𝑤𝑤𝑖𝑖
(4.29)n
Also, we set a unit of measurement for the number of firms n, such that the fixed input
requirement 𝐹𝐹 is given by:
𝐹𝐹= 𝛿𝛿
𝜀𝜀
(4.30)n
Remember that the number of firms in each location is directly proportional to the
manufacturing labor force in this location 𝜆𝜆𝑖𝑖𝛿𝛿𝛿𝛿, such that (4.23) reduces to:
𝑛𝑛𝑖𝑖= 𝜆𝜆𝑖𝑖𝛿𝛿𝛿𝛿
𝐹𝐹𝐹𝐹= 𝜆𝜆𝑖𝑖𝐿𝐿
(4.31)n
139
From this, the price index (4.14) as well as the wage equation can be simply expressed as:
𝐼𝐼𝑗𝑗= °± 𝜆𝜆𝑖𝑖𝐿𝐿º𝑤𝑤𝑖𝑖· 𝑇𝑇𝑖𝑖𝑖𝑖»(1−𝜀𝜀)
𝑅𝑅
𝑖𝑖=1
³
1
1−𝜀𝜀
(4.32)n
𝑤𝑤𝑖𝑖= ȱ𝑌𝑌𝑗𝑗· 𝑇𝑇𝑖𝑖𝑖𝑖
1−𝜀𝜀·
𝑅𝑅
𝑗𝑗=1
𝐼𝐼𝑗𝑗
𝜀𝜀−1É
1𝜀𝜀
(4.33)n
These constitute the first two of three equations that characterize the short-run equilibrium.
What is missing is the income-determining equation, which is defined by the sum of wage
income from manufacturing workers in the region 𝜆𝜆𝑖𝑖𝛿𝛿𝛿𝛿 as well as from farmers 𝜙𝜙𝑖𝑖(1 −𝛿𝛿)𝐿𝐿.42
Hence, the income of a region 𝑖𝑖 is given 𝑌𝑌𝑖𝑖= 𝜆𝜆𝑖𝑖· 𝑤𝑤𝑖𝑖· 𝛿𝛿𝛿𝛿+ 𝜙𝜙𝑖𝑖(1 −𝛿𝛿)𝐿𝐿. Taking into account our
initial simplifications, namely that the manufacturing workforce is immobile across countries
and exogenously set to 𝜙𝜙= 0.25, that the distributions of the manufacturing workforce is given
by ∑
𝜆𝜆𝑟𝑟= 1
4
𝑖𝑖=1
, that the total mass of population is set to 𝐿𝐿= 2 and that the two countries
are of equal size lets us write the income equation in our four region case as:
𝑌𝑌𝑖𝑖= 𝜆𝜆𝑖𝑖· 𝑤𝑤𝑖𝑖· 𝛿𝛿+ (1 −𝛿𝛿)
2
, 0 ≤𝜆𝜆𝑖𝑖≤1
(4.34)n
Where we additional use of our assumption 𝜆𝜆1 + 𝜆𝜆2 = 𝜆𝜆3 + 𝜆𝜆4. And similarly, the price index
simplifies to:
𝐼𝐼𝑗𝑗= °± 𝜆𝜆𝑖𝑖º𝑤𝑤𝑖𝑖· 𝑇𝑇𝑖𝑖𝑖𝑖»(1−𝜀𝜀)
𝑅𝑅
𝑖𝑖=1
³
1
1−𝜀𝜀
, 0 ≤𝜆𝜆𝑖𝑖≤1
(4.35)n
Given that manufactures can be traded across all regions, and our economy is made up of four
regions in total, the short-run equilibrium relationship is expressed by 12 equations (3 for each
region) given in (4.36) through (4.47):
𝑌𝑌1 = 𝜆𝜆1 · 𝑤𝑤1 · 𝛿𝛿+ (1 −𝛿𝛿)
2
(4.36)n
𝑌𝑌2 = 𝜆𝜆2 · 𝑤𝑤2 · 𝛿𝛿+ (1 −𝛿𝛿)
2
(4.37)n
42 Note that given constant returns to scale and perfect competition, the wages for agricultural labor are equal
everywhere and set as the numeraire as seen before.
140
𝑌𝑌3 = 𝜆𝜆3 · 𝑤𝑤3 · 𝛿𝛿+ (1 −𝛿𝛿)
2
(4.38)n
𝑌𝑌4 = 𝜆𝜆4 · 𝑤𝑤4 · 𝛿𝛿+ (1 −𝛿𝛿)
2
(4.39)n
𝐼𝐼1 = Î𝜆𝜆1 · 𝑤𝑤1
1−𝜀𝜀+ 𝜆𝜆2(𝑤𝑤2 · 𝑇𝑇𝐷𝐷)(1−𝜀𝜀) + 𝜆𝜆3(𝑤𝑤3 · 𝑇𝑇𝐷𝐷𝐷𝐷)(1−𝜀𝜀) + 𝜆𝜆4(𝑤𝑤4 · 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷)(1−𝜀𝜀)Ï
1
1−𝜀𝜀
(4.40)n
𝐼𝐼2 = Î 𝜆𝜆1(𝑤𝑤1 · 𝑇𝑇𝐷𝐷)(1−𝜀𝜀) + 𝜆𝜆2 · 𝑤𝑤2
1−𝜀𝜀+ 𝜆𝜆3(𝑤𝑤3 · 𝑇𝑇𝐹𝐹)(1−𝜀𝜀)𝜆𝜆4 + (𝑤𝑤4 · 𝑇𝑇𝐷𝐷𝐷𝐷)(1−𝜀𝜀)Ï
1
1−𝜀𝜀
(4.41)n
𝐼𝐼3 = Î 𝜆𝜆1(𝑤𝑤1 · 𝑇𝑇𝐷𝐷𝐷𝐷)(1−𝜀𝜀) + 𝜆𝜆2(𝑤𝑤2 · 𝑇𝑇𝐹𝐹)(1−𝜀𝜀) + 𝜆𝜆3 · 𝑤𝑤3
1−𝜀𝜀+ 𝜆𝜆4(𝑤𝑤4 · 𝑇𝑇𝐷𝐷)(1−𝜀𝜀)Ï
1
1−𝜀𝜀
(4.42)n
𝐼𝐼4 = Î 𝜆𝜆1(𝑤𝑤1 · 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷)(1−𝜀𝜀) + 𝜆𝜆2(𝑤𝑤2 · 𝑇𝑇𝐷𝐷𝐷𝐷)(1−𝜀𝜀) + 𝜆𝜆3(𝑤𝑤3 · 𝑇𝑇𝐷𝐷)(1−𝜀𝜀) + 𝜆𝜆4 · 𝑤𝑤4
1−𝜀𝜀Ï
1
1−𝜀𝜀
(4.43)n
𝑤𝑤1 = º𝑌𝑌1 · 𝐼𝐼1
𝜀𝜀−1 + 𝑌𝑌2 · 𝑇𝑇𝐷𝐷
1−𝜀𝜀· 𝐼𝐼2
𝜀𝜀−1 + 𝑌𝑌3 · 𝑇𝑇𝐷𝐷𝐷𝐷
1−𝜀𝜀· 𝐼𝐼3
𝜀𝜀−1 + 𝑌𝑌4 · 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷
1−𝜀𝜀· 𝐼𝐼4
𝜀𝜀−1»
1𝜀𝜀
(4.44)n
𝑤𝑤2 = º𝑌𝑌1 · 𝑇𝑇𝐷𝐷
1−𝜀𝜀· 𝐼𝐼1
𝜀𝜀−1 + 𝑌𝑌2 · 𝐼𝐼2
𝜀𝜀−1 + 𝑌𝑌3 · 𝑇𝑇𝐹𝐹
1−𝜀𝜀· 𝐼𝐼3
𝜀𝜀−1 + 𝑌𝑌4 · 𝑇𝑇𝐷𝐷𝐷𝐷
1−𝜀𝜀· 𝐼𝐼4
𝜀𝜀−1»
1𝜀𝜀
(4.45)n
𝑤𝑤3 = º𝑌𝑌1 · 𝑇𝑇𝐷𝐷𝐷𝐷
1−𝜀𝜀· 𝐼𝐼1
𝜀𝜀−1 + 𝑌𝑌2 · 𝑇𝑇𝐹𝐹
1−𝜀𝜀· 𝐼𝐼2
𝜀𝜀−1 + 𝑌𝑌3 · 𝐼𝐼3
𝜀𝜀−1 + 𝑌𝑌4 · 𝑇𝑇𝐷𝐷
1−𝜀𝜀· 𝐼𝐼4
𝜀𝜀−1»
1
𝜀𝜀
(4.46)n
𝑤𝑤4 = º𝑌𝑌1 · 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷
1−𝜀𝜀· 𝐼𝐼1
𝜀𝜀−1 + 𝑌𝑌2 · 𝑇𝑇𝐷𝐷𝐷𝐷
1−𝜀𝜀· 𝐼𝐼2
𝜀𝜀−1 + 𝑌𝑌3 · 𝑇𝑇𝐷𝐷
1−𝜀𝜀· 𝐼𝐼3
𝜀𝜀−1 + 𝑌𝑌4 · 𝐼𝐼4
𝜀𝜀−1»
1
𝜀𝜀
(4.47)n
These 12 equilibrium conditions formalize the notion of centripetal (demand and cost linkages)
as well as centrifugal forces (competition) anticipated in Subsection 4.2.2. Take first, the price
index given in equations (4.40) through (4.43). Consumer prices at one particular location can
be seen as a weighted average of all source location sizes (𝜆𝜆) and prizes (which, given (4.19)
are directly proportional to the wage rate 𝑤𝑤) with weights given by the distance to these
exporting locations (𝑇𝑇), respectively.43 As such, the price index is lower in those regions, where
a higher share of demand is sourced from large (high 𝜆𝜆), low wage (low 𝑤𝑤) and importantly,
nearby locations (low 𝑇𝑇); and of course, most cheaply sourced locally, i.e. when 𝑇𝑇= 1. In other
words, locations with large shares of own or close by manufacturing employment have lower
price indices given that a smaller share of the total consumption needs to be imported; this is
the “price index effect” analytically derived in Fujita et al. (2001). These dynamics describe
the cost (forward) linkage described in Subsection 4.2.2 whereby a larger home market provides
lower consumer prices. As such, moving to a region, i.e. making it larger, displays a self-
reinforcing centripetal force.
43 Consumer prices at one particular location can be seen as a weighted average of all source locations and
their prizes (which, given (4.19) are directly proportional to the wage rate 𝑤𝑤) as well as distance 𝑇𝑇 with
weights given by the relative size of these locations 𝜆𝜆.
141
The wage equations given in (4.44) through (4.47) can be interpreted similarly. In essence,
wages are higher in regions where income 𝑌𝑌, and thereby expenditure, is high or in regions
where these larger markets are more proximate (low 𝑇𝑇). Put simply, firms are able to pay
higher wages if they have better access to large markets. This describes the demand (backward)
linkage anticipated before and indicates that a larger number of workers, and thereby,
consumers, increase the local demand which increases the wage firms are able to pay. Similarly,
as for the cost linkage, this attracts more workers to this region, and also firms, thereby acting
as a self-reinforcing centripetal force. This is described by the “home market effect” (for the
full derivation, see Fujita et al. 2001). Importantly, the wage equation also encapsulates a
centrifugal force which is given by its positive dependence on the price index 𝐼𝐼. As just
established, the price index is lower in larger regions, i.e. those with a higher number of
manufacturing workers, and thereby, varieties. And given that the number of manufactures is
regulated not by output per firm, but by the number of firms, a lower price index automatically
indicates a larger number of competing firms, which exerts a downward pressure on the wages
a firm is able to pay.44 As a results, firms may seek to relocate in order to shelter from
competition allowing them to pay higher wages, which may also draw workers.
In the end, the relative strength of these centrifugal and centripetal forces can be handily
manifested in real wage differentials across regions, which combine the effects on nominal wages
and prices. Formally, the real wages 𝜔𝜔 of regions are given by dividing the total wage income
𝑤𝑤 by the consumer price index of both manufactures 𝐼𝐼 and food 𝐹𝐹 together with their relative
consumption shares 𝛿𝛿, hence 𝜔𝜔𝑖𝑖= 𝑤𝑤𝑖𝑖· 𝐼𝐼−𝛿𝛿(𝑝𝑝𝐹𝐹)−(1−𝛿𝛿). Note that we are able to dismiss the
component of the agricultural good, as it is set as the numeraire. Real wages of all four regions
are then expressed by:
𝜔𝜔1 = 𝑤𝑤1 · 𝐼𝐼1
−𝛿𝛿
(4.48)n
𝜔𝜔2 = 𝑤𝑤2 · 𝐼𝐼2
−𝛿𝛿
(4.49)n
𝜔𝜔3 = 𝑤𝑤3 · 𝐼𝐼3
−𝛿𝛿
(4.50)n
𝜔𝜔4 = 𝑤𝑤4 · 𝐼𝐼4
−𝛿𝛿
(4.51)n
Where the values of the right-hand side are given by the simultaneous solution to the 12 short-
run equilibrium conditions (4.36) through (4.47). In the long run, we assume that workers
respond to the real wage differential across regions by migrating such that the share of
manufacturing workers within the two home and foreign economies, 𝜆𝜆1 and 𝜆𝜆2, as well as 𝜆𝜆3
and 𝜆𝜆4, are endogenously determined. I assume workers to move between regions with the
following dynamics:
44 This also be validated in (4.7) or (4.15), where demand of an individual firm is inversely related to the price
index.
142
𝑑𝑑𝜆𝜆𝑖𝑖
𝑑𝑑𝑑𝑑= 𝛾𝛾
⎩
Ó
Ó
Ó
Ó
⎨
Ó
Ó
Ó
Ó
⎧
𝜔𝜔𝑖𝑖
𝜔𝜔𝑗𝑗
−1 𝑖𝑖𝑖𝑖 0 < 𝜆𝜆𝑖𝑖< 1
𝑚𝑚𝑚𝑚𝑚𝑚Ø0, 𝜔𝜔𝑖𝑖
𝜔𝜔𝑗𝑗
−1Ù 𝑖𝑖𝑖𝑖 𝜆𝜆𝑖𝑖= 1
𝑚𝑚𝑚𝑚𝑚𝑚Ø0, 𝜔𝜔𝑖𝑖
𝜔𝜔𝑗𝑗
−1Ù 𝑖𝑖𝑖𝑖 𝜆𝜆𝑖𝑖= 0
(4.52)n
Hence, for a given real wage differential and spatial configuration 𝜆𝜆𝑖𝑖, workers move between
regions across regions with a particular speed 𝛾𝛾. We now have all the ingredients we need to
define a long-run equilibrium as well. By (4.52), the first type of long-run equilibrium can be
described by a spatial configuration for which real wages across regions are equalized, i.e. a
situation in which workers have no incentive to move. Formally, this is given by a 𝜆𝜆 𝜖𝜖 [0,1] for
which 𝜔𝜔𝑖𝑖/𝜔𝜔𝑗𝑗= 1. One specific case of such is the equal spreading of workers, i.e. for our four-
region model 𝜆𝜆1,2 = 𝜆𝜆3,4 = 0.5 and 𝜔𝜔1,3/𝜔𝜔2,4 = 1. This is also called the “symmetric” or
spreading” equilibrium. The model also admits a second type of a long-run equilibrium, one in
which real wages are not equalized. In these cases, all of the manufacturing workforce is
agglomerated in one of the regions, which represents a corner solution. Formally, such an
equilibrium is given by 𝜆𝜆1,4 = 1 and 𝜆𝜆2,3 = 0, and often referred to as an “agglomerated” or
“core-periphery” equilibrium. To complete the discussion on long-run equilibria, one important
distinction to make is whether such an equilibrium is also a stable one. In general, the stability
of an equilibria depends on whether a small perturbation in the manufacturing workforce at
this spatial configuration triggers dynamics which reinstates the just-left allocation of workers
or not. For the first type of equilibria, the stability is thereby defined by a second condition
which is that the derivative of the real wage differential w.r.t. an infinitesimal change in the
manufacturing workforce is smaller or equal to zero, i.e. formally whether 𝑑𝑑(𝜔𝜔𝑖𝑖/𝜔𝜔𝑗𝑗)/𝑑𝑑𝜆𝜆𝑖𝑖≤0.
Put simply, if migrating from region 𝑗𝑗 to region 𝑖𝑖 increases the real wage differential 𝜔𝜔𝑖𝑖/𝜔𝜔𝑗𝑗,
then the previous equilibrium was not a stable one. In the second type of equilibria, the stability
condition entails that the real wage differential is skewed in favor of the agglomerated region,
such that for 𝜆𝜆1 = 1 , 𝜔𝜔1
𝜔𝜔2 ≥1 and for 𝜆𝜆1 = 0 and 𝜔𝜔1
𝜔𝜔2 ≤1. I analyze the stability of these two
types of equilibria more thoroughly in Appendix C.1.
4.4.5. Spatial Equilibria and Regional Trade Liberalization
The four-region model and the short- as well as long-run equilibrium conditions just developed
lends itself to the comparative static examination of what happens to the forces inducing
agglomeration or dispersion once trade across countries in the EAC is liberalized. In particular,
we can use the solutions to the simultaneous equilibrium conditions (4.36) through (4.47) as
inputs to compute the real-wage differential arising for a given spatial structure and transport
143
costs regiment which dictates the dynamic process towards a stable long-run equilibrium
described in (4.52).
The analysis in this section thereby entails tracking what happens to the (short-run) real
wage differential across regions inside the countries once the costs connecting the two economies
𝑇𝑇𝐹𝐹 are lowered from a former prohibitive level (i.e. autarky) down to levels which mirror those
incurred within the respective countries, i.e. 𝑇𝑇𝐹𝐹= 𝑇𝑇𝐷𝐷.45 This will effectively allow us to analyze
how the process of trade liberalization affects the (stability of) specific (long-run) equilibrium
allocations of workers across regions. Note, however, given that the real wage differential
𝜔𝜔𝑖𝑖/𝜔𝜔𝑗𝑗 depends on twelve simultaneous non-linear equations, the real wage differential is not a
simple function of 𝜆𝜆𝑖𝑖. As such, and as is common in the NEG literature, I will analyze the
dynamics of the spatial equilibria mainly via numerical simulations. This is most efficiently
done by plotting the real wage differential 𝜔𝜔1/𝜔𝜔2 across the full range of potential
manufacturing distributions 𝜆𝜆 𝜖𝜖 [0,1] which may be realized at any point in time. To nonetheless
provide some analytical insights into the numerical results, Appendix C.1 provides a “sustain”
and “break” analysis in the vein of Fujita et al. (2001), which revolves around assessing the
stability of the two specific types of equilibria described above, i.e. “agglomeration” and
“spreading”. Some of the key results of this analysis are discussed in this section as well.
As a final remark on the approach of this section’s analysis, it turns out to be instructive
to compare the results of the model to a more general version of it. To be specific, I will conduct
the simulations additionally for a four-region model with homogenous (or symmetric) intra-
national space. This model mirrors the one described by equations (4.36) through (4.47), but
with a tweak regarding the transport cost structure. This is done by simply setting the three
different types of external iceberg trade costs equal, i.e. 𝑇𝑇𝐹𝐹= 𝑇𝑇𝐷𝐷𝐷𝐷= 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷, such that the two
home regions have identical (international) trade costs to both foreign regions. In the vein of
Figure 4.2, one can think of regions 1 and 2 as well as 3 and 4 in this adjusted model as placed
on a line parallel, rather than perpendicular to the border with roads diagonally connecting
the home and foreign regions, respectively. Note that the model thereby reduces to the one
studied in Monfort and Nicolini (2000), and their conclusions apply analogously. However,
comparing the results from such a model to the present one provides an intuitive reference and
helps in evaluating the additional role heterogeneous intra-national space (from now on
interchangeably referred to as asymmetric space), and thereby unequal access to the newly
integrated foreign markets, plays, particularly for an integration process of two highly unequal
countries, which describes the (stylized) setting of the East African Community’s re-
establishment.
Figure 4.3 below initiates our analysis and plots the real wage differential between regions
1 and 2 across the full range of possible manufacturing distributions λ ϵ [0,1] as well as for
45 In this scenario, the cost associated with trading goods across borders mirrors those incurred when shipping
goods intra-nationally, i.e. 𝑇𝑇𝐹𝐹= 𝑇𝑇𝐷𝐷.
144
three levels of international trade costs 𝑇𝑇𝐹𝐹, respectively. Note that solving for this set
necessitates a choice on the exogenous parameter values given by 𝛿𝛿 and 𝜀𝜀, and the intra-national
trade costs 𝑇𝑇𝐷𝐷. I use values commonly employed in the literature which are given below the
figures. Table C.1 in the Appendix provides additional sensitivity tests using a wider range of
values and includes a discussion thereof. The main interpretations remain. Notice that although
our model entails four endogenous parameters, 𝜆𝜆1, 𝜆𝜆2, 𝜆𝜆3, and 𝜆𝜆4, the plots in Figure 4.3 only
depict two at a time, i.e. is two-dimensional. This is done by setting 𝜆𝜆3 = 𝜆𝜆4 = 0.5, i.e. by
initially assuming an equal distribution of manufacturing in the foreign country. This
assumption is relaxed later in Figure 4.5, when we assess the moderating influence of foreign
economic inequality. Note lastly, that as anticipated earlier, we focus our view on the home
country, i.e. regions 1 and 2, but given symmetry, the results and intuitions apply identically
for regions 4 and 3 Figure C.3 and Figure C.4 of the Appendix also provide the full three-
dimensional plots, which effectively combine the results depicted in Figure 4.3 and Figure 4.5.
Panel A
Panel B
Panel C
Panel D
Figure 4.3: Trade Liberalization and Spatial Equilibria
145
The numerical simulations depicted in Figure 4.3 provide the main insights into the
process of trade liberalization across a four-region economy, i.e. lowering the intra-national
trade costs 𝑇𝑇𝐹𝐹. It depicts the results for two spatial setups (the symmetric and asymmetric
intra-national space) for two sets of parameter values, respectively.46 More specifically, Panels
A and C represent the symmetric case (homogenous intra-national space) for values of the
elasticity of substitution 𝜀𝜀= 6 and 𝜀𝜀= 4. And Panels A and D present the results analogously
for the asymmetric case (heterogeneous intra-national space).
We focus first on the real wage differentials in autarky, i.e. where international trade
costs are prohibitively high 𝑇𝑇𝐹𝐹= ∞, as depicted by the solid line. In the case of low product
differentiation (𝜀𝜀= 6), Panels A and B, we notice that there exists a long-run stable symmetric
equilibrium where the workforce is equally spread across the two home regions for both models,
as can be seen by the negative slope passing-through real wage parity. While this equilibrium
also exists for the case where product differentiation is high (Panel C and D), this equilibrium
is not stable anymore, as can be depicted by the positive slope through the point where 𝜆𝜆1 =
0.5. What happens to this type of equilibria in the home country when the external trade costs
to regions 3 and 4 are lowered? This is depicted by the new short-run equilibrium real wage
differentials given by the dashed (𝑇𝑇𝐹𝐹= 2.00) and dotted lines (𝑇𝑇𝐹𝐹= 𝑇𝑇𝐷𝐷= 1.60).47 As a first
pass through the Panels, and as shown in previous results, lowering the costs to trade with an
external market increases agglomerating tendencies, i.e. increases intra-national inequality (e.g.
Monfort and Nicolini 2000; Paluzie 2001). This can be seen by a general attenuation of the
slopes passing through the symmetric equilibria. Most starkly, in Panel A, the slope concludes
a full rotation from negative to positive values from autarky to free trade. Hence, when the
countries are liberalized, the former stable equilibrium for equal distribution of manufacturing
activity turns out to be unstable. As we defined in the previous section, this is so because an
infinitesimal small shock (increase) to the manufacturing workers in any direction would also
cause a higher real wage skewed towards this region, which would not induce workers to move
back to the symmetric equilibrium. As such, once trade is liberalized, the strength of the force
holding together the equal spreading, i.e. the costs of serving remote markets, weakens. This
may therefore set in motion a cumulative causation for a small increase of consumers in region
1, leading to full agglomeration in region 1, and vice versa, for region 2 if initially moved in the
opposite direction. However, this effect on the slope is generally not as pronounced in the model
with heterogeneous intra-national space. For instance, in Panel B, while the slope is reduced
for higher values of trade liberalization, there still exists a stable equilibrium not leading to a
full core-periphery pattern as it would in Panel A. Remarkably, this long-run stable equilibrium
is brought about at an unequal distribution of the workforce within the home country. That is,
we observe a shift of the curve which cuts the constant parity line parallel to the left. This
46 The two parameter configurations represent the two most common cases seen across the simulations.
47 We thereby implicitly assume a change in the ad valorem tariff of crossing international borders down to
25% and 0%.
146
effectively indicates a stable equilibrium at an unequal distribution across the home regions.
Hence, applied to the case of the EAC, there now exists an increased draw to the border regions,
given that 𝜆𝜆1 reduces from 0.5 in autarky to 0.4 in the free trade scenario, which indicates that
now over half of the manufacturing activity is operating at the border. This is similar to the
result provided in Behrens et al. (2006) and (Crozet and Koenig 2004b), albeit in a 2+1 setup.48
If we move our view to the results in Panel C and D, this result is further corroborated.
In this scenario, the centripetal forces are accentuated as can be seen by positive values of the
slope of 𝜔𝜔1/𝜔𝜔2 throughout. Given that the only amendment is a lower elasticity of substitution
𝜀𝜀, higher product differentiation causes the strength of scale economies to increase (Fujita et
al. 2001; Paluzie 2001). The dynamics are seen in (4.40) through (4.47) in how the strength of
the centripetal and centrifugal forces depends on the parameter 𝜀𝜀. For one, in the price indices,
a lower elasticity of substitution (𝜀𝜀) increases the strength of the love of variety, such that for
any increase in the low-cost access of goods (high 𝜆𝜆, as well as a low 𝑤𝑤 and 𝑇𝑇), the price index
is lower than for higher values of 𝜀𝜀.49 Intuitively speaking, the higher the differentiation between
varieties, the higher the added utility gain of (increased availability of) a further variety to
consumers. Hence, lowering 𝜀𝜀 causes the forward (cost) linkage to intensify. Note however, as
established above, that this also automatically also leads to stiffer product market competition
among varieties, as 𝐼𝐼 is reduced.50 Secondly, in the wage equation, this means that lower values
of 𝜀𝜀 have a negative effect on the wage firms are able to afford, which displays a centrifugal
force. However, counteracting this, remember that any increase in market access (high 𝑌𝑌 and
low 𝑇𝑇) also increases the wages firms are able to afford, and lowering 𝜀𝜀 causes the backward
(demand) linkage to intensify.
As seen by the comparison between top and bottom panels, the forward and backward
linkages are strengthened, i.e. centripetal forces dominate the centrifugal forces caused by a
decrease in 𝜀𝜀.51 As such, a core-periphery pattern is more likely at any level of intra- or
international trade costs. This is seen by positive slopes in both Panel C and D. Again, while
decreases in 𝑇𝑇𝐹𝐹 cause only minor changes in the slope for Panel C, it significantly alters the
equilibrium configuration in the model with heterogeneous intra-national space. However,
different from the case in B, where the cut point with the parity line has shifted to the left, it
has now shifted to the right. This is reconciled with the previous results showing a relative shift
to the border by interpreting this change as decreasing the basin of attraction, which would
48 Note that setting 𝜆𝜆3 = 𝜆𝜆4 = 0.5 is not equal to the case of a foreign country with one region when the two
foreign regions are accessed with different transport costs.
49 To validate this, note that the negative exponent of the entire bracket in (4.40) through (4.43) gets larger,
while the negative exponents of 𝑤𝑤 and 𝑇𝑇 get smaller. Further, from (4.8) and noting that varieties are produced
with the same technology in all locations, which renders the price index as 𝐼𝐼= 𝑝𝑝· 𝑛𝑛1/1−𝜀𝜀. It is easily seen that
𝐼𝐼 is more strongly decreasing in 𝑛𝑛 (varieties) for lower values of 𝜀𝜀 (Fujita et al. 2001).
50 This can also be confirmed in (4.7), i.e. 𝑐𝑐𝑗𝑗= 𝑝𝑝𝑗𝑗
−𝜀𝜀· 𝐼𝐼𝜀𝜀−1 · 𝛿𝛿𝛿𝛿. As established in footnote 49, increasing 𝑛𝑛
decreases 𝐼𝐼 which thereby lowers the demand for any variety. This is also seen in (4.21) and (4.23), whereby
a decrease in 𝜀𝜀 leads to an increase in varieties 𝑛𝑛 at each location.
51 Table C.1 shows that this is the case for all tested parameter configurations.
147
lead to full agglomeration in region 1. In other words, in autarky, it suffices to increase the
manufacturing workforce in region 1 an infinitesimal amount above 50%, for dynamics to unfold
which lead to full agglomeration in 1. In the free trade scenario close to 73% of the
manufacturing workforce would have to be in region 1 for this cumulative causation mechanism
to kick in. Again, Crozet and Koenig (2004) as well as Brülhart et al. (2004) show qualitatively
similar results, albeit for a 2+1 setting.52 This is the second noteworthy departure from the
model with symmetric intra-national space and highlights that relevant different conclusions
arise when the access to foreign regions is unequal.
Hence, the results until now paint two consistent insights for a process of regional market
integration in a spatial layout such as the one present in the EAC. For one, liberalizing trade
across the two countries increases internal agglomeration tendencies. And secondly, that this
agglomeration is more likely to occur in the region bordering the newly accessed markets. For
additional insights into (the moderating forces of) these dynamics, see Section C.1 of the
Appendix. Figure 4.4 reproduces part of the analytics carried out in C.1. As such, it plots the
results of the “sustain” analysis, which essentially evaluates the stability of an agglomerated
equilibrium, i.e. depicts the range of intra-national transport costs 𝑇𝑇𝐷𝐷 for which the
agglomerated equilibrium in the core economic hubs, e.g. region 1, proves sustainable.
Remember that the stability condition of this equilibrium at 𝜆𝜆1 = 1 requires 𝜔𝜔1/𝜔𝜔2 ≥1. Given
that we have seen an increased draw to the border (region 2), we are interested in which range
of values a sustainable agglomeration in the interior (region 1) can be upheld. Again, Panels A
and C provide the case for the symmetric case, while Panels B and D conduct the analysis for
our main model again for the three levels of international transport costs 𝑇𝑇𝐹𝐹. The point where
the line crosses the real-wage differential from below is called the “sustain” point, 𝑇𝑇(𝑆𝑆) and
describes the maximum level of intra-national transport costs for which agglomeration is still
sustainable, i.e. for which 𝜔𝜔2/𝜔𝜔1 ≤1 (note the reversal). This equates to a real wage differential
at 𝜆𝜆1 in Figure 4.3 which stays above the parity line. Beyond this point, agglomeration is not
sustainable anymore, i.e. a case where the line is below parity in Figure 4.3. What happens
when trade is liberalized? In Panel A, the sustain point shifts to the right, which indicates that
agglomeration is able to be upheld for a wider range of domestic transport costs 𝑇𝑇𝐷𝐷. By design,
this result mirrors the one in Monfort and Nicolini (2000) and also what we have seen in Figure
4.3 Panel A, i.e. that decreased cross-border trade costs increase the agglomeration forces. Note
that this is mainly due to a decrease in the slope of the ascending part of the plotted lines. At
these levels, an infinitesimally small increase in 𝑇𝑇𝐷𝐷 increases 𝜔𝜔2/𝜔𝜔1 and liberalization thereby
seems to influence the centrifugal forces (competition) to a larger degree than the centripetal
forces (cost and demand linkages), as seen by the positive slope (Crozet and Koenig 2004a).
52 Brülhart et al. (2004) additionally departs from CES and uses a quasilinear consumer utility in the vein of
(Ottaviano et al. 2002).
148
Panel A
Panel B
Panel C
Panel D
Figure 4.4: Internal Transport Costs and Sustainable Agglomeration
As such, a decrease in international trade costs mainly modulates the strength of the
dispersion forces. In Panel C, these centripetal strengths dominate by reasons given above, such
that the core-periphery pattern is upheld for a larger range of intra-national transport costs.
So much so that in full trade scenario, there exists no sustain point and agglomeration in region
1 is never broken. Again, the results for our main model with heterogeneous intra-national
space provide different conclusions. While the range of transport costs for which a core-
periphery pattern is upheld also decreases in 𝜀𝜀 (compare Panels B and D), trade liberalization
works towards the opposite, i.e. puts negative pressure on the full agglomeration in region 1,
as seen by the negative shift of the sustain point 𝑇𝑇(𝑆𝑆) to the left. As such, regional market
integration decreases the range of values for which agglomeration away from the border region
can be upheld. This time, the change in the slope occurs mainly for the descending part,
149
indicating that trade liberalization decreases centripetal forces for region 1.53 These results
essentially confirm analytically what is depicted numerically Figure 4.3, i.e. that there is an
increased draw to the border with increased trade liberalization. However, one interesting
aspect is that, for any level of regional trade integration 𝑇𝑇𝐹𝐹, agglomeration in region 1 is more
likely to be upheld in the case of high product differentiation. Next to stronger scale effects,
one possible way to interpret this result may be to refer to the heightened competitive pressures
of an increased number of firms from abroad, keeping domestic firms in the interior, where they
are sheltered (Crozet and Koenig 2004b).54 See the discussion of Section C.1.3 of the Appendix.
So far, we have confirmed the results of previous symmetric 2 + 2 settings (Monfort and
Nicolini 2000) and extended those from asymmetric 2 + 1 layouts to an economy with four
regions. What is left to investigate in our unique 2 + 2 setting is the influence of foreign
economic inequality on these initial results, given that we have so far set 𝜆𝜆3 = 𝜆𝜆4 = 0.5. We
now relax this assumption and discuss additional results for the full range of spatial
configurations λ ϵ [0,1] in the foreign country. Given this added dimension, Figure 4.3 turns
three-dimensional which makes it a bit cumbersome to evaluate at first sight (see Figure C.3
and Figure C.4 of the Appendix). To make it more accessible, for the moment, we restrain
ourselves to assessing the influence of a varying foreign manufacturing distribution on our two
types of long-run equilibria depicted in Figure 4.3. As such, Figure 4.5 plots the combinations
of 𝜆𝜆1 and 𝜆𝜆4 where the real wage differential 𝜔𝜔1/𝜔𝜔2 is equalized, i.e. give the contour lines of
the plane spanned by the two endogenous variables as given in Figure C.3.55 The arrows in the
graphs indicate the stability of the equilibrium, not depictable in contour lines. Arrows pointing
towards the line indicate a stable equilibrium, given that economic forces (the real wage
differential) lead consumers back to the original allocation, and vice versa for arrows pointing
away from the line. As is now common, Panel A and C shows the result for the homogenous
2 + 2 model whereas Panel B and D depicts our asymmetric case. Given that the foreign spatial
configuration can only exert influence when trade costs are not prohibitive, we restrict ourselves
to analyze the case for which 𝑇𝑇𝐹𝐹= 1.60.56 To no surprise, Panel A and C shows a vertical line
at 𝜆𝜆1 = 0.5. This is because when the home country is equally spread, and both regions have
equal access to the foreign market via 𝑇𝑇𝐹𝐹, there is no difference in the relative real wages of
the two regions. Hence, shifting shares of the workforce in the foreign regions doesn’t affect the
existence of such an equilibrium.57 Panels B and D paint a wholly different picture, and show
53 Note importantly, that given the unequal intra-national space, a reversal of the analysis, i.e. evaluating the
stability of an agglomerated equilibrium at the border with 𝜆𝜆1 = 0 and 𝜔𝜔1/𝜔𝜔2 ≤1 would render the shift of
the lines as in Panels A and C, i.e. would move the sustain point to the right.
54 This is seen in (4.23) as an increase in product differentiation increases the number of varieties at a location,
solely by an incerase in the number of producers (at a location).
55 Remember that 𝜆𝜆2 = (1 −𝜆𝜆1) and 𝜆𝜆3 = (1 −𝜆𝜆4).
56 The results for all three levels of international transport costs are provided in Figure C.5 and Figure C.6.
57 However, what may in fact happen, as shown in Monfort and Nicolini (2000), for a specific parameter
configuration, is that the stability of this spreading equilibrium depends on the foreign distribution
(interdependent equilibrium).
150
Panel A
Panel B
Panel C
Panel D
Figure 4.5: Foreign economic Inequality and Spatial Equilibria
that in the presence of heterogeneous intra-national space, foreign spatial inequality has a
moderating impact on the domestic (agglomeration vs. dispersion) forces induced by trade
liberalization. This result was first shown by (Behrens et al. 2006), albeit for a setup in which
a “gate” exists in only one of the countries and for a different setup of consumer preference
structures (see footnote 34).
For the case where a long-run stable equilibrium exists (𝜀𝜀= 6), we see that a changing
share of workers in the interior of the foreign country modulates the domestic allocation for
which this equilibrium is reached. Panel D corroborates this view for the case where trade
liberalization has led to a core-periphery pattern as the only stable equilibrium. In general, the
higher the foreign spatial distribution is skewed towards the interior (a higher 𝜆𝜆4), the lower
the share of workers in region 1 needed for both types of equilibria depicted in Panel B and D
in Figure 4.3. What is the implication of this result? In the case of less intensive scale effects
(𝜀𝜀= 6), the draw to the border is further increased (Panel B). On the other hand, when scale
effects are large and full agglomeration is the only stable equilibrium (𝜀𝜀= 4), this result is
151
reversed (Panel D): while there is still an increased draw to the domestic border in free trade,
as can be seen by values for 𝜆𝜆1 above 0.5 on the x-axis, this draw is decreasing in 𝜆𝜆4. Hence,
given a stark regional inequality in the foreign country which is outlined by an economically
strong interior (high 𝜆𝜆4), the basin of attraction needed for domestic agglomeration at the
border is decreased when product differentiation is high. This is intuitively plausible, as it may
signal a decreased relevance of sheltering from foreign competition further away from the
border, the effect of which is particularly pronounced for environments with a larger number
of competing firms (low 𝜀𝜀). In the contrary scenario (high 𝜀𝜀), market access considerations may
prevail over this competition effect, leading to increased agglomeration at the border. Note that
a similar point is made (Behrens et al. 2006), who show that for the case in which one of the
countries is outlined by heterogeneous intra-national space, the general agglomeration tendency
in the domestic country is increased when the foreign country hosts a lower number of firms at
the border. See Section C.1 for a more thorough discussion on these possible mechanisms.
In sum, the NEG model developed in this section and the counterfactual exercise of an
increased market integration performed through it holds three main insights. First, given
heterogeneous intra-national space, progressive trade liberalization draws economic activity to
the border, i.e. the real wages at border regions are relatively higher than in the interior when
compared to autarky. Second, general agglomeration tendencies increase the freer trade is,
although the agglomeration is more likely to occur at the border. And third, foreign spatial
inequality has non-negligible impacts on these domestic effects, such that a core-periphery
pattern in the foreign economy may attenuate or reinforce the first and second results. This
result holds relevant insights for the case of the East African Community, where the countries
integrating are in fact all outlined by large interior hubs positioned away from internal borders.
With these deliberations in mind, the next section tests for reduced form evidence on
these potential spatial effects of regional market integration in the EAC.
4.5 Empirical Strategy
The theoretical exercise of the previous section motivates the empirical strategy. As is seen
from the simulations, lowering trade costs among EAC members is predicted to increase the
draw to the border, i.e. to increase relative welfare in the regions with better access to the new
markets. As discussed in Section 4.2.2, this result is corroborated both by previous theoretical
models as well as by empirical evidence from both developed and developing settings.58
However, what we also saw in the simulations is that this draw may be attenuated or reinforced
depending on the economic (in)equality present in the foreign economy. Given the particular
spatial layout of the EAC member countries anticipated in Section 4.3 and Subsection 4.4.2,
58 Note also that a pure reference to new trade theory is not strictly necessary to render an increased impact
of trade closer to borders. It has been shown in other developing settings that price pass through is highest
directly at the border and decays perpendicular to it (e.g. Nicita 2009; Cali 2014; Atkin and Donaldson 2015).
152
trade liberalization among them presents a fitting empirical case on which to study these
dynamics, i.e. testing whether the re-establishment of the EAC did increase the relative welfare
of border regions in comparison to the pre-existing economic hubs, and also provides a setting
with which to identify household welfare effects of trade integration.
Note that so far, we have measured the attractiveness of border regions with real wages,
and particularly, real wage differentials. However, we can readily translate these real wages into
household welfare, as first established using indirect utility in (4.12), by simply noting that
food was set as the numeraire.59 As such, the real wages discussed broadly in Section 4.4
encapsulate what we envision as household welfare in the simplest form, which is, the income
consumers earn and the prices they face (see e.g. Deaton 1997; Fujita et al. 2001; Winters 2002;
Brülhart et al. 2012). The comparative statics tested theoretically thereby translate to the
empirics in a reduced-form manner and revolve around assessing what happens to households’
welfare (indirect utility) across space following a change in the trade costs among EAC members
from a former prohibitive level down to levels of trade costs that mirror those of the type within
the domestic country, i.e. only given by the geographic distance between locations.60 To
operationalize this, I employ a difference-in-differences (DiD) specification comparing the
changes in welfare of households living relatively closer to borders, 𝜔𝜔2, with those of households
living relatively closer to the economic hubs 𝜔𝜔1, before and after the re-establishment of the
EAC in 2001.61 To flexibly allow for treatment across space a the potential draw to the border,
I model this relationship nonparametrically, i.e. employ a continuous treatment intensity which
is captured by households’ (road) distances to nearest EAC border crossings. The estimating
equation therefore reads:
𝑌𝑌𝑖𝑖𝑖𝑖𝑖𝑖= 𝛼𝛼+ 𝛽𝛽1𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖+ 𝛽𝛽2𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖+ ∑
𝛽𝛽3,𝑡𝑡(𝛾𝛾𝑡𝑡× 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖) +
{𝐶𝐶𝐶𝐶}
𝑡𝑡={𝐸𝐸𝐸𝐸𝐸𝐸}
f ∑
𝛽𝛽4,𝑡𝑡(𝛾𝛾𝑡𝑡× 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖)
{𝐶𝐶𝐶𝐶}
𝑡𝑡={𝐸𝐸𝐸𝐸𝐸𝐸}
+ 𝑋𝑋′
𝑖𝑖+ 𝛿𝛿𝑐𝑐𝑐𝑐/𝑖𝑖/ℎ+ 𝑒𝑒𝑖𝑖𝑖𝑖𝑖𝑖
(4.53)n
𝑌𝑌𝑖𝑖,𝑐𝑐,𝑡𝑡 represents the respective welfare indicator of individual 𝑖𝑖 living in country 𝑐𝑐, surveyed at
survey-sampling period 𝑡𝑡. 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝑖𝑖 is the inverse, relative within-country distance to the nearest
EAC border crossing [0,1], such that a value of 1 indicates individuals in the sample living
closest to the border in the sample, and value of 0 those furthest away. 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖 is a dummy {0,1}
indicating individuals living within 50 kilometers of the three preeminent interior
agglomerations, namely Nairobi in Kenya, Dar es Salaam in Tanzania and Kampala in
59 For a given set of exogenous parameter values, the indirect utility function (4.12) then reduces to a function
of income, which is varying across space only in the nominal manufacturing wages 𝑤𝑤 (see (4.36) through
(4.39), and consumer prices for manufactures 𝐼𝐼.
60 Note that transport costs between regions within countries are assumed constant throughout the same time
period.
61 The empirical specifiction is similar to Eberhard-Ruiz and Moradi (2019), however with distinct regard for
urban vs. rural differences.
153
Uganda.62 𝛾𝛾𝑡𝑡 is an indicator for the respective integration period i.e. switching to 1 for the free
trade period (EAC) between 2001 and 2004, the customs union period (CU) between 2005 and
2009, as well as the common market period (CM) after 2010, respectively.63 Therefore, under
the assumptions that no other concomitant policy or shock has induced a differential trend in
outcomes between these regions during the same time period, 𝛽𝛽3 and 𝛽𝛽4 represent estimates of
the effect of the EAC on border regions as well as on interior agglomerations. Specifically, they
give estimates of the differential effect of households living at the border compared to those
living furthest away, and those living in interior agglomerations to those living in the auxiliary
(rest of the country), both compared before and after the EAC was established. 𝛽𝛽3 and 𝛽𝛽4 can
thus be seen as a test on the theoretical predictions, i.e. whether the EAC led to larger relative
increases in welfare in border regions, i.e. ∆𝜔𝜔1/∆𝜔𝜔2 < 1, given by a 𝛽𝛽3 ≠0 which also satisfies
𝛽𝛽3 > 𝛽𝛽4, rather than the opposite, i.e. in preexisting interior agglomerations, ∆𝜔𝜔1/∆𝜔𝜔2 > 1,
given by 𝛽𝛽4 ≠0 for which 𝛽𝛽3 < 𝛽𝛽4.64 These estimates therefore also indicate if we should expect
dispersion of the previously concentrated economic activity rather than concentration as
proposed by the endogenous adjustment process in (4.52). 𝑿𝑿 represents a matrix of individual-
level control variables which allows us to account for all influences potentially conflating the
relationship between access to (new) markets and household welfare. 𝛿𝛿 captures country-time
fixed effects such that identification in the cross-sectional datasets comes from variation within
individual member countries in specific survey-periods in time. For the results produced from
the household panel, identification comes from changes within household/individuals over time,
such that 𝛿𝛿 represents household-, respectively, individual fixed effects. Standard errors are
constructed by allowing for spatial correlation of errors, i.e. Conley standard errors are used
(Conley 1999, 2010). I additionally check for the clustering of errors at the level of the survey
enumeration area, i.e. at the survey cluster level.65 Binary dependent variables are estimated
with a simple Linear Probability Model (LPM) specification.66
4.6 Data
I employ a distinct set of longitudinal, geo-referenced household-level surveys that were sampled
in all three founding members before and after the establishment of the EAC. First, I make
62 I also test the lower distance thresholds 25km and 10km in the robustness tests of Subsection 4.7.2.
63 In the sample, the first post-EAC responses we measure come from the year 2002, which is indicated in the
regression results.
64 Of course, this statement is true only for the maximum effect (change) of border distance, i.e. going from
the largest distance to the smallest distance in the sample. In-sample prediction, such as an interquartile range
bound between 0 and 1 is arguably more appropriate as a comparison, where 𝛽𝛽3 must be more than
infinitesimally larger than 𝛽𝛽4.
65 The cut-off for Conley standard errors is chosen by the function of the “fixest” package in R, which ensures
a large enough sample size within a certain distance cutoff and is additionally robust to sub-sampling. The
results are robust to sensitivity checks of the Conley standard errors by increasing and reducing the distance
cutoff. Results can be obtained from the author.
66 Results for binary dependent variables estimated via Probit yields qualitatively identical and quantitatively
similar marginal effects.
154
use of the complete set of available Demographic and Health Surveys (DHS). DHS are cross-
sectional, household-based surveys which are representative at both the national- as well as
regional level, and collect a broad array of information on topics such as demographics,
education, employment and occupation, as well as fertility and family planning (Croft et al.
2018).67 The main respondents are women of reproductive age (15-49), but DHS also provides
information on men and children living in the sampled households, as well as on household-
specifics such as consumer durables and wealth assets in possession. To increase the variable
size and sample space, I pool these additional data from the Men- and Household recodes also,
leading to a main sample of 227,860 individuals living in 140,943 households across 7,962 survey
locales interviewed between 1999 and 2020. Later extensions and robustness checks expand this
sample to include non-GPS survey rounds sampled from 1988 onwards, leading to an extended
sample size of 332,725 individuals living in 203,150 households.68 Note that the main sample
also includes special survey rounds such as the AIDS Indicator Survey (AIS), the Malaria
Indicator Survey (MIS), as well as the Knowledge, Attitudes and Practices Survey (KAP), to
gather a higher frequency of survey years. While all variables used across the paper are
consistently available in these survey years, too, I provide an additional robustness that drops
these survey waves in Subsection 4.7.2.69
Second, I make use the geo-referenced Afrobarometer (AFB) survey rounds, which span
a timeframe of 18 years (from 1999 to 2017) across seven survey waves, i.e. rounds 1 through
7 (Afrobarometer 2019).70 Afrobarometer surveys are representative at the national level, and
the main respondents are adults of the sampled households. They carry individual- and
household-level information on basic characteristics, socio-demographics as well as own
(economic) living conditions, household assets, and additionally, provide information on
individuals’ sentiments as well as opinions towards the economy, democracy, governance and
society. Afrobarometer fits geo-coordinates (latitude and longitude) to respondents at the level
of their respective enumeration area, and the sampling procedure aims for eight
individuals/households per EA (BenYishay et al. 2017). The Afrobarometer adds information
on 38,644 individuals (households) living in 3,414 geo-referenced localities across Kenya,
Uganda and Tanzania to the sample, and additionally, provides the opportunity to test specific
sentiments and attitudes towards free trade, which I turn to in the section on robustness and
extensions (4.7.2).
Lastly, I supplant the analysis with information from the Kagera Health and Development
67 More precisely, the first level administrative subdivision, most often referred to as regions, districts, provinces
or states.
68 The surveys were sampled in Kenya in 1989, 1993, 1998, 2003, 2008-09, 2014, 2015 (MIS), and 2020 (MIS),
in Tanzania in 1991-92, 1994 (KAP) 1995, 1996, 1999, 2003-04 (AIS), 2004-05, 2007-08 (AIS), 2010, 2011-12
(AIS), 2015-16, and 2017, and in Uganda in 1988-89, 1995-96, 1995, 2000-01, 2006, 2009 (MIS), 2011, 2011
(AIS), 2014-15 (MIS) 2016 and 2018-19 (MIS).
69 With the exception of Employed Work {0,1}.
70 Surveys were sampled in 2000-2001 (only Tanzania and Uganda), 2002-03, 2005, 2008, 2011-12, 2014-15 and
2017.
155
Survey (KHDS) (World Bank and University of Dar es Salaam 1994, 2004, 2010). The KHDS
is a representative panel originally sampled from Kagera, a GADM-1 administrative region of
Tanzania bordering Uganda in the northwest. The panel collected detailed information on
households’ and individuals’ wealth and poverty dynamics, such as employment, salary, (non-)
durable assets as well as food- and non-food consumption, all for which values in constant
(deflated) Tanzanian Shilling are provided (Beegle et al. 2006; De Weerdt et al. 2010). The
KHDS also includes information on migration decisions of individuals as well as community-
level variables such as the price of commodities in local markets. The KHDS set out by
interviewing 6,356 individuals living in 915 households spread across 51 sampling clusters in
the first four yearly survey waves between 1991 and 1994 (round 1).71 All of the initially sampled
households (rather, the individuals living within those households) were sought to be re-
contacted in the succeeding two survey rounds in 2004 and 2010 (rounds 2 and 3), respectively.
The tracking of individuals was highly successful, the sample evincing re-contact rates of 80%
for individuals and over 90% for singular households.72 Importantly, the number of administered
households (and individuals) grew significantly over the sample timespan, as all members
residing in (new) households of original respondents were fully included in the sample, and also
tracked in later survey years.73 As such, the KHDS is able to add information on 21,696 distinct
individuals – interviewed a minimum of one-, and a maximum of six times – whose households
are spread across 2,019 survey locales in Tanzania and Uganda. Importantly, out of the 6,356
original survey respondents sampled between 1991-1994, 4,430 individuals were successfully
(re)-interviewed in 2004, and 3,848 were surveyed in all three rounds, including 2010.
Figure 4.6 visually depicts this distinct set of geo-referenced data by plotting the sample
enumeration areas of households from each of these three sources across East Africa. Notice
that the map also depicts enumeration areas of contiguous EAC-accession as well as non-
accession countries. These data will be employed in the extensions and robustness tests of
Subsection 4.7.2.
4.6.1. Dependent Variables
As anticipated above, the real wages analyzed in theory are proxied by household welfare i.e.
the income consumers earn and the prices they face (Deaton 1997; Winters 2002). Given the
usual data restrictions of household surveys, i.e. a lack of precise wage and price data, I proxy
welfare by a set of intensive and extensive labor market outcomes (work, employment, and
income) as well as consumption measures (food and non-food consumption, durable as well as
non-durable assets).
71 Not all of the households were interviewed in all of the first four waves.
72 And over 90% for those cases where at least one of the original household members was aimed to be re-
interviewed.
73 As such, the number of singular households contained in the survey expanded from 915 in the first round
(1991 to 1994) to 2,719 in 2004 and 3,314 in 2010 (De Weerdt et al. 2010).
156
Figure 4.6: Sample Coverage
To measure the levels of consumption in the Demographic and Health Surveys, I make
use of the Wealth Index {1,5}, which is a DHS-constructed index placing households on a
relative scale of wealth within their respective sample. The construction of the index is based
on an array of consumer durables, such as the construction of dwelling, sanitation facilities as
well as possessions such as a TV, motor vehicles etc. (Rutstein and Johnson 2004). I
countercheck these results with the Comparative Wealth Index established by the DHS, which
facilitates the comparison of the wealth scores underlying the wealth indices across countries
and samples (Rutstein and Staveteig 2014). I additionally construct the International Wealth
Index (IWI) as established in Smits and Steendijk (2015) as a further attempt to make
157
household’s wealth ranking more comparable across surveys.74 Concerning labor market
outcomes as the second dimension, I test the variable Employed Work {0,1}, which indicates
whether the respondent worked for someone outside of the household (conditionally on having
work). In later extensions for these labor market results, I also test Worked in last Year {0,1},
which is the baseline measurement indicating whether survey respondents were pursuing some
activity aside from housework within the last calendar year, on which 2) is conditioned on, and
test whether this activity was also remunerated, i.e. Paid in Cash {0,1}. Lastly, I test the
Occupational Type {1,3} of work, which places all activities categorized within the survey
schedule from “agrarian” (1), “worker” (2) and “professional” (3) activities.75 This may be
regarded as a test on the skill-intensity of the respondent’s occupation.76
Concerning the Afrobarometer, the level of (basic) household consumption is measured
by Lived Poverty [0,4], which is constructed by averaging individuals’ responses to the three
separate questions: “Over the past year, how often, if ever, have you or anyone in your family
gone without: Enough clean water for home use” / “[…]: Enough food to eat” / “[…]: Medicines
or medical treatment?”.77 The response values range from “never” (0), “just once or twice” (1),
“several times” (2), “many times” (3) and “always” (4). Similar to the DHS, the primary test
on individuals’ labor market outcome is measured via Employed Work {0,1}. Again, I test
whether individuals Worked Last Year {0,1} as well as their Occupation Type {1,3} to explore
an extended set of labor market outcomes. Relatedly, I employ Cash Income Deprivation {0,4}
which is asked in the same way as the components of the lived poverty index and indicates how
often individuals within a household have gone without a cash income within the past year.
One of the key characteristics of the Afrobarometer surveys is the component containing
opinions, attitudes, and sentiments on individual, political, as well as domestic and
international economic topics. As such, I am able to test the variable Support for: Regional
Integration {1,5} which evaluates the strength of the support for free movement and trade.78
Further, I test whether individuals living closer or further from the border assess the Ease of
Crossing Borders {1,4} as more or less difficult, how much they evaluate the EAC as well as
the African Union (AU), as the supranational trade facilitator of the continent, in helping their
74 The IWI exploits information from the entire universe of developing countries household surveys to construct
factor loadings of specific household wealth items.
75 For instance, “workers” are occupations such as traders, artisans, or unskilled manual labor. “Professional”
is comprised of lawyers, accountants and teachers.
76 A simple regression of Occupation Level {1,3} on the individual characteristics age, age squared, years of
education as well as a female dummy and country-time fixed effects shows that each year of education increases
the index by 0.06 units and a standard error of 0.0045.
77 Contrary to the Afrobarometer’s use, I don’t include the question on Cash Income droughts, as I test it as
a separate outcome for the labor market outcomes. Further, I am unable to include the response to having
enough fuel to cook food, as responses to this variable are not available in pre-EAC survey rounds.
78 The question asked in Round 6 of the Afrobarometer probes the support by agreeing with either of the two
following mutually exclusive statements. Statement 1: “People living in the sub-region should be able to move
freely across international borders to trade or work”. Statement 2: “because foreign workers take away jobs,
and foreign traders sell their goods at very low prices, governments should protect their own citizens and limit
the cross-border movement of people and goods” (Afrobarometer 2019).
158
country {0,3}, and evaluate whether they would like having an Immigrant as a Neighbor {1,5}.
I supplant these variables with further subjective assessments of the Present vs. Past: Life
Satisfaction {1,4}, as well as their Present vs. Past: Living Standards of People {1,4}.
The Kagera Health and Development Survey (KHDS) allows us more detailed access into
the consumption and income dimensions of households and the individuals therein. First, I test
changes in the Annual per capita Household Consumption which expresses aggregate food and
non-food consumption in constant, i.e. deflated, 2010 Tanzanian Shilling (‘000 TZS).79 Food
items are constituted of both purchased as well as home-produced food, non-food items are
comprised of expenses on items such as clothing, schooling, services like haircuts, or utilities.80
I also test for differences in Food- and Non-Food Consumption separately, to identify potential
systematic differences across the two and to countercheck the results on lived poverty measured
in the Afrobarometer more precisely. Secondly, I test for changes in household wealth similar
to the DHS using the Value of Durable Assets as well as the Value of the occupied Dwelling of
a survey household.81 Both measures are given in deflated, 2004 Tanzanian Shilling (‘000 TZS)
as this component was last administered in the survey wave of 2004. Concerning income, I
proceed in similar to the previous two surveys and ask whether the respondent has Employed
Work {0,1}, whether they have Salaried Work {0,1} and the Occupation Type {1,3}. In
extensions, I also test the overall likelihood of having any kind of work via Worked last Year
{0,1}, as well as employed individuals’ Monthly Salary in deflated 2004 Tanzanian Shilling (‘000
TZS). Similarly to the Afrobarometer, the KHDS contains an array of subjective assessments
and evaluations. As such, I test for the main motives of migrating to the current place of
residence, and whether movers found Paid (formal) Employment {0,1} right after migrating to
the current location. Again, these measures are supplanted by assessments of their Subjective
HH. Wealth {1,5} both at the survey time (2004) as well as 10 years before (1994) and the
general Life Satisfaction on a Ladder {1,9}.
As a last empirical investigation into the spatial corollaries anticipated in Section 4.4, I
test for changes in the extent of agglomeration at respondents’ geographic locations
(enumeration areas) across all three surveys. To do this, I merge granular population data
provided by the Gridded Population of the World (CIESIN 2017) to evaluate the EAC’s effect
on Population Density (sdz.), a measure of the total number of persons per square kilometer
at the specific geography, to all unique enumeration areas of the respective samples. 82 Assessing
changes in population density across space before and after the EAC is a direct test of the
long-run dynamics induced by the agglomeration vs. dispersion forces discussed in Section 4.4.
In another vein, population density may be useful as a more general indicator suggested to
79 The total annual household consumption is distributed equally across all household members.
80
For
the
full
information
on
the
construction
of
the
aggregate
consumption,
see
https://microdata.worldbank.org/index.php/catalog/2251/download/ 34035.
81 Durable goods include e.g. radios, refrigerators, telephones. The figure for the value of the occupied dwelling
represents an estimate of the head of the household.
82 To facilitate interpretation, I standardize the value at a mean of 0 and a standard deviation of one.
159
capture “underlying differences in productivity and quality of life” (Breinlich et al. 2014; 733),
which displays a relevant outcome for policy interventions such as market integration.83
4.6.2. Independent Variables
The main explanatory variable of interest 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 [0,1] is measured by calculating the shortest
road distance from each household’s enumeration area to the nearest (within country) internal
EAC border crossing (depicted in Figure 4.6).84 To circumvent endogeneity in the construction
of roads, I only use major roads, i.e. motorways, trunk- and primary roads as provided by
OpenStreetMap (OSM 2022), which can be tracked back to the pre-EAC era. Border crossings
are defined as points where these major roads connect to both sides of the border. To assess
the sensitivity of the results to the specific distance calculation, I also construct beeline (as the
crow flies) distances from all enumeration areas to both the border crossings as well as to the
nearest possible point on the borderlines spanned by two EAC country pairs. Shapefile data
for country administrative areas, i.e. the boundaries of which I use come from the Center for
Spatial Sciences at the University of California (GADM 2020). To test the effects on the pre-
existing economic hubs, I employ 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 {0,1}, which is a dummy indicator switching to one for
households located within 50km of the country’s respective economic hub (i.e. Dar es Salaam,
Nairobi, and Kampala).85 To also provide a test on the general tendency of agglomeration
derived in Section 4.4, irrespective of border distance, I employ Agglomeration {0,1}, which
switches to 1 for households living within 50km of an “urban center” demarcated as such in
the year 2000. The list of urban centers I use is provided by the European Commission (2019).
Lastly, I also make use of the Urban {0,1} dummy, which is a survey-specific classification of
the level of urbanization at the specific location.86
I also include an array of controls to account for influences which potentially conflate the
relationship between household welfare and spatial aspects over time. As such, I include the
individual-level covariates Age, Age squared, a dichotomous indicator of gender, Female {0,1},
as well as individuals’ Educational Attainment. I additionally account for potentially correlated
geographic influences of development across distance and closely follow Henderson et al. (2018)
with a set of important physical geographic features. I therefore include the location’s Elevation
(Farr et al. 2007), Ruggedness (Nunn and Puga 2012) as well as agricultural characteristics
such as the number of Growing Days (Ramankutty et al. 2002) as well as average long-term
83 See for instance Rappaport and Sachs (2003).
84 I measure distances using the projection of coordinates along the earth’s ellipsoid (using WGS 84, EPSG
7030).
85 Later robustness tests relax this distance cutoff and also test areas within 25km and 10km, respectively.
Notice that for the KHDS survey, there is no data for households living in these hubs in the pre-EAC era, as
the survey was initially sampled in Kagera only. For these earlier cases, the dummy switches to 1 for individuals
living in “Bukoba” the urban capital of Kagera.
86 Urban stems either from country census information (DHS) or on the assessment of sample enumerators
(Afrobarometer). See e.g. https://www.idhsdata.org/idhs-action/variables/URBAN#comparability_section
and
https://www.afrobarometer.org/wp-content/uploads/2022/07/AB_R9.-Survey-Manual_eng_FINAL
_20jul22.pdf.
160
Monthly Temperature and Monthly Rainfall (Fick and Hijmans 2017). In later robustness
checks, I additionally control for the location’s Malaria Ecology (Sachs et al. 2004), Absolute
Latitude, as well as household’s distance to Navigable Rivers, as well as Major Lakes and Major
Harbors.87 Lastly, I add country-year fixed effects to control for time-specific influences as well
as country-specific influences at specific points in time, such as the Kenyan Post-Election Crisis
of 2007-2008, and additionally include household-, respectively, individual fixed effects for
estimations using the Kagera Health and Development Survey.
Table C.2 of the Appendix provides summary statistics for the dependent- and
independent variables categorized by sample source. The table also provides first insights into
the distribution of outcome variables across space by grouping values into (border) distance
quartiles and reporting a separate mean of outcomes within the three core agglomerations (see
panel b). As expected from the facts outlined in Section 4.3 the countries are highly polarized.
For instance, compared to the overall mean in the sample, individuals living in the core have
between 0.8 and 2 additional years of schooling, which may provide a reason for the higher
occupation type which is also shifted 0.5 units in favor of in the core. Also, individuals living
in the economic hubs have a 20 percentage points higher likelihood of employed work, and a
24% decrease in lived poverty.
4.7 Results
4.7.1. Main Results
Table 4.1 presents the first set of results estimated via regression equation (4.53). Importantly,
these results set out the analysis by reporting an aggregate “pre vs. post” difference-in-
differences effect across all post-integration periods estimated by setting 𝛾𝛾𝑡𝑡= 1 for all responses
collected in or after 2002.88 The effects differentiated across the specific post-integration periods,
i.e. the test on the temporal evolution of the EAC towards a Customs Union (CU) as well as
a Common Market (CM), are established from Table 4.2 onwards. Note lastly, that the results
reported in all tables of this section are restricted to the reporting of the two DiD estimates 𝛽𝛽3
and 𝛽𝛽4, only. This is because we are primarily concerned with the relative development of
border vs. interior agglomerations, and because the identification strategy is ultimately bound
to estimate the differential effects of regional market integration only.89
We first focus on results produced from the two nationally representative household-
87 The criteria for rivers’ “navigability” as well as the importance of lakes (“major”) is defined as in Henderson
et al. (2018), i.e. I select all natural rivers within size categories 1-5 (scale 1-7) as defined in Natural Earth
(2019) and lakes with a surface area of over 5,000 sq. kilometers (Lehner and Döll 2004). Concerning harbors,
I define all large and medium sized ports listed in the World Port Index (WPI) as “major harbors” (NGA
2019) as in (Wild and Stadelmann 2022), who provide recent individual-level evidence that access to such
harbors is a robust predictor of living standards and household welfare.
88 2002 is the earliest available post-EAC survey year for the samples employed.
89 Given that the control group is merely treated less intensely, however effectively located in the same
intervention jurisdiction, the estimation differences out any effect that is common to both regional entities.
161
surveys, the Afrobarometer and the DHS for which the table depicts estimates on three specific
outcomes, each representing a distinct dimension of and contributing to spatial inequality. To
facilitate the comparison of effects across samples, the dependent variables are consistently
grouped into consumption, income and agglomeration categories. Concerning the consumption
Table 4.1: Aggregate Difference-in-Differences Estimates
Consumpt.
Income
Agglom.
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[3.14]
[0.19]
[0.00]
Border [0,1] × EAC 1[t ≥ 2002]
0.114
0.076*
0.111
-0.479**
-0.060
0.069
(0.278)
(0.041)
(0.094)
(0.202)
(0.041)
(0.094)
0.681*
0.066
0.238
0.018
0.143***
0.464
Core {0,1} × EAC 1[t ≥ 2002]
-0.362***
-0.045
0.758**
0.363***
0.023
0.799***
(0.069)
(0.062)
(0.336)
(0.121)
(0.040)
(0.118)
0.0000
0.4746
0.0240
0.0028
0.5737
0.0000
Individual Controls
YES
YES
NO
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
183,250
71,738
7,692
R-Squared
0.13
0.14
0.25
0.34
0.16
0.29
R-Squared -Within
0.10
0.10
0.23
0.33
0.11
0.28
Dependent Variable
Afrobarometer
DHS
dimension reported in columns (1) and (4), the results paint a stark picture. Both the AFB
and the DHS provide no evidence of relative increases in household welfare in regions closer to
internal EAC borders following the establishment of the EAC. In fact, the results show that
consumption was differentially affected in the negative direction, as given by a statistically
significant, relative shift in household wealth by -0.115 (=-0.479×0.24), when comparing
households located at the median distance in the sample (236 km, which corresponds to a value
of 0.76 for Border [0,1]) with households living closest to the border. This represents a reduction
Notes: The results in each column are produced by a separate regression. In columns (1) through (3), data come
from the Kenya, Uganda and Tanzania Afrobarometer surveys rounds 1 through 7 sampled between 2000 and 2020.
In columns (4) through (6), data come from the Kenya, Uganda and Tanzania Demographic and Health surveys
(DHS) sampled between 1999 and 2020. The sample mean of the respective dependent variable is given in brackets
above the estimates. For the DHS variable measured at the household level, Wealth Index in column (4), the answer
from the main survey respondents (women) is used. For the coefficients on Population Density in columns (3) and
(6), the level of observation is given by the unique sample enumeration areas of the respective survey. EAC 1[t ≥
2002] switches to one for individuals sampled from 2002 and onwards. All regressions include individual-level
controls for respondents’ age, gender, as well as education, and also include the geographic controls average monthly
temperature, average monthly rainfall, elevation, ruggedness, and the number of growing days. The regressions also
include country-year fixed effects. Binary dependent variables are estimated through a Linear Probability Model
(LPM). The standard errors reported allow for spatial correlation, i.e. Conley standard errors are used. ***, **, *
represents significance at the 1, 5 and 10 percent level, respectively.
162
in wealth quintiles by 3.7% relative to the overall sample mean. Concerning the complementary
DiD estimate for the core agglomerations, we observe the opposite effect. Compared to the rest
of the country, households living in Nairobi, Dar es Salaam or Kampala increased their position
on the wealth quintile by 0.363 units (12%) and reduced the occurrence of lived poverty by
0.362 (32%). Concerning income, the results do not provide systematic evidence for a change
in the likelihood of having employed work, only showing a marginally significant increase (at
the 10% level) by 1.9 percentage points (8%) on the interquartile range for border households
in the Afrobarometer. Turning to the evidence on agglomeration patterns (columns 3 and 6),
we see that the population density of preexisting core agglomerations is increased starkly
following the re-establishment of the EAC by three-quarters of a standard deviation, compared
to household locations in the rest of the country.
In sum, this first set of results does not provide evidence in favor of the main prediction
of Section 4.4, i.e. that relative household welfare (real wages) is increased in border regions.
Households and individuals living closer to borders did not experience (greater) relative welfare
gains after the EAC was established, compared to individuals living further away. This result
is consistent throughout the outcome variables tested in both the Afrobarometer as well as the
DHS samples. Inasmuch as trade liberalization weakens agglomeration tendencies in the
primate cities, which was the second prediction of Section 4.4, the findings strongly negate this
notion, as we can observe large differential increases in household consumption and population
density in the core agglomerations predating the EAC.
I now turn to a more nuanced assessment of these aggregate effects and estimate the full
set of period-specific difference-in-differences estimates anticipated in equation (4.53). Table
4.2 reports these results, which effectively expand the simple DiD effect of Table 4.1 to three
separate estimates for border and interior regions, which compare outcomes across space in the
initial free trade regiment (EAC), the customs union (CU), and the common market era (CM)
all relative to the same pre-EAC period. As such, the estimates are directly comparable and
allow an insight into the temporal evolution of the mean increases in spatial inequality across
the EAC shown in Table 4.1. While these temporally differentiated effects confirm the average
effect shown in Table 4.1, the results provide three interesting additional insights. First, border
regions did not seem to have benefitted differentially more compared to more distant regions
in any of the EAC’s time periods. Importantly, this result is true even in the early years
following the re-establishment, which goes against Eberhard-Ruiz and Moradi (2019) who show
that growth of nightlights in the EAC was differentially higher for cities closer to the borders
in the initial periods. In fact, the present results show that it is more likely that regions closer
to the border experienced negative relative welfare reductions following re-establishment. Three
out of the four significant estimates show reductions in the DHS wealth index as well as
employed work opportunities (columns 4 and 5). The only positive effect depictable is a 9.8
163
Table 4.2: Difference-in-Differences across three Integration Periods
Consumpt.
Income
Agglom.
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[3.14]
[0.19]
[0.00]
Border [0,1] × EAC 1[2002-2004]
-0.083
0.050
0.098
-0.284
-0.129
0.104
(0.296)
(0.058)
(0.071)
(0.312)
(0.089)
(0.136)
0.779***
0.396***
0.166**
0.362
0.144
0.447***
Border [0,1] × CU 1[2005-2009]
0.172
0.060
0.064
-0.372*
-0.138**
0.084
(0.292)
(0.065)
(0.081)
(0.224)
(0.054)
(0.089)
0.555***
0.355***
0.428***
0.096
0.011
0.348***
Border [0,1] × CM 1[t ≥ 2010]
0.142
0.098**
0.119
-0.547***
-0.034
0.060
(0.282)
(0.039)
(0.101)
(0.202)
(0.040)
(0.095)
0.615**
0.012
0.241
0.007
0.384***
0.523***
Core {0,1} × EAC 1[2002-2004]
-0.261***
-0.013
-0.191***
0.433***
0.073*
0.146
(0.077)
(0.050)
(0.054)
(0.141)
(0.040)
(0.278)
0.001
0.797***
0.000
0.002
0.071*
0.599
Core {0,1} × CU 1[2005-2009]
-0.470***
-0.006
0.728**
0.314***
0.037
0.770***
(0.097)
(0.045)
(0.332)
(0.105)
(0.039)
(0.177)
0.000
0.897***
0.028
0.003
0.334***
0.000
Core {0,1} × CM 1[t ≥ 2010]
-0.338***
-0.076
1.083***
0.361**
0.006
0.956***
(0.065)
(0.069)
(0.375)
(0.142)
(0.041)
(0.111)
0.0000
0.2725
0.0039
0.0109
0.8791
0.0000
Individual Controls
YES
YES
NO
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
183,250
71,738
7,692
R-Squared
0.13
0.14
0.27
0.34
0.16
0.30
R-Squared -Within
0.11
0.10
0.25
0.33
0.11
0.29
Dependent Variable
Afrobarometer
DHS
percentage point increase in employed work in the CM period, which may or not be suggestive
evidence of a common market protocol, i.e. free movement of labor potentially benefiting
individuals closer to new markets. Second, while the positive welfare effect for households living
in the core agglomerations is present for all years, the effect is either reduced (column 4) or
Notes: The results in each column are produced by a separate regression. In columns (1) through (3), data come
from the Kenya, Uganda and Tanzania Afrobarometer surveys rounds 1 through 7 sampled between 2000 and 2020.
In columns (4) through (6), data come from the Kenya, Uganda and Tanzania Demographic and Health surveys
(DHS) sampled between 1999 and 2020. The sample mean of the respective dependent variable is given in brackets
above the estimates. For the DHS variable measured at the household level, Wealth Index in column (4), the answer
from the main survey respondents (women) is used. For the coefficients on Population Density in columns (3) and
(6), the level of observation is given by the unique sample enumeration areas of the respective survey. EAC 1[2002-
2004] switches to one for individuals sampled from 2002 to and including 2004, CU 1[2005-2009] for individuals
sampled from 2005 and including 2009, and CM 1[t ≥ 2010] for individuals sampled from 2010 onwards. All
regressions include individual-level controls for respondents’ age, gender, as well as education, and also include the
geographic controls average monthly temperature, average monthly rainfall, elevation, ruggedness, and the number
of growing days. The regressions also include country-year fixed effects. Binary dependent variables are estimated
through a Linear Probability Model (LPM). The standard errors reported allow for spatial correlation, i.e. Conley
standard errors are used. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
164
constant (column 1) over the progressive deepening of the EAC, and even vanishes for one of
the estimates (column 5). The estimates for changes in population density show the opposite
dynamic, which constitutes the third noticeable insight of Table 4.2. While in the early EAC
period, the results either show non-identifiable or even negative differential effects (column 6
and 3, respectively) in population density for survey locales close to the interior hubs, the
estimates on later periods provide evidence of growing agglomeration tendencies. Expressed
quantitatively, and in comparison to the rest of the country, the agglomeration in the economic
hubs grew more strongly by a magnitude of about one standard deviation when measuring
differences in the common market period to years preceding the EAC. Reconciling this
observation with the positive welfare effects of agglomerations (columns 1 and 4), which predate
these positive responses of population density in time, the finding provides suggestive evidence
for individuals responding with the dynamic proposed in (4.52) of Section 4.4, i.e. respond to
positive welfare differentials with migration inflows.90 Given that estimated welfare differences
stagnate or even decrease in the periods after these large population inflows (column 4), these
findings may also suggest that by moving, individuals endogenously regulate welfare differences
(downward). The loss of significance on the positive employed work outcome (column 5) from
the CU period onwards may provide some (weak) evidence for this in the form of an increase
in the elasticity of labor supply due to migration. One interesting aside on these first set of
results is their connection to the discussed issue of “urbanization without growth” (Fay and
Opal 2000). The effects show for increased consumption and agglomeration tendencies in
primate cities, without concomitant expansion of labor market opportunities, they provide
evidence on “urbanization without growth”, i.e. favor the notion of “consumption-” rather than
“production cities” associated with the urbanization of developing economies. Given their
discussed low shares of resource dependency, this goes against the results of (Gollin et al. 2016),
who illustrate resource dependency as a driving difference between these two distinct processes.
Extended Labor Market Results. The results on the remaining set of labor market
outcomes introduced in the Section 4.6 are presented in Table C.9 of the Appendix. The table
shows evidence on the differential development of having any type of activity outside of
household work within the last year (columns 1 and 4), whether conditionally on such an
activity, one is paid in cash compared to in-kind or no compensation, and, at which level of the
occupational skill dimension, i.e. Occupation Type {1,3} the work is situated.
Overall, the results across variables and samples are mixed and provide no systematic
evidence in favor of or against spatial inequalities following the establishment of the EAC. The
most robust result is a positive change in the likelihood of having work outside of household
activities for individuals closer to border regions. However, the effect is decreasing in absolute
terms and in statistical significance over time (column 4). Also, evidence on the type or
90 The effects on consumption are strongest in the EAC and CU years (columns 1 and 4), the effects on
population density in the periods directly succeeding these eras (columns 3 and 6).
165
remuneration of this work shows that the likelihood of it being low-skilled and unpaid work is
larger (compare columns 5 and 6). Concerning core agglomerations, there is some weak evidence
on having work (column 4) and being remunerated for it, household’s occurrence of
experiencing monetary droughts increased in the CU period (column 3). However, given the
inconsistency of these results, further interpretation of these results remains contentious. The
next section looks at these issues arguably more fittingly, using the panel dataset of the sample.
As anticipated above, one related possibility which cannot be isolated by the empirics is that
a quick inflow of population attenuates a potentially positive labor market effect given the
oversupply of it. This issue is better assessed with the panel dataset of the sample with which
one may partly entangle these effects and to which we turn to next.
Kagera Health and Development Survey. Table 4.3 provides the final set of main results,
reporting estimates produced from the Kagera Health and Development Survey (KHDS). A
few notes towards the interpretation of results in comparison to the Afrobarometer and the
DHS. First, given the timeframe of the survey, I am unable to provide longer-term evidence as
the third and final sampling round was conducted in 2010 and Table 4.3 thereby reports
estimates on the effect of the EAC and CU periods only. Second, some survey items were not
administered in 2010, such that values cannot be estimated for these time periods either. Third,
concerning the interpretation of the Core {0,1} dummy. Given the spatial limitations of the
KHDS survey in the first waves between 1991-1994, I am unable to measure outcomes for
households in Dar es Salaam prior to 2004. Hence, the DiD estimate on Core {0,1} is given by
comparing differences of individuals over time living in Kagera’s urban capital “Bukoba” in
addition to those individuals who have later moved to (or were initially sampled in) Dar es
Salaam or Kampala in the second and third rounds. The robustness checks in Subsection 4.7.2
removes the latter group and thereby provide a test on differential spatial sorting and on the
general tendency of agglomeration predicted in Section 4.4.91 Fourth and last, as a panel, the
KHDS allows the inclusion of household-, respectively, individual fixed effects, and
identification thereby stems from changes within households or individuals across time.92
Columns (1) and (2) depict the first set of results which test for differential changes in
per capita consumption of food and non-food items as well as the value of durable items in the
household. Both figures are expressed in deflated Tanzanian Shilling (TZS) priced in constant
2010 and 2004 levels, respectively. Households living in core agglomerations increased their
consumption following the establishment of the EAC compared to households in the auxiliary,
while households living closer to the regions bordering the new markets do not evince
statistically significant differences in either integration period. Quantitatively, each individual
living in households in core agglomerations consumes an extra of over 170,320 TZS worth of
91 I pick up concerns about spatial sorting in Subsection 4.7.2 on Robustness Checks, Validity Tests, and
Extensions.
92 Depending on whether questions are administered at the household- or individual-level.
166
food and non-food items more than in the pre-EAC period, compared to the development of
the rest of the country over the same time span. In relative terms, this indicates an increase of
Table 4.3: Difference-in-Differences using the KHDS Survey
Agglom.
Annual p.c.
Consumpt.
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border [0,1] × EAC 1[2004]
-853.376
187.152
0.581*
0.071
0.107
-0.848**
(1072.010)
(754.481)
(0.331)
(0.049)
(0.261)
(0.384)
0.426
0.8043
0.079
0.1532
0.681
0.027
Border [0,1] × CU 1[2010]
-1317.845
0.432
-0.010
-1.631***
(1154.386)
(0.305)
(0.404)
(0.586)
0.254
0.156***
0.981***
0.005
Core {0,1} × EAC 1[2004]
170.320***
706.607***
-0.013
0.026***
0.134***
0.891***
(42.328)
(192.323)
(0.023)
(0.008)
(0.021)
(0.070)
0.000
0.0003
0.560
0.0019
0.000
0.000
Core {0,1} × CU 1[2010]
275.036***
-0.039
0.069
0.891***
(57.368)
(0.025)
(0.067)
(0.099)
0.0000
0.1187
0.3013
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.43
R-Squared -Within
0.12
0.09
0.04
0.01
0.14
0.41
Consumption
Dependent Variable
Kagera Health and Development Survey (KHDS)
Income
Notes: The results in each column are produced by a separate regression. Data come from the Kagera
Health and Development Surveys (KHDS) collected in four waves across 1991-1994, as well as one wave in
2004 and 2010, respectively. In columns (1) and (2) outcome variables represent aggregate household
information provided by the head of the household provided by the head of the household, in columns (3)
through (5) they are administered on an individual level. For the coefficient on Population Density in
column (6), the level of observation is given by the unique sample enumeration areas available in the sample.
Certain indicators were not sampled in the survey wave of 2010, which is why there is no estimate given
for these columns. The sample mean of the respective dependent variable is given in brackets above the
estimates. Border [0,1] is the inverse, relative within country distance to the nearest Border crossing. Core
{0,1} is a dummy indicating individuals living in the core agglomeration of their respective country (i.e.
Dar es Salaam and Kampala). For the initial KHDS survey waves “Bukoba” - the capital of Kagera
represents the core agglomeration. EAC 1[2004] switches on for individuals (re-)sampled in 2004. CU
1[2010], switches on for individuals (re-)sampled in 2010, the second re-interview period of the KHDS. All
regressions include individual-level controls for respondents’ age, gender, as well as education, and also
include the geographic controls average monthly temperature, average monthly rainfall, elevation,
ruggedness, and the number of growing days. The regressions also include an indicator whether the
household is living in proximity to (former) refugee camps. The regressions testing household-level
outcomes, columns (1) and (2), include household fixed effects, the regressions testing individual-level
outcomes, columns (3) through (5), include individual fixed effects. All regressions include country-year
fixed effects. Binary dependent variables are estimated through a Linear Probability Model (LPM). The
standard errors reported allow for spatial correlation, i.e. Conley standard errors are used. ***, **, *
represents significance at the 1, 5 and 10 percent level, respectively.
167
roughly 31% compared to the average of the sample mean. This figure grows to almost 50% in
the post-EAC years. Households also increase the consumption of consumer durables, owning
a stock of goods in value of over 706,607 TZS higher than the comparison group. While I
cannot identify positive effects on a higher likelihood of having employed work in
agglomerations, there is a weakly significant effect at borders. A change from the median
distance to borders in the sample to the minimum distance leads to an increase in the likelihood
of employed work by 5.0 (=0.581×0.086) percentage points. However, the remaining labor
market outcomes paint a more consistent picture compared to the cross-sectional results. That
is, individuals in core agglomeration profit differentially more from the establishment of the
EAC both in terms of extensive and intensive labor market outcomes. Living in core
agglomerations compared to the periphery increases the likelihood of having salaried work by
2.6 percentage points and raises the occupation type in which the individual works by 0.134
units. In other words, the skill gap in work done by individuals in capital cities versus
peripheries increased by another 11% relative to the sample mean.
As a further exploration into these dynamics Table C.10 of the Appendix provides the
findings on the remaining labor market outcomes introduced in the data Section 4.6. The
findings show no differential effects across space for the likelihood of having worked in the past
year or showing an increase in monthly salary.93 Table C.10 also tests an expanded set of
consumption components, i.e. tests for differences in food vs. non-food consumption and tests
the value of the occupied dwelling of the household. With the exception of row 3 of column
(2), all results confirm the evidence shown in Table 4.3, namely that households proximate to
borders do not gain differentially more than interior regions but those living in the core
agglomerations, which show strong, positive effects on consumption and wealth. Concerning
the last dimension of and contributing to spatial inequality, population density, we see
quantitatively highly similar and qualitatively identical results compared to the cross-sectional
samples. The disparity in population density is growing, by up to half a standard deviation as
measured in 2010. Evincing strengthened agglomeration patterns.94 In sum, the results from
the panel setting support the evidence produced from the cross-sectional samples.
4.7.2. Robustness Checks, Validity Tests and Extensions
Robustness Checks. As a first insight into the stability of the presented results, I conduct
an array of robustness tests summarized in Table C.3 through Table C.8 of the Appendix and
briefly discussed here. As before, all of the results are produced using regression specification
(4.53), unless otherwise indicated. The tables also report the baseline coefficients from Table
93 Interestingly, regressions relaxing the strict individual fixed effects show a strong differential increase of
wages in agglomerations and a decrease of them at borders (significant at the 10 and 5 percent level).
94 An explanation of the larger and statistically more significant effect in the initial EAC period is that this
estimate measure changes in the tail end of the EAC period (i.e. in 2004), in contrast to the earlier samples,
which average changes between 2001 and 2004.
168
4.2 in the top rows with which to compare the results of the adapted estimations. Given the
lack of effects for border regions throughout the samples and variables tested, the table confines
the reporting of the three difference-in-difference coefficients on Core {0,1}. The full results
including the DiD estimates for Border [0,1] can be accessed in the source tables which are
referred to in these summary tables at each instance of the respective test.
To begin, (a) allows for the clustering of standard errors at the enumeration area level
instead of implementing Conley standard errors. (b) Removes all individual and geographic
controls from the regression.95 (c) Adds the extended set of geographic controls anticipated in
Subsection 4.6.1, namely locations’ Absolute Latitude, Malaria Ecology (Sachs et al. 2004),
Navigable Rivers, Major Lakes and Major Harbors to control for other trade-related influences
and adds the dummy non-EAC ≤ 100km {0,1} as well as the interaction with all EAC period
dummies to net out effects potentially stemming from a change (loss) in market access at non-
EAC borders. (d) Employs the sample survey weights provided by the Afrobarometer and DHS,
accounting for the pooling across countries and years by standardizing the weights for each
country-survey round pair.96 (e) Excludes low-precision localities. For the Afrobarometer, this
is implemented by dropping all observations for which the AidData precision code is above 2
(AidData 2017).97 In the DHS survey, I drop all observations for which coordinates are not
generated from a GPS receiver used by the fieldworker.98 This test cannot be conducted for the
KHDS as there is no distinction in the precision of GPS locales. (f) Replaces the Core {0,1}
dummy in specification (4.53) with Agglomeration {0,1} which switches to 1 for households
living within 50km of an “urban center” demarcated as such in the year 2000. The list of urban
centers I use are provided by the European Commission (2019).99 This can be seen as a general
test on the agglomeration tendency of the results in Subsection 4.4.5 irrespective of border
distance or proximity to the capital cities. As a further test on this, I also try the Urban {0,1}
dummy attached to the surveys.100 (g) Reduces the spatial cut-off criteria for living in core
agglomerations to 25km and (h) to 10km, respectively. (i) Splits the CM period into two
dummies, namely CM 1[2010-2014] and add a post-CM time period post-CM 1[t ≥ 2015] to
provide a test on the hypothesized transitory shift to a welfare-equalizing equilibrium as
discussed in the previous section.101 (j) Excludes the individuals in the sample which did not
live in the survey location at least three years before the establishment of the EAC (before
95 For the KHDS survey, I cluster observations at a specific geographic delineation, which is based on 2 decimal
places of latitude-longitude combination, i.e. raster of slightly larger than 1 square km. at the equator.
96 I.e. transform the weights such that they sum to 1 for each pair.
97 The scale ranges from 1-8. Using precision code 1 leads to a loss of data in the range of 40-50%.
98 See https://dhsprogram.com/Methodology/upload/MEASURE-DHS-GPS-Data-Format.pdf.
99 The definition reads: “The spatially-generalized high-density clusters of contiguous grid cells of 1 km2 with
a density of at least 1,500 inhabitants per km2 of land surface or at least 50% built -up surface share per km2
of land surface, and a minimum population of 50,000.” (European Commission 2019; 13). For my purposes, I
use a minimum population threshold of 100,000.
100 Results can be requested from the author.
101 Notice that “post-CM” has no further meaning other than nomenclature. The common market of the EAC
has continued to persist.
169
1999), i.e. excludes “post-EAC Migrants”.102 (k) Excludes DHS’ special survey rounds including
the AIS, KAP and MIS surveys, leading to a sample of 141,879 individuals living in 88,196
households located across 5,110 survey locales and interviewed between 1999 and 2016. (m)
“logs” relevant dependent variables, i.e. transforms the variables in the way log(1+y), for
outcomes expressed in constant Tanzanian Shilling (only Table C.5).
While the general upshot of all these sensitivity checks is that all previous conclusions
and interpretations hold, there are some interesting takeaways for two specific tests which also
have bearings on the theoretical results. For one, by splitting up the CM years into two periods
(j), we notice that the convergence anticipated in Table 4.2 is at least partly corroborated.
While the Afrobarometer shows seemingly unaltered results compared to the years 2010-2014,
estimates produced with the DHS survey, which include three more survey years from 2017 to
and including 2020, drop both in size as well as significance. Most notably, the coefficient on
population density is halved compared to the “early” CM period. And secondly, when
exchanging the Core {0,1} dummy with a wider selection of urbanities (f), the effects are either
non-significant, significantly weaker, or point in the opposite direction. Particularly interesting
are the effects on population density which are insignificant for most of the estimates, and
significantly lower in magnitude (by an order of 4) when compared to the developments in the
core agglomeration. Only for the KHDS are results similar in magnitude and significance. This
may be reconciled by the fact that the dummy is a test on “urbanities” rather than economic
hubs existent in 2000, and that “urbanities” are endogenous to the outcome. However, given
that endogenous formation of agglomeration is precisely what theory dictates, I take this
observation seriously and evaluate this effect in the context of the main theoretical model, i.e.
with heterogeneous intra-national space. I do so by estimating a triple-difference specification
which tests for a differential effect for (endogenous) agglomerations at border regions, which is
effectively done by interacting the previous treatment interaction Border [0,1] × EAC [t ≥
2002] with the Urban {0,1} dummy provided in the Afrobarometer and DHS. The results are
shown in Table C.11 of the Appendix. The combined effect in row three provides no convincing
evidence that agglomeration newly establishes at borders, and that relative welfare increases
at those agglomerations are differentially higher. Only the DHS wealth index is significant at
common levels (column 4), providing evidence of reduced relative wealth of households located
in urbanities at the border. Panel b) tests the extended set of labor market outcomes. Again,
most results are insignificant or strongly negative (columns 5 and 6). This confirms the previous
results which did not show a draw to the border.
Next to this battery of tests on the stability of the DiD coefficients on Core {0,1} I also
provide relevant robustness test on Border [0,1], which are summarized in Table C.6 through
Table C.8. (a) Replaces distances calculated over road with a simple beeline (“as the crow
flies”) distance to EAC border crossings. (b) Approaches in similar but estimates effects for the
102 Because of missing migration information, this test is only possible for the DHS and KHDS sample.
170
nearest (beeline) distance on the entire borderline formed by two EAC member countries. (c)
“logs” the original road distance measure, in the fashion log(1+x). Relatedly, Table C.82 and
Table C.83 provide results for a more flexible distance specification by entering the continuous
distance to internal EAC border crossings and core agglomerations as well as their squared
value as explanatory variables. (d), (e) and (f) replace the continuous treatment indicator
Border [0,1] with dummy indicators of living within 100km, 50km, and 25km to the Border
Crossings. In sum, the previous findings of non-identifiable, or negative, differences across
border distance before and after the EAC’s re-establishment remains robust.103 The only slight
deviations from the results presented before are shown in the border dummy specifications (d)
through (f) which show evidence of positive effects on consumption for the narrowest distance
bandwidth (≤ 25km). We explore some caveats with these effects in the extensions below.
One last test on the robustness of the results is provided in Table C.20 and Table C.21
where I test for the composite items of the two indices employed, i.e. Lived Poverty [0,4] and
Occupation Type {1,3}. As expected, the results are robust to the decomposition. Interestingly,
in line with positive effects at borders in the CM period we see a shift of occupations away
from agrarian jobs to semi-formal engagements, at least for the Afrobarometer and KHDS
surveys, respectively. However, this effect is also apparent for individuals in core
agglomerations, and larger in size. The last noteworthy comment regards the alternative
consumption and wealth indices, the International Wealth Index and the Comparative Wealth
Index. What is more strongly confirmed in these results is that the effect of increased
consumption in agglomerations is a one-off effect which is decreased strongly across time. Again,
this provides suggestive evidence on population inflows to equalize welfare differences due to
negative agglomeration externalities (e.g. congestion, housing market).
Validity Tests. The fundamental assumption behind the employed difference-in-differences
design requires that absent policy change, the spatial disparity in households’ welfare within
the EAC countries would have evolved “in parallel” i.e. continued their relative pre-intervention
trajectories. In other words, for our estimates to represent a causal relationship, nothing other
than the policy of regional market integration should have induced a differential welfare change
across space in the timespan (shortly) the before and after the re-establishment. While in
practice ultimately never verifiable, I provide three distinct pieces of evidence that may
strengthen our confidence in this assumption.
Before more formal tests are discussed, there needs to be an initial check on other
(concomitant) policy measures with the potential to influence the economic geography of the
respective countries. In the timespan between 1995 and 2010, the most relevant policies I was
able to identify mainly concentrated on trade facilitation. For instance, the Northern Corridor
Transport Improvement Plan (NCIP) of 2004 aimed to improve transport infrastructure to
103 One exception is the now positive effect on Occupation Type {1,3} for border regions in the CM period.
171
facilitate trade integration. I do not deem this investment as undermining the results, as the
completion of project goals aimed within the NCIP were temporally lagging the main results
presented here by a large margin (see World Bank 2016). A related concern is that the
facilitation of “one-stop-border-posts” (OSBPs) may lead to differential success of integration
across border. However, as in the case of the NCIP most OSBPs were erected many years after
the large increases first set in in our results (Cadot et al. 2015; EAC 2015).104 Another initial
concern may be displayed by “Export Processing Zones” (EPZs) or “Special economic Zones”
(SEZs) in or near the core agglomerations defined in the paper. While all of the three member
countries actively promote SEZs, the timing as well as the spatial pattern relief concerns of an
effect entirely attributable to such developments. For instance, in Kenya, the majority of EPZs
are outside of Nairobi, many of them in the port of Mombasa not included in the Core
Definition.105 In Tanzania, the operation of EPZs is possible since the ratification of the EPZ
act of 2002 and the SEZ act of 2006, respectively. However, while data on firms operating under
such licenses is untransparent with several contradicting reports on the absolute number, they
agree that the general impact to industrialization was small (Andreoni et al. 2022). Concerning
the spatial dimension, most recent data suggests that the majority of firms operating under an
SEZ license are outside of Dar es Salaam (Kinyondo et al. 2016; Andreoni et al. 2022).106 Lastly,
the distribution of EPZs in Uganda does in fact evince a stark regional disparity skewed towards
the Central region enclosing Kampala (UFZA 2022). However, SEZ only began operating after
2014, which is the year the Ugandan Free Zones Act was ratified, rendering our results robust
to this development.
Placebo Tests. The first formal test against the difference-in-differences assumption is
presented in Table 4.4. The results shown are produced in the identical way as they were in
Table 4.1, but use data on contiguous “placebo countries”, i.e. estimate a differential change
across pre and post-EAC time periods for non-EAC countries bordering Kenya, Tanzania and
Uganda. The countries available in the data are Malawi, Mozambique, and Zambia in the
Afrobarometer. The DHS expands the countries to include Ethiopia and Rwanda.107 In these
countries Border [0,1] represents the inverse, relative within-country road distance to the
nearest major road crossings with an EAC country and Core {0,1} identifies individuals living
within the respective core agglomeration (i.e. Addis Abeba in Ethiopia, Kigali in Rwanda,
Lilongwe in Malawi, Lusaka in Zambia and Maputo in Mozambique). The results provide
104 I also estimate a potential heterogeneity across border regions in Table C.17 of the Appendix, which is
discussed at the end of this subsection. The findings provide no evidence of such an influence.
105 While many of them are in the Machakos county, next to Nairobi, road and beeline distances are above the
commonly used threshold of 50km. Also note that the results are robust to the narrower spatial delineation
(i.e. 25 and 10km), if one assumes laborers to commute to these EPZs.
106 Even though the largest of the SEZ zones are located in the Dar es Salaam-Bagamoyo corridor. Bagamoyo
is in similar distance to Dar es Salaam as Machakos to Nairobi in Kenya.
107 We include data from Rwanda only until 2005, given that the country joined the EAC in 2007.
172
Table 4.4: Placebo Tests – DiD in contiguous non-Intervention Countries
Consumpt.
Income
Agglom.
Consumpt.
Income
Agglom.
Panel a)
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.25]
[0.24]
[0.00]
[3.20]
[0.17]
[0.00]
Border [0,1] × EAC 1[t ≥ 2002]
0.179
0.062
-0.031
0.682***
0.015
-0.176
(0.143)
(0.075)
(0.159)
(0.137)
(0.039)
(0.128)
0.210
0.409***
0.848
0.000
0.696***
0.167
Core {0,1} × EAC 1[t ≥ 2002]
-0.142
-0.040
0.773**
0.130
-0.067
-0.111
(0.103)
(0.041)
(0.303)
(0.167)
(0.087)
(0.400)
0.169
0.321
0.011
0.436
0.438
0.781
Observations
28,541
18,994
2,329
234,346
126,587
9,026
R-Squared
0.09
0.19
0.23
0.30
0.14
0.27
R-Squared -Within
0.05
0.14
0.22
0.29
0.13
0.25
Panel b)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[0.49]
[1.62]
[2.12]
[0.67]
[1.48]
[0.44]
Border [0,1] × EAC 1[t ≥ 2002]
0.022
0.027
-0.579***
-0.033
-0.055
-0.106
(0.074)
(0.138)
(0.141)
(0.081)
(0.104)
(0.120)
0.766***
0.842***
0.000
0.682***
0.598***
0.375***
Core {0,1} × EAC 1[t ≥ 2002]
-0.038
-0.016
-0.361***
0.083*
0.048
0.040
(0.055)
(0.068)
(0.073)
(0.048)
(0.068)
(0.091)
0.493
0.810
0.000
0.086
0.479
0.664
Observations
19,191
12,715
28,274
310,644
187,498
196,153
R-Squared
0.18
0.31
0.11
0.17
0.24
0.18
R-Squared -Within
0.10
0.27
0.06
0.15
0.22
0.13
Individual Controls
YES
YES
NO/YES
YES
YES
NO/YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Dependent Variable
Afrobarometer
DHS
Afrobarometer
DHS
Income
Notes: This table conducts a “placebo” analysis by testing for a spatially differentiated effect across contiguous,
non-EAC countries within the time frame of the EAC’s establishment and expansion. As such, in columns (1)
through (3), data come from the Malawi, Mozambique and Zambia Afrobarometer surveys rounds 1 through 7
sampled between 1999 and 2018. In columns (4) through (6), data come from the Ethiopia, Malawi, Mozambique,
Rwanda, and Zambia Demographic and Health surveys (DHS) sampled between 2000 and 2019. The sample mean
of the respective dependent variable is given in brackets above the estimates. EAC 1[t ≥ 2002] switches to one for
individuals sampled from 2002 onwards. All regressions include individual-level controls for respondents’ age,
gender, as well as education, and also include the geographic controls average monthly temperature, average
monthly rainfall, elevation, ruggedness, and the number of growing days. The results in each column and panel are
produced by a separate regression. The regressions also include country-year fixed effects. Binary dependent
variables are estimated through a Linear Probability Model (LPM). The standard errors reported allow for spatial
correlation, i.e. Conley standard errors are used. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively.
173
suggestive evidence that the differential welfare change in economic hubs was not a larger,
regional trend within the timeframe under evaluation. More generally, the results do not show
any consistent evidence of spatial inequalities increasing or decreasing over time. While there
are positive effects on population density in the Afrobarometer, the arguably more precise
spatial estimates produced from the DHS does not confirm this finding. Further, while there
some improvements in wealth at borders (column 4 of panel a), column 3 in panel b) shows
that the occurrence of cash income droughts was similarly strongly decreased in core
agglomerations.108 The rest of the results show no hint towards a differential development
following the EAC, which strengthens the confidence in attributing the previous estimates to
the EAC. For completeness, Table C.12 of the Appendix shows the companion results which
splits the EAC 1[t ≥ 2001] dummy into the three integration periods. The results from this
more nuanced regression do not provide interpretations different than the ones just discussed.
One interesting aspect, however, is that, in the DHS sample, many of the significant effects for
core agglomerations are only apparent in the EAC period. Given that they almost universally
imply a negative impact and show no accompanied increase at borders, I deem these results as
non-indicative of other region-wide processes influencing the results from the EAC countries.109
Pre-Trends. One weakness of the datasets employed is the narrow timeframe of pre-
integration periods. While a difference-in-difference approach implemented on data shortly
before and shortly after the intervention may alleviate concerns of other concomitant policies
driving the results, it excludes the possibility to net out potentially longer-term, unit-specific
time trends from the identified treatment effects. Also, it does not allow a test on the pre-
intervention evolvement of relative welfare within the countries. Given that this is a crucial test
concerning the validity of the results, I try to circumvent this data restriction by drawing on
region-based estimates from the country, i.e. include non-GPS coded survey rounds of the DHS
which extend back to 1988.110 While there is no finely gridded information on the location of
respondents available, DHS provides regional-based information, which identifies individuals’
residence on the GADM-1 level.111 I use this information and construct a new Core Region
{0,1} dummy indicating whether individuals live in the capital city region. Note that while
many of the surveys in pre-GPS years do not provide granularity of the measure employed
previously, given their political and economic importance, Nairobi and Dar es Salaam were
nonetheless demarcated as their own region at finer levels.112 As such, the use of this Core
Region {0,1} dummy likely captures much of what is also measured by the GPS-based Core
108 Going from the median distance of the sample to the border (912km) renders an effect size of -0.257
(=0.597×0.43).
109 Four out of five estimates imply negative welfare changes at core agglomerations.
110 See section on Data sources 4.6.
111 See https://spatialdata.dhsprogram.com/boundaries/.
112 For Kampala in Uganda, this was only done in the 1988-1989 survey for non-GPS surveys. After this year
and until 2002, Kampala was located in the “central” region from which point on it was its own district again.
174
{0,1} dummy used in the main estimations.113 Concerning the definition of border regions, the
matter is not as straightforward. Many of the regions defined in the DHS, particularly in the
pre-GPS years, could be considered both border as well as interior regions given their vast
extent to the inland of countries (see footnote 111). Hence a dummy categorization as used for
capital cities will not likely suffice to capture true border households. I therefore try to improve
upon a simple dummy with the following steps. First, I assign households in all survey rounds
a regional correspondence for which boundaries are consistently available from as early on as
possible. This yields 7, 20 and 4 consistently demarcated regions for the years 1988 until 2020
for Kenya, Tanzania and Uganda, respectively.114 Second, using all available GPS samples, I
retrieve the mean, modal as well as median values of road distances within these boundaries as
central tendencies of the distribution within them. Third, I assign these values to all households
nested within a specific region, especially now also for the non-GPS households. Finally, I
encode those households living in regions ranking in the 10th percentile of these boundary-based
distances within their country as a 1 in the dummy Border Region {0,1}.115 Under the
assumption that the population distribution within regions has not dramatically changed
between 1988 and the latter two decades, this arguably allows a more precise ordering of border
to non-border regions within the sample.
Table 4.5 presents the result using these two region-based indicators, using the mean as
the central tendency for the Border Region {0,1} dummy. Panel a) relates to Table 4.1 and
shows the aggregate difference-in-differences estimates before and after the establishment of
the EAC, with the omitted time period, i.e. the reference group now defined from 1988-2000.116
Panel b) explicitly tests for pre-trends by introducing a pre-EAC 1[1996-2000] dummy which
tests for differential changes in border-, respectively, capital city regions over the time period
before the establishment of the EAC. The reference group for all estimates shown in this panel
has therefore changed to pre-EAC 0[1988-1995]. Panel a) confirms the result shown in Table
1 from the GPS-based measurements. There are no indications of a positive relative household
welfare change across all tested outcomes.117 And further, we see the same strongly positive
effects of welfare for households living in the core agglomerations.118 Turning to the test of
parallel trends in panel b), we see no robust indication of an unequal trend of core vs. peripheral
regions. Only one of the estimates is significant at conventional 5%-levels (column 6). However,
113 Indeed, for survey rounds where GPS information is available, the Pearson correlation coefficient between
Core {0,1} ≤ 50km (the main dummy) and Core Region {0,1} is 0.694, for Core {0,1} ≤ 25km it increases to
0.749 and for Core {0,1} ≤ 10km it correlates to 0.681.
114 For one round, the 1991-92 Tanzania DHS, I am confined to 6 regions.
115 For survey rounds where GPS information is available, 61.6% (54.5%) of individuals scoring a 1 for this
dummy live within 100km of the (granularly) calculated road distance. To compare, a simple border
categorization leads to a value of 37.5%. The results are also robust to using the 20th percentile.
116 Standard errors are clustered at the “region” level and are further robust to clustering at the “cluster” level
also.
117 Indeed, column 6 shows evidence of a negative effect of cash employment.
118 Note that Population Density, as location-based outcome variable, is constructed in the same fashion as
the region-based distances.
175
Table 4.5: Region-Based Estimates and Pre-Tests
Consumpt.
Income
Agglom.
Panel a)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.12]
[0.19]
[0.00]
[0.76]
[1.52]
[0.53]
Border Region {0,1} × EAC 1[t ≥ 2002]
0.021
-0.033
-0.104
-0.082
-0.042
-0.133**
EAC 0[1988-2000]
(0.170)
(0.022)
(0.137)
(0.063)
(0.051)
(0.055)
0.904**
0.156*
0.452
0.200***
0.419***
0.023***
Core Region {0,1} × EAC 1[t ≥ 2002]
0.386***
0.090***
2.203*
-0.037
0.089*
-0.002
EAC 0[1988-2000]
(0.122)
(0.022)
(1.223)
(0.054)
(0.050)
(0.030)
0.0035
0.0003
0.0815
0.4978
0.0824
0.9383
Observations
258,820
104,440
11,841
236,646
140,613
136,163
R-Squared
0.28
0.16
0.54
0.21
0.25
0.18
R-Squared -Within
0.27
0.10
0.53
0.18
0.16
0.08
Panel b)
Border Region {0,1} × pre-EAC 1[1996-2000]
0.098
0.024
0.084
-0.296**
0.014
-0.063
pre-EAC 0[1988-1995]
(0.125)
(0.029)
(0.082)
(0.121)
(0.060)
(0.049)
0.436***
0.413***
0.315***
0.020
0.817***
0.208
Border Region {0,1} × EAC 1[t ≥ 2002]
0.088
-0.017
-0.068
-0.304**
-0.031
-0.171**
EAC 0[1988-1995]
(0.200)
(0.032)
(0.120)
(0.134)
(0.078)
(0.065)
0.663
0.593***
0.575***
0.030
0.693***
0.013
Core Region {0,1} × pre-EAC 1[1996-2000]
0.440*
0.022
0.131
-0.245*
0.099
-0.185***
pre-EAC 0[1988-1995]
(0.223)
(0.026)
(0.261)
(0.121)
(0.102)
(0.052)
0.058
0.408***
0.621
0.051
0.340**
0.001
Core Region {0,1} × EAC 1[t ≥ 2002]
0.652***
0.102***
2.255*
-0.211
0.147
-0.094**
EAC 0[1988-1995]
(0.229)
(0.023)
(1.294)
(0.137)
(0.087)
(0.045)
0.008
0.000
0.091
0.133
0.102
0.043
Observations
258,820
104,440
11,841
236,646
140,613
136,163
R-Squared
0.28
0.16
0.54
0.21
0.25
0.18
R-Squared -Within
0.27
0.10
0.53
0.18
0.16
0.08
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
NO
NO
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Dependent Variable
DHS (Region-based)
Income
Notes: This table makes use of the non-GPS survey rounds of the Demographic and Health Surveys (DHS) sampled
before 1999 and additionally conducts “pre-tests” towards the difference-in-differences approach. The data thereby
come from the full sample of Kenya, Tanzania, and Uganda DHS surveys sampled between 1988 and 2004, making
use of AIS, KAP and MIS rounds as well. The sample mean of the respective dependent variable is given in brackets
above the estimates. Border Region {0,1} switches to one for individuals living in a region with a median road
distance to Border crossings below the 10th percentile of all (within-country) GPS-border distances in the sample.
Core Region {0,1} is a dummy indicating individuals living in the region which hosts the core agglomeration of
their respective country (i.e. Nairobi, Dar es Salaam and Kampala). EAC 1[t ≥ 2002] switches to one for individuals
sampled from 2002 onwards. Pre-EAC 1[1996-2000] switches to one for individuals sampled in survey years between
1996 and including 2000. As such, in panel a), the reference group of the estimates are comprised of individuals
sampled in the full pre-EAC period, i.e. from 1991 to 2000, while in panel b), the reference group is formed by
individuals sampled between 1988 and including 1995. Hence, the DiD estimate on “pre-EAC” in panel b)
represents the pre-test. The results in each column and panel are produced by a separate regression. All regressions
include individual-level controls for respondents’ age, gender, as well as education. The regressions also include
country-year fixed effects. Binary dependent variables are estimated through a Linear Probability Model (LPM).
The standard errors reported allow for clustering at the “region” level. ***, **, * represents significance at the 1,
5 and 10 percent level, respectively.
176
the coefficient signals a sign opposite to what is expected, and then attenuates for the EAC
estimate which would be consistent with a positive welfare effect in core agglomerations. I
conduct the same estimations with the placebo-countries analyzed in Table 4.4, and the results
are reported in Table C.13. As expected from the earlier findings of the contiguous countries,
the results do not provide consistent evidence of a growing spatial inequality, which further
negates the notion of an ongoing trend towards more pronounced economic inequality in the
larger East African region. Importantly, the results are fully robust to using the median or
mode road distances to the border as well. They are additionally robust to using the simpler
border categorization. Appendix Table C.14 and Table C.15 also provide the results on the
temporally disaggregated DiD effects. All of the previous conclusions remain. The added
temporal dimension of the region-based samples allows for one further (placebo) test which is
presented in Table C.16, Figure C.9 and Figure C.10.119 The estimates shown are produced
from the identical regression specification as for Table 4.1, but now the single temporal
treatment dummy switches to one for varying (placebo) intervention years which are plotted
on the x-axis, respectively. The estimates therefore test for differential changes across space
when one would not expect such changes if it was indeed the re-establishment of the EAC that
has influenced the spatial pattern during the true intervention years. Importantly, the sample
used for the estimation is re-centered at the respective (placebo) intervention year such that
(actual) post-EAC data does not influence the estimates too strongly. The estimation always
uses the earliest possible survey waves for which entries of the respective outcome are variable
but truncates the data to the right. For example, in the case of the placebo treatment year
1994 the sample includes household-level data from 1988 to 1997. For the placebo treatment
year 1995, the sample is expanded to include 1998, and so on. The results shown are in
accordance with the true intervention kicking in the early 2000s. There are almost no significant
differential trends for border or core agglomerations throughout the 1990s. For the few that
are, coefficient stability may be a concern, and is likely due to the rather infrequent sampling
in the (very) early sampling years.120 Effects become significant from 2000 onwards with a peak
of point coefficient size in 2002.
Columns (4) through (6) of Table C.16, as well as Figure C.9 and Figure C.10 provide
the last piece of evidence in this regard, by plotting these same difference-in-difference estimates
together with their 95% confidence interval against those estimated from the sample of placebo-
countries defined above.121 Note that for legibility, the variables in the plots are transformed
into z-scores and can therefore not be directly compared to Table C.16 concerning coefficient
119 To improve legibility, the figures are split in that Figure C.9 analyzes the effects on core agglomerations,
and Figure C.10 on border regions.
120 For instance, for the Wealth Index the first available data is provided in the Kenya DHS 1993, with the
next available samples available in Uganda in 1995, Tanzania in 1996 and again, Kenya in 1998.
121 Note that the plot starts with the 1997 estimate given that data for the placebo countries as only few
samples provide the main dependent variables before this date. Again, Wealth Index is available only in Malawi
in 1992 Rwanda in 1992, and Zambia in 1996. 1997 is the first year for which repeat observations in a country
are available.
177
size, albeit against each other. As can be depicted in Figure C.9, placebo estimates do not
evince the same “hump” as those from the EAC countries in the first years following the
establishment of the EAC. Figure C.10 shows that a discontinuous border effect is apparent
neither in EAC- nor in placebo countries. What is interesting in the results on placebo countries
is that they show evidence of a positive effect on consumption in borders in the early EAC
timeframe paired with negative effects at their core agglomerations following these years. A
thorough investigation of this is out of the scope of this paper, but needs to be mentioned,
nonetheless. In sum, these results strengthen our confidence that our estimates reflect the effect
of regional market integration and do not conflate them with longer-term trends over time or
region-wide effects at a specific point in time.
Spatial Sorting. The last validity test offered addresses concerns about the potential spatial
sorting of individuals within the intervention years, which would systematically influence the
measured treatment. Specifically, there exists the possibility that skilled individuals positively
select, i.e. move, into border-, respectively, capital city regions within treatment years in
attempt to profit from the policy or other complementarities such as more productive firms, or
joint productivity (Kremer 1993). This would lead to an upward bias of the observed results.
Given that most of the results are drawn from cross-sectional households, entirely excluding
this possibility is not possible. However, I offer four-part evidence that this may be unlikely.
First, using regression specification (4.53), I test for the duration of having lived at the current
locale using the KHDS survey to test for differential migration into border or capital city
regions following the establishment of the EAC as well as the common market (CM) protocol.
The results are shown in column 6, panel b) of Table C.10 in the Appendix and provide no
evidence on differential migration movements of sample respondents across regions induced by
the EAC. Importantly, individuals are not more likely to have recently migrated to the primate
city before and after the EAC compared to other locales. Second, as seen in the robustness
checks, the results are (highly) robust to excluding “post-EAC migrants”, i.e. individuals that
moved to the respective region of residence less than 3 years before the EAC was operational.
Importantly, this is also the case for the results from the KHDS panel, which excludes the
possibility that “movers” drive the aggregate effects observable in this sample. Third, using the
KHDS panel survey once more, I specifically test for spatial sorting of high-skilled individuals
into the capital cities of Dar es Salaam and Kampala by regressing an indicator of having
moved to these cities onto personal characteristics such as education, gender and age as well
as country-time fixed effects.122 I also regress a dummy of living in these capital cities on this
set of covariates. In both cases, I find no systematic evidence of selection into economic hubs,
with only the coefficient on education marginally significant (p values of [0.097] and [0.078],
respectively) and small effect sizes in the area of 0.5 and 0.9 percentage points (likelihood
122 There are no individuals in the KHDS which have moved to Nairobi over the survey timeframe.
178
increases) for each year of schooling.123 Fourth, note that I consistently include relevant
individual-level controls such as age, gender and education which may conflate an EAC effect
with differential spatial sorting. Note also that the inclusion of which do not alter the results
dramatically (compare the baseline coefficients to panel (b) in Table C.3 through Table C.8,
respectively). In addition, as is discussed in the extensions provided in the next paragraph, I
test for the main motivation of migrating to the current place of residence for KHDS sample
respondents. The effect does not show evidence for more or less economically motivated
migration before and/or after the EAC’s re-establishment.
In sum, these results do not provide evidence in favor of a negative selection out of, and
a positive selection into, border- as well as capital city regions, respectively, which could drive
the results identified at the aggregate.124 And while none of the validity tests can completely
eradicate concerns about a potential ongoing trend, they do not undermine the findings to a
degree which would cast significant doubts about the nature of the main results, nor the validity
of the identification strategy employed.
Extensions. I provide three extensions to the existing set of results, which provide further
insights into the nature of the findings. First, given the partly significant coefficients using
border dummies (see panels (e) to (f) in Table C.6 through Table C.8), I explore a
heterogeneous effect across the different EAC border segments. I do so by replacing the
continuous measure of border distance Border [0,1] with dummies switching to 1 for households
living within 50km of border distance to the three country-border pairs from both country
directions. As such, I again estimate regression equation (4.53), albeit with separate difference-
in-differences effects for TZA-UGA {0,1}, TZA-KEN {0,1}, and KEN-UGA {0,1}, together
with Core {0,1}. Note that this also constitutes a test on a parametric specification and
identification of effects. The results are presented in Table C.17. The findings provide mixed
evidence. Concerning the Tanzania-Uganda border section, Afrobarometer as well as DHS do
not agree on the direction of the effect. Note that this particular border represents a special
case, as it hosted a large inflow of refugees from Rwanda in the decade before the EAC to the
south, i.e. in Kagera (Tanzania), which first caused a decrease in average living standards in
this region and an increase in the years thereafter. The suggested reasons for positive
turnaround in welfare over time include directly targeted programs (e.g. infrastructure) and
agglomeration externalities as given by the longer-term settlement of the refugee population
(Maystadt and Duranton 2019). In an additional robustness test (not shown here), I control
for the distance to these (former) refugee camp sites and observe that the effects for these
border dummies diminish significantly.125 Paired with the fact that the panel estimates of the
123 Results can be obtained from the author.
124 Of course, many of these tests rely on the recall of individuals, which constitutes a general weakness of
household surveys.
125 Results can be obtained from the author.
179
main results, drawn mainly from the Kagera region, provide no evidence for a differential effect
across border distance, these partly significant border dummies at this segment likely do not
represent EAC-induced effect across distance.126 Mostly negative effects on consumption
together with some income measures are observed at the Tanzania-Kenya border (columns 1,
7, 8 and 11) and mixed effects are depictable for the income components at the Kenya-Uganda
border section (columns 4, 8 and 9) as well as positive effect on population density which is,
however, less than 7% of the differential effect when compared to the core.
The second extension makes use of the opinion polling of the Afrobarometer and KHDS
surveys to evaluate (potentially altered) sentiments towards free trade and individuals’
subjective wellbeing. Table C.18 presents these results showing the component of the
Afrobarometer in panel a) and the results drawn from the KHDS survey in panel b). Concerning
subjective wellbeing, there is no uniform direction of differential effects across space. While
individuals at borders evaluate their life satisfaction as worse than 5 years prior, they assess
the general standard of living as higher.127 The rest of the variables tested are only administered
in one of the survey rounds, which is why there is no DiD estimate possible. In these cases, we
test simple differences between border- and core agglomeration regions to the rest of the
country, having arguably benefited more from trade integration. While individuals living in
border regions deem it as easier to cross international borders in order to work and trade in
foreign countries, the rest of the tested opinions show for non-significant differences in border
regions.128 Most importantly, individuals situated at borders, or in capital cities at that, do not
support the free movement of labor across countries (column 3), nor do they assess the EAC
or the African Union (AU) as trade-facilitating entities as more or less helpful to their country
(panel a), columns 5 and 6).129 And lastly, individuals in capital cities show for a higher
126 Another concern may be the sample size available for these border dummies in the Afrobarometer, for
which most of the effects appear. For instance, in the pre-EAC periods, the TZA-UGA dummy (≤ 50km)
contains only 67 individuals, 73 for the KEN-UGA segment, and 92 for TZA-KEN. In the DHS, this increases
to 200, 275 and 343, respectively.
127 Specifically, it sets a prior to compare the current living standards to the former military rule. The survey
question reads: “We are going to compare our present system of government with the former system of military
rule. Please tell me if the following things are better or worse now than they used to be. People have an
adequate standard of living.” The response values range from ‘Much worse’ (1), ‘Somewhat worse’ (2), ‘No
change’ (3), ‘Somewhat better’ (4) and ‘Much better’ (5). I remove the observations valued ‘Don’t know’.
Regarding the question in column (1), the survey question reads: “When you look at your life today, how
satisfied do you feel compared with five years ago?”. The response values range from ‘Much less satisfied’ (1),
‘Slightly less satisfied’ (2), ‘About the same’ (3), ‘Slightly more satisfied’ (4) and ‘Much more satisfied’ (5). I
remove the observations valued ‘Don’t know’. In Tanzania, this question is asking the respondents to compare
their life to one year ago.
128 The survey question reads: “In your opinion, how easy or difficult is it for people in
[West/South/East/North/Central] Africa to cross international borders in order to work or trade in other
countries, or haven’t you heard enough to say?”. The response values range from ‘Very difficult’ (1), ‘Difficult’
(2), ‘Easy’ (3), ‘Very Easy’ (4) and ‘Never try’ (7). I remove the observations valued ‘Never Try’.
129 The survey question of column (3) reads: “Which of the following statements is closest to your view? Choose
Statement 1 or Statement 2. Statement 1: People living in [West/South/East/North/Central] Africa should
be able to move freely across international borders in order to trade or work in other countries. Statement 2:
Because foreign migrants take away jobs, and foreign traders sell their goods at very cheap prices, governments
should protect their own citizens and limit the cross-border movement of people and goods.” The response
180
tendency to dislike immigrants or foreign workers as neighbors.130 This may be evidence for
increased experience with competition on the labor market, but given the general insignificance
of the other results, this result is not further interpreted. Moving to the results on simple
differences in the Kagera Health and Development Survey, we see some indication of what is
found in the main results.131 That is, individuals in border regions have a lower likelihood of
having had paid (formal) employment directly after migrating there. Going from the median
border distance in the sample (103 km) to respondents directly at the border decreases the
likelihood of paid (formal) employment after the move by 5.6 (4.1) percentage points. The rest
of the results indicating higher or lower subjective household life satisfaction or wealth are
generally insignificant with the exception of a positive effect on the current assessment of the
households as rich rather than poor in core agglomerations.
The third and last extension tries to shed some light on the non-differential effects of the
trade policy across border distance. Without reference to NEG theory, studies have shown that
following trade, the pass-through of price changes following trade liberalization decays strongly
in border distance (see e.g. Nicita 2009 for the case of Mexico in NAFTA). However, there is
also evidence that borders in Africa are particularly “thick”, and that trade agreements may
not result in improvements from such deviations in the law of one price, depending on local
characteristics such as differences in ethnic makeup, language, or informal credit practices
across borders (see e.g. Versailles 2012; Aker et al. 2014; Brenton et al. 2014). To test for a
distance penalty in price pass-through, I make use of the price questionnaire in the Kagera
Health and Development Survey of rounds 1 and 2 which provide market prices of various food
items. The prices are expressed per unit (e.g. per kilogram) and averaged across rainy and dry
seasons to control for periodic fluctuations. To test for the border pass-through, I estimate
regression equation (4.53), and assess the logged price of four homogenous, heavily consumed
and, of course, traded goods across the three countries and thereby check for nominal price
differences before and after the establishment of the EAC with regards to border distance, and
as usual, core agglomerations. The results are presented in Appendix Table C.19. As is seen,
there are almost no differential effects of price changes across border distance. Only the price
for Millet (a cereal grain), decreased slightly more at survey locales closer to borders. Finally,
we see the now common effect for core agglomerations, which show statistically significant price
values range from ‘Agree very strongly with Statement 1’ (1), ‘Agree with Statement 1’ (2), ‘Agree with
Statement 2’ (3), ‘Agree very strongly with Statement 2’ (4), ‘Agree with Neither’ (5) and ‘Don’t know’ (7).
I recode 5 to represent the median value. I remove the observations valued ‘Don’t know’.
The survey question on the variable in column (4) and (5) reads: In your opinion, how much do each of the
following do to help your country, or haven’t you heard enough to say? [EAC/African Union]. The response
values range from ‘Don’t help’ (0), ‘Help a little’ (1), ‘Help somewhat’ (2), ‘Help a lot’ (3) and ‘Don’t know’
(7). I remove the observations valued ‘Don’t know’.
130 The survey question reads: “For each of the following types of people, please tell me whether you would
like having people from this group as neighbors, dislike it, or not care: Immigrants or foreign workers.”
‘Strongly dislike’ (1), ‘Somewhat dislike’ (2), ‘Would not care’ (3) and ‘Somewhat like’ (4), ‘Strongly like’ (5)
and ‘Don’t know’ (9). I remove the observations valued ‘Don’t know’.
131 There are weak findings for an increased likelihood of any type of activity besides housework.
181
decreases for all items tested. The relative size of the effects is in the range of 2-8%. Why there
is no positive price change at borders is up for debate. What can be said from previous literature
(see above) is that integration of markets across (East) Africa, is that borders remain a
hinderance.
Reconciliation with Theory. As a final note on the empirical results observed, I attempt to
reconcile them with the (contrary) predictions of the theoretical simulations in Section 4.4.
Importantly, while I provide some deliberations on the possible (parameter) configurations that
may bring about such results, they remain speculative. The reduced-form empirics employed
in this chapter are not apt to inform on specific parameter coefficients which would permit
isolating the key factor driving these observed results. Nonetheless, keeping in mind the
theoretical predictions, the results shown in this section hint towards three broad possibilities.132
The first trivial one is that the cross-border activity triggered by regional market integration
of the EAC did not induce a large enough (market access) shock to render proximity to new
markets as relevant enough to break a general agglomeration tendency towards interior hubs.
Second, the interior economic hubs display too strong of an attraction that even in the event
of a fully integrated market, agglomeration in Nairobi, Dar Es Salaam and Kampala is
sustained. This latter scenario can be intuitively displayed by referring to Panel D in Figure
4.3, where a large enough share of either previous manufacturing distribution or population
inflows renders a fully agglomerated equilibrium of interiors possible even for the scenario of
full trade liberalization, i.e. a 𝑇𝑇𝐹𝐹= 𝑇𝑇𝐷𝐷= 1.60. Depending on the degree of integration, the
necessary share of manufacturing positioned in hubs prior to liberalization setting in motion
such an equilibrium has a minimum of 50% and a maximum of 73% in the full free trade
scenario. Importantly, this is the case only for the foreign market which is equally balanced.
As is established in Panel D of Figure 4.5, this necessary labor share is decreasing in foreign
spatial inequality, where 78% depicts the scenario when all foreign activity is situated at the
border and less than 68% when the foreign economy is fully agglomerated in the interior.133 As
such, given the significant polarization of all member countries before the establishment of the
EAC (see Section 4.3) one can imagine a scenario in which liberalization does not suffice to
break agglomeration, because countries open up to similarly spatially unequal foreign markets.
And third, the dynamics indicated by the dwindling differences of AFB and DHS
consumption and agglomeration measures over time displayed throughout the results,
particularly in Table 4.2, Table C.4, Table C.16 and Table C.21 may also point towards a
transitory shift to a new welfare equalizing equilibrium. The mechanics can be displayed in
Panel B of Figure 4.3 and are as follows. With a starting point of equal (manufacturing) labor
distribution, trade liberalization sets in, which lowers the real wage differential of core
132 Of course, only if the assumptions of the difference-in-differences employed hold. These assumptions have
been tested in Subsection 4.7.2.
133 For results in all three trade scenarios, see Figure C.4.
182
agglomerations. Labor immediately responds with outflow as seen in column (3) of Table 4.2.
However, these outflows are too large such that they quickly offshoot the long-run stable
equilibria between 40% and 50% and render real wages in interior hubs as exceeding those at
borders. The subsequent response is a migration inflow back into core agglomerations until
welfare equalization at a new (lower) equilibrium share of labor is reached. Three factors go
against this possibility, however: First, an equal or less than equal spatial distribution of labor
and welfare in favor of region 2 was not a likely initial position of all three countries before the
EAC. Second, population inflows into border regions are not observed in Table 4.2. And third,
most importantly, while differential changes in welfare and population decrease across time,
they do remain positive in all periods post-liberalization, which is only possible in a move
towards a fully agglomerated equilibrium. One last possibility, of course, is that there is no
practical spatial heterogeneity, i.e. border regions do not have better access to the new
markets.134 In such a scenario, depicted in Panel A and C in Figure 4.3, the result is fully robust
with trade setting in motion full agglomeration, as depicted by the rotating lines in all
parameter configurations. Agglomeration is more likely in the regions which hosted the majority
of workers in pre-EAC years. Note that there is yet another scenario which may generate the
observed results. That is, if market integration among the three partner countries increased
trade diversion, i.e. trade with the rest of the world, instead of intra-regional trade, one may
observe clustering in the interior hubs, which also host the significant links to these
international markets (Marchand 2012). The increasing agglomeration after the customs union
protocol may be a hint at this. However, this is only vague evidence, and needs to be thoroughly
investigated against the specific external tariff structure of the EAC. Exploring this is beyond
the scope of the present paper.
4.8 Conclusion
This paper investigates the impact of the re-establishment of the East African Community
(EAC) on household welfare using three distinct sets of longitudinal, geo-referenced household-
level surveys from the three founding members Kenya, Tanzania and Uganda. I formally derive
the potential impact of the EAC on households from a canonical New Economic Geography
(NEG) model with heterogeneous intra-national space and test the predictions through a
reduced-form difference-in-differences specification with treatment intensity given by
households’ road distance to internal EAC border crossings. I therefore treat the re-
establishment of the EAC in 2001 – and the expansion to a customs union and common market
in 2005 and 2010, respectively – as a regional policy intervention having differential effects on
individual households governed by their geo-spatial location within the countries.
The results show that households and individuals living closer to the internal EAC border
did not experience a relative increase in welfare following the re-establishment, as measured by
134 One may think of shifting transport types i.e. using airports or harbors rather than roads.
183
an array of consumption indices as well as intensive and extensive employment outcomes.
Rather, the results hint at the strengthened concentration of activity in the pre-existing,
interior economic hubs as evinced by the strong differential and economically relevant increases
in household welfare across the measured dimensions as well as subsequent inflows of
population. As such, these findings go against the general prediction of the theoretical
simulations and also against the hypothesis prominently outlined in Krugman and Elizondo
(1996), which predicted a dispersion of the concentrated economic activity of developing
countries following liberalization. My results are also in contrast to other recent empirical
findings, which have regularly documented regions closer to the new market (potential) to profit
from the less costly access to them (for an overview see Brülhart 2011). Most notably, the
present results are at odds with Eberhard-Ruiz and Moradi (2019) who provide (remote
sensing) evidence that the EAC’s re-establishment led to a positive (one-off) effect on city
growth for cities closer to the internal EAC borders.
While in disagreement with the general prediction of the model and these previous
contributions, the model allows a potential insight into the reasons why such results may come
about. The theoretical results show that the spatial configuration of the foreign economy has
a non-negligible moderating force on the strength of dispersion induced by trade liberalization
in the domestic country. As such, the presence of economically dominating interiors in all
partner countries potentially weakens the dispersion force of market integration to the point
where a regional pattern outlined by stark inequalities as existent prior to the EAC remains a
possible outcome even after regional trade liberalization. Such explanations remain contentious,
however, as the reduced-form empirics are ultimately not apt to inform on specific parameter
coefficients, which would permit isolating key factors driving these observed effects.
One last, potentially policy-relevant contribution of the present paper is the connection
to the widely discussed issue of “urbanization without growth” (Fay and Opal 2000). Given
that the results show for increased consumption and agglomeration tendency in primate cities,
without a concomitant expansion of labor market opportunities, they provide evidence in favor
of developing countries’ metropoles growing to “consumption-” rather than “production cities”
(Gollin et al. 2016), and that this process may be influenced by regional market integration.
184
4.9 References
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2001. ‘The Colonial Origins of
Comparative Development: An Empirical Investigation’. American Economic Review
91 (5): 1369–1401.
Afrobarometer, [Kenya, Tanzania, Uganda] [Round 1, 2, 3, 4, 5, 6, 7]. 2019. ‘Afrobarometer
Data’.
AidData. 2017. ‘Geocoding Methodology, Version 2.0.1.’ Williamsburg, VA: AidData at
William & Mary.: AidData Research and Evaluation Unit.
Aker, Jenny C., Michael W. Klein, Stephen A. O’Connell, and Muzhe Yang. 2014. ‘Borders,
Ethnicity and Trade’. Journal of Development Economics 107 (March): 1–16.
Andreoni, Antonio, Julian Boys, and Ole Therkildsen. 2022. ‘The Political Economy of
“Specialism” in Tanzania: How to Make Export Processing Zones Work via Conditional
Special Licensing’. In . Anti-Corruption Evidence Working Paper 046. London, UK.:
SOAS University of London. Anti-Corruption Evidence (ACE) Research Consortium.
Atkin, David, and Dave Donaldson. 2015. ‘Who’s Getting Globalized? The Size and
Implications of Intra-National Trade Costs’. w21439. Cambridge, MA: National Bureau
of Economic Research.
Barros, Laura, and Inmaculada Martínez-Zarzoso. 2022. ‘Systematic Literature Review on
Trade Liberalization and Sustainable Development’. Sustainable Production and
Consumption 33 (September): 921–31.
Bartik, Timothy J. 1991. Who Benefits from State and Local Economic Development Policies?
W.E. Upjohn Institute.
Beegle, Kathleen, Joachim De Weerdt, and Stefan Dercon. 2006. ‘Kagera Health and
Development Survey 2004 Basic Information Document’. The World Bank,
Washington, DC.
Behrens, Kristian, Carl Gaigné, Gianmarco I P Ottaviano, and Jacques-François Thisse. 2003.
‘Inter-Regional and International Trade: Seventy Years After Ohlin’. In CEPR
Discussion Paper No. 4065. Paris & London: CEPR Press.
Behrens, Kristian, Carl Gaigné, Gianmarco I.P. Ottaviano, and Jacques-François Thisse. 2006.
‘Is Remoteness a Locational Disadvantage?’ Journal of Economic Geography 6 (3):
347–68.
———. 2007. ‘Countries, Regions and Trade: On the Welfare Impacts of Economic
Integration’. European Economic Review 51 (5): 1277–1301.
Bickenbach, Frank, Eckhardt Bode, Peter Nunnenkamp, and Mareike Söder. 2016. ‘Night
Lights and Regional GDP’. Review of World Economics 152 (2): 425–47.
Blanchard, O., and Lawrence F. Katz. 1992. ‘Regional Evolutions’. Brooking Papers on
Economic Activity 23 (1): 1–76.
Bluhm, Richard, and Gordon C. McCord. 2022. ‘What Can We Learn from Nighttime Lights
for Small Geographies? Measurement Errors and Heterogeneous Elasticities’. Remote
Sensing 14 (5): 1190.
Borraz, Fernando, Daniel Ferrés, and Máximo Rossi. 2013. ‘Assessment of the Distributive
Impact of National Trade Reforms in Brazil’. The Journal of Economic Inequality 11
(2): 215–35.
Brakman, Steven, Harry Garretsen, and Charles van Marrewijk. 2019. ‘New Economic
Geography After Thirty Years’. In Handbook of Regional Science, edited by Manfred
M. Fischer and Peter Nijkamp, 1–23. Berlin, Heidelberg: Springer Berlin Heidelberg.
Brakman, Steven, Harry Garretsen, and Charles van Marrewijk. 2020. An Introduction to
Geographical and Urban Economics: A Spiky World. Third edition. Cambridge, United
Kingdom: Cambridge University Press.
185
Breinlich, Holger, Gianmarco I.P. Ottaviano, and Jonathan R.W. Temple. 2014. ‘Regional
Growth and Regional Decline’. In Handbook of Economic Growth, 2:683–779.
Amsterdam.
Brenton, Paul, Alberto Portugal-Perez, and Julie Régolo. 2014. ‘Food Prices, Road
Infrastructure, and Market Integration in Central and Eastern Africa’. World Bank
Policy Research Working Paper 7003, Policy Research Working Papers, , August.
Brülhart, Marius. 2011. ‘The Spatial Effects of Trade Openness: A Survey’. Review of World
Economics 147 (1): 59–83.
Brülhart, Marius, Olivier Cadot, and Alexander Himbert. 2017. ‘Trade Integration and
Spatially Balanced Development: Implications for Uganda and Rwanda’. International
Growth Centre Working Paper, no. F-43406-CCP-1.
———. 2019. ‘Let There Be Light: Trade and the Development of Border Regions’. FERDI
Working Paper, No. P251.
Brülhart, Marius, Céline Carrère, and Federico Trionfetti. 2012. ‘How Wages and Employment
Adjust to Trade Liberalization: Quasi-Experimental Evidence from Austria’. Journal
of International Economics 86 (1): 68–81.
Brülhart, Marius, Matthieu Crozet, and Pamina Koenig. 2004. ‘Enlargement and the EU
Periphery: The Impact of Changing Market Potential’. The World Economy 27 (6):
853–75.
Cadot, Olivier, Alexander Himbert, and Marie-Agnès Jouanjean. 2015. ‘Trade Facilitation and
Concentration: Evidence from Sub-Saharan Africa’. Shaping Policy for Development -
ODI Research Reports.
Cali, Massimiliano 2014. ‘Trade Boom and Wage Inequality: Evidence from Ugandan Districts’.
Journal of Economic Geography 14 (6): 1141–74.
Castilho, Marta, Marta Menéndez, and Aude Sztulman. 2012. ‘Trade Liberalization, Inequality,
and Poverty in Brazilian States’. World Development 40 (4): 821–35.
CIESIN. 2017. ‘Gridded Population of the World, Version 4 (GPWv4): Population Density,
Revision 11’. Center For International Earth Science Information Network-CIESIN-
Columbia University. Socioeconomic Data and Applications Center (SEDAC).
Conley, T.G. 1999. ‘GMM Estimation with Cross Sectional Dependence’. Journal of
Econometrics 92 (1): 1–45.
Conley, Timothy G. 2010. ‘Spatial Econometrics’. In Microeconometrics, edited by Steven N.
Durlauf and Lawrence E. Blume, 303–13. London: Palgrave Macmillan UK.
Crozet, Matthieu, and Pamina Koenig. 2004a. ‘Trade Liberalization and the Internal
Geography of Countries’. In Multinational Firms’ Location and the New Economic
Geography, by Jean-Louis Mucchielli and Thierry Mayer, 3311. Edward Elgar
Publishing.
———. 2004b. ‘EU Enlargement and the Internal Geography of Countries’. Journal of
Comparative Economics 32 (2): 265–79.
De Weerdt, Joachim, Kathleen Beegle, Helene B. Lilleor, Stefan Dercon, Kalle Hirvonen,
Martina Kirchberger, and Sofya Krutilov. 2010. ‘Kagera Health and Development
Survey 2010: Basic Information Document. Study Paper 46’. Copenhagen: Rockwool
Foundation Working.
Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach to
Development Policy. Baltimore, MD: Published for the World Bank [by] Johns Hopkins
University Press.
Dixit, Avinash K., and Joseph E. Stiglitz. 1977. ‘Monopolistic Competition and Optimum
Product Diversity’. The American Economic Review 67 (3): 297–308.
186
EAC. 1999. ‘Treaty for the Establishment of the East African Community’. African Yearbook
of International Law Online / Annuaire Africain de Droit International Online 7 (1):
419–509.
———. 2015. ‘Report of the Committee on Communication, Trade and Investments for the
on-Spot Assessment on the EAC One Stop Border Posts (OSBPs) in EAC Partner
States’. Arusha, Tanzania: East African Legislative Assembly.
Eberhard-Ruiz, Andreas, and Alexander Moradi. 2019. ‘Regional Market Integration in East
Africa: Local but No Regional Effects?’ Journal of Development Economics 140
(September): 255–68.
Edmonds, Eric V., Nina Pavcnik, and Petia Topalova. 2010. ‘Trade Adjustment and Human
Capital Investments: Evidence from Indian Tariff Reform’. American Economic
Journal: Applied Economics 2 (4): 42–75.
Emran, M. Shahe, and Zhaoyang Hou. 2013. ‘Access to Markets and Rural Poverty: Evidence
from Household Consumption in China’. Review of Economics and Statistics 95 (2):
682–97.
Engel, Jakob, Deeksha Kokas, Gladys López Acevedo, and Maryla Maliszewska. 2021. The
Distributional Impacts of Trade: Empirical Innovations, Analytical Tools, and Policy
Responses. Trade and Development. Washington, DC: World Bank.
Erten, Bilge, Jessica Leight, and Fiona Tregenna. 2019. ‘Trade Liberalization and Local Labor
Market Adjustment in South Africa’. Journal of International Economics 118 (May):
448–67.
European Commission. 2019. Description of the GHS Urban Centre Database 2015: Public
Release 2019 : Version 1.0. LU: European Commission Joint Research Centre.
Publications Office.
Ezcurra, Roberto, and Alba Del Villar. 2021. ‘Globalization and Spatial Inequality: Does
Economic Integration Affect Regional Disparities?’ The Annals of Regional Science 67
(2): 335–58.
Ezcurra, Roberto, and Andrés Rodríguez-Pose. 2014. ‘Trade Openness and Spatial Inequality
in Emerging Countries’. Spatial Economic Analysis 9 (2): 162–82.
Farr, Tom G., Paul A. Rosen, Edward Caro, Robert Crippen, Riley Duren, Scott Hensley,
Michael Kobrick, et al. 2007. ‘The Shuttle Radar Topography Mission’. Reviews of
Geophysics 45 (2): RG2004.
Fay, Marianne, and Charlotte Opal. 2000. Urbanization without Growth: A Not so Uncommon
Phenomenon. Vol. 2412. World Bank Publications.
Feyrer, James. 2019. ‘Trade and Income—Exploiting Time Series in Geography’. American
Economic Journal: Applied Economics 11 (4): 1–35.
Fick, Stephen E., and Robert J. Hijmans. 2017. ‘WorldClim 2: New 1‐km Spatial Resolution
Climate Surfaces for Global Land Areas’. International Journal of Climatology 37 (12):
4302–15.
Frankel, Jeffrey A., and David Romer. 1999. ‘Does Trade Cause Growth?’ American Economic
Review 89 (3): 379–99.
Fujita, Masahisa, Paul R. Krugman, and Anthony J. Venables. 2001. The Spatial Economy:
Cities, Regions, and International Trade. 1. paperback ed. Cambridge, Mass.: MIT
Press.
Fujita, Masahisa, and Tomoya Mori. 2005. ‘Frontiers of the New Economic Geography*’.
Papers in Regional Science 84 (3): 377–405.
Fukase, Emiko. 2013. ‘Export Liberalization, Job Creation, and the Skill Premium: Evidence
from the US–Vietnam Bilateral Trade Agreement (BTA)’. World Development 41
(January): 317–37.
187
GADM. 2020. ‘Global Administrative Areas [Digital Geospatial Data]’. Center for Spatial
Sciences at the University of California, Davis.
Gallup, John Luke, Jeffrey D. Sachs, and Andrew D. Mellinger. 1999. ‘Geography and
Economic Development’. International Regional Science Review 22 (2): 179–232.
Gibson, John, Susan Olivia, and Geua Boe‐Gibson. 2020. ‘NIGHT LIGHTS IN ECONOMICS:
SOURCES AND USES 1’. Journal of Economic Surveys 34 (5): 955–80.
Gibson, John, Susan Olivia, Geua Boe-Gibson, and Chao Li. 2021. ‘Which Night Lights Data
Should We Use in Economics, and Where?’ Journal of Development Economics 149
(March): 102602.
Giovannetti, Giorgia, Enrico Marvasi, and Arianna Vivoli. 2021. ‘The Asymmetric Effects of
20 Years of Tariff Reforms on Egyptian Workers’. Economia Politica 38 (1): 89–130.
Giovannetti, Giorgia, Marco Sanfilippo, and Arianna Vivoli. 2022. ‘Import Tariff
Liberalization, Employment, and Gender in Ethiopia’. The Journal of Development
Studies 58 (11): 2366–92.
Glaeser, Edward L. 2014. ‘A World of Cities: The Causes and Consequences of Urbanization
in Poorer Countries’. Journal of the European Economic Association 12 (5): 1154–99.
Goldberg, Pinelopi Koujianou, and Nina Pavcnik. 2007. ‘Distributional Effects of Globalization
in Developing Countries’. Journal of Economic Literature, 75.
Gollin, Douglas, Remi Jedwab, and Dietrich Vollrath. 2016. ‘Urbanization with and without
Industrialization’. Journal of Economic Growth 21 (1): 35–70.
Hanson, Gordon. 1994. ‘Localization Economies, Vertical Organization and Trade’. American
Economic Review 86 (5).
Hanson, Gordon. 1997. ‘Increasing Returns, Trade and the Regional Structure of Wages’. The
Economic Journal 107 (440): 113–33.
Hazlewood, Arthur. 1979. ‘THE END OF THE EAST AFRICAN COMMUNITY: WHAT
ARE THE LESSONS FOR REGIONAL INTEGRATION SCHEMES ?’ JCMS:
Journal of Common Market Studies 18 (1): 40–58.
Helpman, Elhanan. 1998. ‘The Size of Regions’. Topics in Public Economics: Theoretical and
Applied Analysis, 33–54.
Henderson, J. Vernon. 1974. ‘The Sizes and Types of Cities’. The American Economic Review
64 (4): 640–56.
———. 1982. ‘Systems of Cities in Closed and Open Economies’. Regional Science and Urban
Economics 12 (3): 325–50.
Henderson, Vernon J., Tim Squires, Adam Storeygard, and David Weil. 2018. ‘The Global
Distribution of Economic Activity: Nature, History, and the Role of Trade’. The
Quarterly Journal of Economics 133 (1): 357–406.
Hirte, Georg, Christian Lessmann, and André Seidel. 2020. ‘International Trade, Geographic
Heterogeneity and Interregional Inequality’. European Economic Review 127 (August):
103427.
IMF. 2023. ‘Direction of Trade Statistics (DOTS)’. 2023.
Jedwab, Remi, and Dietrich Vollrath. 2015. ‘Urbanization without Growth in Historical
Perspective’. Explorations in Economic History 58 (October): 1–21.
Kaahwa, Wilbert T. K. 2003. East African Community: EAC Treaty and Challenges to the
Community. Occasional Papers (East African Community Secretariat), no. 3. Arusha,
Tanzania: East African Community Secretariat in conjunction with the German Agency
for Technical Co-operation.
Kanbur, Ravi, and Xiaobo Zhang. 2005. ‘Fifty Years of Regional Inequality in China: A
Journey Through Central Planning, Reform, and Openness: FIFTY YEARS OF
REGIONAL INEQUALITY IN CHINA’. Review of Development Economics 9 (1): 87–
106.
188
Kenya Central Bureau of Statistics. 2001. ‘Statistical Abstract: 2001’. Nairobi, Kenya.: Kenya
Central Bureau of Statistics, Ministry of Finance and Planning.
———. 2003. ‘Statistical Abstract: 2003’. Nairobi, Kenya.: Kenya Central Bureau of Statistics,
Ministry of Finance and Planning.
Kenya National Bureau of Statistics. 2011. ‘Statistical Abstract: 2011’. Nairobi, Kenya.: Kenya
National Bureau of Statistics.
———. 2013. ‘Basic Report on The 2010 Census of Industrial Production (CIP)’. Nairobi,
Kenya.: Kenya National Bureau of Statistics, Ministry of Industrialization.
Kinyondo, Abel, Carol Newman, Finn Tarp, and UNU-WIDER. 2016. The Role and
Effectiveness of Special Economic Zones in Tanzania. 122nd ed. Vol. 2016. WIDER
Working Paper. UNU-WIDER.
Kis-Katos, Krisztina, and Robert Sparrow. 2015. ‘Poverty, Labor Markets and Trade
Liberalization in Indonesia’. Journal of Development Economics 117 (November): 94–
106.
Kremer, Michael. 1993. ‘The O-Ring Theory of Economic Development’. The Quarterly Journal
of Economics 108 (3): 551–75.
Krugman, Paul, and Anthony. J. Venables. 1995. ‘Globalization and the Inequality of Nations’.
The Quarterly Journal of Economics 110 (4): 857–80.
Krugman, Paul. 1980. ‘Scale Economies, Product Differentiation, and the Pattern of Trade’.
The American Economic Review 70 (5): 950–59.
———. 1991. ‘Increasing Returns and Economic Geography’. Journal of Political Economy 99
(3): 483–99.
Krugman, Paul, and Raul Livas Elizondo. 1996. ‘Trade Policy and the Third World
Metropolis’. Journal of Development Economics 49: 137–50.
Krugman, Paul R. 1979. ‘Increasing Returns, Monopolistic Competition, and International
Trade’. Journal of International Economics 9 (4): 469–79.
Lehner, Bernhard, and Petra Döll. 2004. ‘Development and Validation of a Global Database
of Lakes, Reservoirs and Wetlands’. Journal of Hydrology 296 (1–4): 1–22.
Mansori, Kashif S. 2003. ‘The Geographic Effects of Trade Liberalization with Increasing
Returns in Transportation’. Journal of Regional Science 43 (2): 249–68.
Marchand, Beyza Ural. 2012. ‘Tariff Pass-through and the Distributional Effects of Trade
Liberalization’. Journal of Development Economics 99 (2): 265–81.
———. 2019. ‘Inequality and Trade Policy: The Pro‐Poor Bias of Contemporary Trade
Restrictions’. Review of Income and Wealth 65 (S1). Martínez-Zarzoso, Inma, Leila
Baghdadi, and Hendrik Wiard Kruse. 2016. ‘Is Tunisian Trade Policy Pro-Poor?’ In
1033.
Maystadt, Jean-François, and Gilles Duranton. 2019. ‘The Development Push of Refugees:
Evidence from Tanzania’. Journal of Economic Geography 19 (2): 299–334.
McCaig, Brian. 2011. ‘Exporting out of Poverty: Provincial Poverty in Vietnam and U.S.
Market Access’. Journal of International Economics 85 (1): 102–13.
McCaig, Brian, and Margaret S. McMillan. 2020. ‘Trade Liberalisation and Labour Market
Adjustment in Botswana’. Journal of African Economies 29 (3): 236–70.
Monfort, Philippe, and Rosella Nicolini. 2000. ‘Regional Convergence and International
Integration’. Journal of Urban Economics 48 (2): 286–306.
Monfort, Philippe, and Tanguy van Ypersele. 2003. ‘Integration, Regional Agglomeration and
International Trade’. In CEPR Discussion Paper No. 3752, 32. Paris & London: CEPR
Press.
Mugomba, Agrippah T. 1978. ‘Regional Organisations and African Underdevelopment: The
Collapse of the East African Community’. The Journal of Modern African Studies 16
(2): 261–72.
189
Nathan Associates. 2011. ‘Corridor Diagnostic Study of the Northern and Central Corridors of
East Africa’. In Volume 1: Main Report. Alexandria, VA, USA.
Natural Earth. 2019. ‘Natural Earth. Free Vector and Raster Map Data.’
NGA, Natioal Geospace-Intelligence Agency. 2019. ‘World Port Index 27th Edition’. United
States Government.
Nicita, Alessandro. 2009. ‘The Price Effect of Tariff Liberalization: Measuring the Impact on
Household Welfare’. Journal of Development Economics 89 (1): 19–27.
Nicita, Alessandro, Marcelo Olarreaga, and Guido Porto. 2014. ‘Pro-Poor Trade Policy in Sub-
Saharan Africa’. Journal of International Economics 92 (2): 252–65.
Nunn, Nathan, and Diego Puga. 2012. ‘Ruggedness: The Blessing of Bad Geography in Africa’.
Review of Economics and Statistics 94 (1): 20–36.
OSM. 2022. OpenStreetMap Database. OpenStreetMap contributors. Cambridge, United
Kingdom: OpenStreetMap Foundation.
Ottaviano, Gianmarco, Takatoshi Tabuchi, and Jacques-Francois Thisse. 2002. ‘Agglomeration
and Trade Revisited’. International Economic Review 43 (2): 409–35.
Paluzie, Elisenda. 2001. ‘Trade Policy and Regional Inequalities*’. Papers in Regional Science
80 (1): 67–85.
Pavcnik, Nina. 2017. ‘The Impact of Trade on Inequality in Developing Countries’. w23878.
Cambridge, MA: National Bureau of Economic Research.
Porto, Guido G. 2006. ‘Using Survey Data to Assess the Distributional Effects of Trade Policy’.
Journal of International Economics 70 (1): 140–60.
Puga, Diego. 1999. ‘The Rise and Fall of Regional Inequalities’. European Economic Review,
32.
Ramankutty, Navin, Jonathan A. Foley, John Norman, and Kevin McSweeney. 2002. ‘The
Global Distribution of Cultivable Lands: Current Patterns and Sensitivity to Possible
Climate Change: Cropland Sensitivity to Climate Change’. Global Ecology and
Biogeography 11 (5): 377–92.
Rappaport, Jordan, and Jeffrey D. Sachs. 2003. ‘The United States as a Coastal Nation’.
Journal of Economic Growth 8: 5–46.
Redding, Stephen J. 2022. ‘Trade and Geography’. In Handbook of International Economics,
5:147–217. Elsevier.
Redding, Stephen J., and Esteban Rossi-Hansberg. 2017. ‘Quantitative Spatial Economics’.
Annual Review of Economics 9 (1): 21–58.
Redding, Stephen J, and Daniel M Sturm. 2008. ‘The Costs of Remoteness: Evidence from
German Division and Reunification’. American Economic Review 98 (5): 1766–97.
Rodrik, Dani. 2002. ‘Institutions, Integration, and Geography: In Search of the Deep
Determinants of Economic Growth’. In In Search of Prosperity: Analytical Country
Studies on Growth, Princeton, NJ. Princeton University Press.
Rutstein, Shea O., and Kiersten Johnson. 2004. ‘The DHS Wealth Index’. In DHS Comparative
Reports No. 6. Calverton, Maryland, USA: Macro International.
Rutstein, Shea O., and Sarah Staveteig. 2014. ‘Making the Demographic and Health Surveys
Wealth Index Comparable’. ICF International Rockville, Maryland, USA.
Sachs, Jeffrey D. 2003. ‘Institutions Don’t Rule: Direct Effects of Geography on Per Capita
Income’. w9490. Cambridge, MA: National Bureau of Economic Research.
Sachs, Sonia Ehrlich, Andrew Mellinger, Pia Malaney, Jeffrey Sachs, Anthony Kiszewski, and
Andrew Spielman. 2004. ‘A Global Index Representing the Stability of Malaria
Transmission’. The American Journal of Tropical Medicine and Hygiene 70 (5): 486–
98.
Samuelson, Paul A. 1952. ‘The Transfer Problem and Transport Costs: The Terms of Trade
When Impediments Are Absent’. The Economic Journal 62 (246): 278.
190
Schiff, Maurice W., and L. Alan Winters. 2003. Regional Integration and Development.
Washington, DC : [s.l.]: World Bank ; Oxford University Press.
Smits, Jeroen, and Roel Steendijk. 2015. ‘The International Wealth Index (IWI)’. Social
Indicators Research 122 (1): 65–85.
Tanzania National Bureau of Statistics. 2004. ‘Employment and Earnings Survey: 2001
Analytical Report’. Dar es Salaam, Tanzania: National Bureau of Statistics, Ministry
of Planning, Economy and Empowerment. The United Republic of Tanzania.
———. 2006. ‘Population Census: 2002 Analytical Report’. Dar es Salaam, Tanzania: National
Bureau of Statistics, Ministry of Planning, Economy and Empowerment. The United
Republic of Tanzania.
———. 2007. ‘Employment and Earnings Survey: 2002 Analytical Report’. Dar es Salaam,
Tanzania: National Bureau of Statistics, Ministry of Planning, Economy and
Empowerment. The United Republic of Tanzania.
———. 2010. ‘Annual Survey of Industrial Production and Performance: 2008 Statistical
Report’. Dar es Salaam, Tanzania: National Bureau of Statistics, Ministry of Industry,
Trade and Marketing. The United Republic of Tanzania.
———. 2018. ‘Annual Survey of Industrial Production: 2016 Statistical Report’. Dar es Salaam,
Tanzania: National Bureau of Statistics, Ministry of Industry, Trade and Investment.
The United Republic of Tanzania.
Thünen, Johann H. von. 1826. Der Isolierte Staat in Beziehung Auf Landwirtschaft Und
Nationalökonomie. Hamburg: Perthes.
Topalova, Petia. 2010. ‘Factor Immobility and Regional Impacts of Trade Liberalization:
Evidence on Poverty from India’. American Economic Journal: Applied Economics 2
(4): 1–41.
UFZA. 2022. ‘Uganda Free Zone Authority Annual Report 2021/2022’. Kampala, Uganda.
Uganda Bureau of Statistics. 2003. ‘Uganda Business Register: 2001/2002, Report’. Kampala,
Uganda: Uganda Bureau of Statistics, Directorate of Business Industry, Agriculture and
Energy Statistics.
———. 2006. ‘Uganda Business Inquiry: 2000/2001’. Entebbe, Uganda: Uganda Bureau of
Statistics.
———. 2007. ‘Uganda Business Register: 2006/2007, Report’. Kampala, Uganda: Uganda
Bureau of Statistics, Directorate of Business Industry, Agriculture and Energy
Statistics.
———. 2012. ‘Statistical Abstract: 2012’. Kampala, Uganda: Uganda Bureau of Statistics.
———. 2016. ‘The National Population and Housing Census 2014 – Main Report’. Kampala,
Uganda: Uganda Bureau of Statistics.
UNCTAD. 2022. ‘Trade Analysis and Information System (TRAINS)’. United Nations
Conference on Trade and Development.
United Nations Development Programme. 2011. ‘Regional Integration and Human
Development: A Pathway for Africa’. New York, NY: United Nations Development
Programme.
UNSD. 2022. ‘International Merchandise Trade Statistics (COMTRADE)’. United Nations
Statistics Division.
UNU-CRIS. 2019. ‘Regional Integration Knowledge System (RIKS)’.
Versailles, Bruno. 2012. ‘Market Integration and Border Effects in Eastern Africa’, 48.
Vo, Thang T., and Dinh X. Nguyen. 2020. ‘Impact of Trade Liberalization on Household
Welfare: An Analysis Using Household Exposure-to-Trade Indices’. Social Indicators
Research, September.
191
Wild, Frederik, and David Stadelmann. 2022. ‘Coastal Proximity and Individual Living
Standards: Econometric Evidence from Georeferenced Household Surveys in sub‐
Saharan Africa’. Review of Development Economics 26 (4): 1883–1901.
Winters, L. Alan. 2002. ‘Trade Liberalisation and Poverty: What Are the Links?’, 29.
Winters, L. Alan, and Antonio Martuscelli. 2014. ‘Trade Liberalization and Poverty: What
Have We Learned in a Decade?’ Annual Review of Resource Economics 6 (1): 493–512.
Winters, L. Alan, Neil Mcculloch, and Andrew Mckay. 2004. ‘Trade Liberalization and
Poverty: The Evidence So Far’. Journal of Economic Literature, 64.
World Bank. 2009. World Development Report 2009: Reshaping Economic Geography. 2009.
Washington, D.C.: World Bank.
———. 2016. ‘Implementation Completion and Results Report (IDA-39300 IDA-45710) for the
Northern
Corridor
Transport
Improvement
Project’.
Report
ICR00002771.
Washington, D.C: The World Bank.
———. 2020. The African Continental Free Trade Area: Economic and Distributional Effects.
Washington, DC: World Bank.
———. 2022. ‘Data.Worldbank.Org’. 2022.
World Bank, and University of Dar es Salaam. 1994. Kagera Health and Development Survey
1991-1994 (Wave 1 to 4 Panel). Ref. TZA_1991_KHDS_v01_M. Dataset Downloaded
from the World Bank Microdata Library (Www.Microdata.Worldbank.Org). World
Bank and University of Dar es Salaam.
———. 2004. Kagera Health and Development Survey 2004. Ref. TZA_2004_KHDS_v01_M.
World Bank and University of Dar es Salaam.
———. 2010. Kagera Health and Development Survey 2010. World Bank and University of
Dar es Salaam.
WTO. 2019. ‘East African Community (EAC) Trade Policy Review: Report by the Secreteriat’.
WT/TPR/S/384. Geneva, Switzerland: World Trade Organization.
Zeng, Dao-Zhi, and Laixun Zhao. 2010. ‘Globalization, Interregional and International
Inequalities’. Journal of Urban Economics 67 (3): 352–61.
192
193
Appendix C
Regional Market Integration and Household
Welfare: Spatial Evidence from the East African
Community
194
C.1 Analytical Insights to the Model
As anticipated in Subsection 4.4.5, this section provides analytical insights into the simulation
results of the main text. As in the NEG tradition, the analysis revolves around checking the
stability of two specific equilibria of the model, namely the “spreading” equilibrium, i.e. where
the real wage differential 𝜔𝜔1/𝜔𝜔2 = 1 and 𝜆𝜆1 = 𝜆𝜆2 = 0.5, and the “agglomerated” equilibrium,
where all manufacturing is concentrated in one of the regions such that 𝜆𝜆1 = 1 and 𝜔𝜔1/𝜔𝜔2 ≥1
or 𝜆𝜆1 = 0 and 𝜔𝜔1/𝜔𝜔2 ≤1. The analytical evaluation is thereby concerned with assessing the
stability of both equilibria to varying internal transport costs, i.e. at which level of 𝑇𝑇𝐷𝐷 spreading
is broken and agglomeration sustainable, hence the name “sustain” and “break” analysis. As
such, the treatment closely follows the exposition in Chapter 5 of Fujita et al. (2001), albeit
for a modified spatial layout and focusing on the influence of external trade costs rather than
internal ones only. Note that the analysis is constrained to these two cases because at these
points, the system of non-linear equations reduces to a more tractable set which simplifies the
analysis. However, given that 𝜆𝜆 varies all the way from 0 to 1, and choices have to be made on
the other parameter values, (the stability of) further equilibria may depend on many such
combinations, the main analysis of this paper relies on the numerical simulations of Subsection
4.4.5 in order to give a full picture of the long-run dynamics. As first introduced in Subsection
4.4.5, we compare the analytical results for the setting with heterogeneous (asymmetric) intra-
national space to the ones drawn from a 2 + 2 setting with homogenous (symmetric) intra-
national space. Note, that as mentioned in Subsection 4.4.5, the analysis of the latter set-up
mirrors the one fist shown in (Monfort and Nicolini 2000).
C.1.2 Symmetry Breaking
We start by analyzing the robustness of a symmetric equilibrium, that is, the configuration in
which 𝜆𝜆1 = 𝜆𝜆2 = 0.5 and 𝜔𝜔1/𝜔𝜔2 = 1. From the discussion in Subsection 4.4.4, and visually
depictable in Figure 4.3, we know that this equilibrium is stable if migrating in either direction
leads to a lower real wage in the destination region than in the origin. Stated more generally,
for the symmetric equilibrium to be a stable one, the slope of the total differential with respect
to 𝑑𝑑𝑑𝑑i has to satisfy:
𝑑𝑑𝜔𝜔𝑖𝑖
𝜔𝜔𝑗𝑗
𝑑𝑑𝜆𝜆𝑖𝑖
≤ 0.
(C.1)n
Before we start deriving an expression for (C.1), notice that Figure 4.3, specifically, the
differences across Panels A and B as well as C and D, already hold the insight insofar as a
symmetric equilibrium can be upheld during trade liberalization within our heterogeneous 2 +
2 setting. The simulations show that any move away from autarky (𝜏𝜏≠∞) also entails a move
away from the symmetric distribution of manufacturing as an equilibrium. This is observable
195
by the shift of the cut point to the left (Panel B) and to the right (Panel D). Hence, contrary
to Panels A and C, the relative share of the manufacturing workforce across regions in the first
type of equilibrium is dependent on external transport costs. As such, for the “symmetry
breaking” analysis, we are limited to the 2 + 2 setting with homogenous intra-national space.
In this setup, the equal distribution is always a possible equilibrium, independent of the
(external) transport costs. This is explicated in the following steps. First note that when
𝜆𝜆𝐻𝐻(𝑜𝑜𝑜𝑜𝑜𝑜),𝐹𝐹(𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜) = 0.5, income 𝑌𝑌𝐻𝐻,𝐹𝐹= 0.5 and from this, the wage reduces to 𝑤𝑤𝐻𝐻,𝐹𝐹= 1, and
this is true for all (home and foreign) regions, hence the drop of the indices.135 We can confirm
this by plugging these values into (4.36) through (4.44), and solving.
𝑌𝑌= 𝛿𝛿
2 + (1 −𝛿𝛿)
2
= 0.5
(C.2)n
𝐼𝐼= [0.5 + 0.5𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐹𝐹
1−𝜀𝜀]
1
1−𝜀𝜀
(C.3)n
Plugging these two results into the wage equation leads to
𝑤𝑤= º0.5𝐼𝐼𝜀𝜀−1 + 0.5𝐼𝐼𝜀𝜀−1𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝐼𝐼𝜀𝜀−1𝑇𝑇𝐹𝐹
1−𝜀𝜀»
1
𝜀𝜀, or
𝑤𝑤= ê0.5 + 0.5𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐹𝐹
1−𝜀𝜀
𝐼𝐼1−𝜀𝜀
ë
1𝜀𝜀
= 1
(C.4)n
𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔𝑔 𝑡𝑡ℎ𝑎𝑎𝑎𝑎 𝐼𝐼1−𝜀𝜀= 0.5 + 0.5𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐹𝐹
1−𝜀𝜀
Note that the subscripts can be dropped as income, price indices and therefore wages are equal
in all regions at the symmetric equilibrium. Note that the existence of this equilibrium
configuration does not depend on foreign trade costs 𝑇𝑇𝐹𝐹.136 From this set of manipulations,
specifically (C.3) and (C.4), it is easily seen that real wages across regions are equal. With
these results in mind, we are able to proceed with a crucial simplification in the derivation of
the total differential. Namely, at this symmetric equilibrium, a change in one of the endogenous
variables for one region requires the identical change for the other region in the opposite
direction (Fujita et al. 2001). This can be confirmed, for instance, by computing the total
derivatives of the two income equations for regions 1 and 2, plugging in the equilibrium values,
and checking whether 𝑑𝑑𝑌𝑌1 + 𝑑𝑑𝑌𝑌2 = 0.
135 Note that the superscripts 𝐻𝐻 and 𝐹𝐹 indicate identical values for all home (1 and 2) and foreign regions (3
and 4), respectively.
136 Accordingly, the existence of a symmetric equilibrium in the homogenous 2 + 2 region case is independent
of the foreign labor distribution.
196
𝑑𝑑𝑌𝑌1 = 𝑑𝑑𝜆𝜆1𝑤𝑤1𝛿𝛿+ 𝑑𝑑𝑤𝑤1𝜆𝜆1𝛿𝛿
(C.5)n
𝑑𝑑𝑌𝑌2 = −𝑑𝑑𝜆𝜆1𝑤𝑤2𝛿𝛿+ 𝑑𝑑𝑤𝑤2(1 −𝜆𝜆1)𝛿𝛿
(C.6)n
Where we made use that 𝜆𝜆2 = (1 −𝜆𝜆1). If we now plug in the equilibrium values derived in
(C.2) through (C.4) and assuming analogously that 𝑑𝑑𝑤𝑤1 = −𝑑𝑑𝑤𝑤2, gives
𝑑𝑑𝑌𝑌1 = 𝑑𝑑𝑑𝑑1𝛿𝛿+ 𝑑𝑑𝑤𝑤10.5𝛿𝛿 𝑎𝑎𝑎𝑎𝑎𝑎 𝑑𝑑𝑌𝑌2 = −𝑑𝑑𝑑𝑑1𝛿𝛿−𝑑𝑑𝑤𝑤10.5𝛿𝛿
(C.7)n
which satisfies 𝑑𝑑𝑌𝑌1 + 𝑑𝑑𝑌𝑌2 = 0.
(C.8)n
This confirms that 𝑑𝑑𝑌𝑌𝐻𝐻≡𝑑𝑑𝑌𝑌1 = 𝑑𝑑𝑌𝑌2. Hence, the total derivate of the income at the symmetric
equilibrium can be written as
𝑑𝑑𝑌𝑌𝐻𝐻= 𝑑𝑑𝑑𝑑𝑑𝑑+ 𝛿𝛿
2 𝑑𝑑𝑑𝑑
(C.9)n
And equally for the foreign country such that 𝑑𝑑𝑑𝑑≡𝑑𝑑𝑌𝑌𝐻𝐻= 𝑑𝑑𝑌𝑌𝐹𝐹. This operation can be
confirmed for all other equilibrium equations, i.e. for price indices 𝑑𝑑𝑑𝑑 and wages 𝑑𝑑𝑑𝑑.137
Importantly, the same intuition applies to the total differential of the real wage equations also,
such that it suffices to assess only the change in the real wage of one of the two foreign or
home regions, and (C.1) effectively boils down to 𝑑𝑑𝜔𝜔𝑖𝑖
𝐻𝐻,𝐹𝐹/𝑑𝑑𝜆𝜆𝑖𝑖
𝐻𝐻,𝐹𝐹, e.g. given by
𝑑𝑑𝜔𝜔𝑖𝑖/𝑑𝑑𝜔𝜔𝑗𝑗
𝑑𝑑𝜆𝜆𝑖𝑖
≡ 𝑑𝑑𝑑𝑑
𝑑𝑑𝑑𝑑 = 𝑑𝑑𝑑𝑑· 𝐼𝐼−𝛿𝛿−𝑑𝑑𝑑𝑑· 𝑤𝑤· 𝐼𝐼−(1+𝛿𝛿) · 𝛿𝛿
𝑑𝑑𝑑𝑑
(C.10)n
After some manipulations, which involves plugging in (C.9) into the equations for 𝑑𝑑𝑑𝑑𝐻𝐻,𝐹𝐹 and
𝑑𝑑𝑑𝑑𝐻𝐻,𝐹𝐹, and solving these four equations as a system, we arrive at expressions to plug into
(C.10) which are solely dependent on the exogenous parameter values 𝛿𝛿 and 𝜀𝜀 as well the
iceberg trade costs 𝑇𝑇𝐷𝐷 and 𝑇𝑇𝐹𝐹.138 As this expression hinges on the totals differentials from all
four regions, the expressions is unwieldly compared to a 2-region NEG model. Hence, to
facilitate interpretation, the results from this “break” analysis are provided graphically in
Figure C.1 for a given set of parameter values and the three levels of external transport costs
137 I do not show them here because they grow relatively large as they are additionally dependent on changes
in the endogenous variables of the foreign country.
138 Also using 𝐼𝐼= í0.5 + 0.5𝑇𝑇𝐷𝐷
(1−𝜀𝜀) + 𝑇𝑇𝐹𝐹
(1−𝜀𝜀)î
1
1−𝜀𝜀.
197
𝑇𝑇𝐹𝐹. Note that we are now also able to assess the stability of this equilibrium to a range of
internal transport costs which is not simply set to 𝑇𝑇𝐷𝐷= 1.60 as previously. As in the NEG
tradition, Table C.1 additionally provides the “break” values 𝑇𝑇𝐷𝐷(𝐵𝐵) together with the ones
derived in b), i.e. the “sustain” values 𝑇𝑇𝐷𝐷(𝑆𝑆), for a range of parameter values 𝛿𝛿 and 𝜀𝜀 and the
three levels of trade liberalization 𝑇𝑇𝐹𝐹. Figure C.1 provides the results to the break analysis for
the four-region model with homogenous intra-national space, plotting (𝑑𝑑𝜔𝜔2/𝑑𝑑𝜔𝜔1)/𝑑𝑑𝜆𝜆2 across an
increasing value of intra-national trade costs 𝑇𝑇𝐷𝐷 for three different values of international trade
costs 𝑇𝑇𝐹𝐹, separately. Note that by symmetry, we are free to choose the direction of effects in
any of the two home or foreign regions. We focus on the real wage differential of region 2,
however, as it facilitates the comparison with the sustain analysis in b) and given that we have
seen the increased draw towards border regions in the simulations discussed in Section 4.4.
Panel A
Panel B
Figure C.1: Internal transport costs and symmetric equilibrium
Note first, that with zero transport costs, i.e. 𝑇𝑇𝐷𝐷= 1, there is no difference in the (real)
wage across regions such that the total differential is zero at the origin.139 Increasing the intra-
national transport costs from this point on in Panel A shows that up until a level of transport
costs of 𝑇𝑇𝐷𝐷= 1.47, the symmetric equilibrium is unstable, given that a move away from region
1 increases real wages at the destination, i.e. in region 2, i.e. the total differential is positive.140
In other words, in this scenario, the cost- and demand linkages of agglomerating are strong
enough to render the cost of serving the demand of region 1 at a distance as profitable. For
any increase in transport costs beyond this point, this is not true anymore and manufacturing
activity spreads out. The dashed and dotted curves show the effect of trade liberalization,
which is to increase the range of transport costs within which a symmetric equilibrium is
139 The basic NEG setup hinges on positive transport costs, i.e. a 𝑇𝑇𝐷𝐷≠1 which essentially means that
economies are not rendered identical in all respects.
140 You can retrieve the precise value from Table C.1.
198
unstable. Why is this the case? We need to analyze a couple of (countervailing) effects step by
step in order to interpret the likely effects (e.g. Crozet and Koenig 2004a; Brülhart et al. 2004;
Brülhart 2011). Note at first that liberalizing trade, i.e. decreasing 𝑇𝑇𝐹𝐹 down from prohibitive
levels, causes the components in both the price index (4.40)-(4.43) as well as the wage equation
(4.44)-(4.47) that are dependent on external markets 3 and 4 to make up a larger component
of the overall 𝐼𝐼 and 𝑤𝑤 at the respective location. This results in several dynamics. First, it
lowers producers’ need to locate close to consumers in the home country as a larger share of
their sales come from abroad, i.e. the demand linkage is lowered. Secondly, it analogously
decreases consumer’s need to locate near producers in the home country as a larger share of
their demand now stems from abroad, i.e. the cost linkage is also lowered (Crozet and Koenig
2004b). As such, lowering 𝑇𝑇𝐹𝐹 essentially reduces agglomerating tendencies inside the domestic
country which is the well-known result put forward in Krugman and Elizondo (1996). Note
that this also means that the moderating force of these cost and demand linkages as given by
the internal transport costs 𝑇𝑇𝐷𝐷 is weakened also, as can be depicted by an attenuation of the
slopes in Figure C.1. But why is agglomeration in this present model more likely then? There
are two crucial differences to the model in Krugman and Elizondo (1996) which turn this result
around. For one, as Crozet and Koenig (2004a) point out, Krugman and Elizondo (1996) do
not model an immobile agricultural sector, the demands of which act as a spreading force, and
secondly, they explicitly model congestion costs of agglomerations (such as rent or commuting).
These congestion costs are independent of trade costs, hence decreasing trade costs does not
lower the centrifugal tendency of them. On the contrary, the dispersion force of the type of
model I as well as (Monfort and Nicolini 2000) employ, immobile farmers, is crucially dependent
on the trade costs to serve them. Together with the key result of the original core NEG model
(Krugman 1991), i.e. that the strength of the centrifugal force given by these farmers falls faster
in (international) transport costs than the strength of the centripetal force, this may display
one reason why the result is turned towards agglomeration in our case (see also the discussion
in Brülhart 2011). However, there is one further potential reason why agglomeration tendencies
may be increased by opening up to external markets, which is increased competition of firms
from abroad (Crozet and Koenig 2004b). Remember from Section 4.4 that the dispersion force
stems from increased competition given by the positive relationship between the price index 𝐼𝐼
and the break-even wage rate firms are able to afford to pay, as seen in (4.44) through (4.47).
Thereby, in a similar line of argument as above, from the point of producers, decreasing 𝑇𝑇𝐹𝐹
lowers the relative importance of domestic competition, such that sheltering away from local
firms is less important given the new competition foreign firms pose (Crozet and Koenig 2004a;
Brülhart et al. 2004). By looking at the results in Figure 4.3 above as well as Figure C.1, it
seems that this decreased competition effect dominates the decreased agglomeration forces.
Additionally, when looking at Panel B in Figure C.1, we see this effect amplified up to a point
where the does not even exist a break point anymore, i.e. 𝑇𝑇𝐷𝐷= ∅, and agglomeration is the
199
only long-term stable equilibrium in the case of free trade. Note that the difference between
Panel A and B makes intuitive sense, given that the only change between the two is the reduced
elasticity of substitution, from 𝜀𝜀= 6 to 𝜀𝜀= 4. This change increases the forward and backward
linkages, by reasons given in Subsection 4.4.5, leading to an increased agglomerating force
(Fujita et al. 2001).
Notice how these results compare to the simulations in Figure 4.3, in which we have set
𝑇𝑇𝐷𝐷= 1.60. At this point on the x-Axis (Figure C.1), spreading is never sustainable for an 𝜀𝜀=
4, as is confirmed by the positive slope in Figure 4.3 Panel B. And for an 𝜀𝜀= 6 only stable
when trade liberalization has not fully concluded yet, e.g. a value 2.00 ≥ 𝑇𝑇𝐹𝐹> 1.60 (Figure
4.3 Panel A). Table C.1 encapsulates these results at one glance. We see that a decrease in the
elasticity of substitution consistently shifts the break point to the right, i.e. increases the range
of values for which spreading is unsustainable. Notice, also that an increase in 𝛿𝛿, i.e. an increase
in the share of income devoted to manufactures, has the identical effect.
How do these deliberations compare to the model with heterogeneous intra-national
space? Panel B in Figure 4.3 shows that for the case with a lower product differentiation (𝜀𝜀=
6), an equilibrium where economic activity is spread out is more likely at higher degrees of
trade liberalization than in the homogenous case; albeit with higher shares of economic activity
placed at the border. In this case, it seems that the competition effect from abroad does not
yet seem to fully dominate the local one and its spreading tendency.141 Intuitively, firms and
consumers now also profit from increased agglomeration, but there is a bias towards
agglomerating in the vicinity of the newly accessed markets, i.e. in region 2. This notion is
further confirmed seen in Panel D of Figure 4.3, where, as in the case with homogenous intra-
national space, decreased product differentiation (𝜀𝜀= 4) causes the curve to reverse, rendering
full agglomeration as the only stable, long-run equilibrium, but now this is more likely to
happen at the border region 2, compared to the interior region 1.
C.1.3 Sustainable Agglomeration
We now turn to the “sustain” analysis. In 4.4.4, we have already established that the stability
of this equilibrium trivially depends on the condition 𝜔𝜔𝑖𝑖/𝜔𝜔𝑗𝑗≥1 if 𝜆𝜆𝑖𝑖= 1. As such, we need to
derive an expression for the real wage differential at this point which, as in C.1.2, depends only
on the parameter values 𝛿𝛿, 𝜀𝜀 as well as, importantly, the different types of iceberg trade costs
𝑇𝑇𝐷𝐷, 𝑇𝑇𝐹𝐹, 𝑇𝑇𝐷𝐷𝐷𝐷, 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷. For this analysis, we are able to derive analytical solutions for both spatial
layouts, i.e. the homogenous as well as the heterogeneous layout of trade costs. As in C.1.2,
the first step entails plugging in the equilibrium values for 𝜆𝜆𝑖𝑖, i.e. 𝜆𝜆1 = 𝜆𝜆4 = 1 and
correspondingly, 𝜆𝜆2 = 𝜆𝜆3 = 0, and noting that the wage equations of regions 1 and 4 reduce to
141 Although, notice the slope does in fact decline in Figure 4.4 Panel B slightly with progressing trade
liberalization.
200
𝑤𝑤1 = 𝑤𝑤4 = 1.142 To see this, note that the income equations (4.36) through (4.39) in this spatial
configuration are given by n
𝑌𝑌1,4 = (1 + 𝛿𝛿)
2
(C.11)n
𝑌𝑌2,3 = (1 −𝛿𝛿)
2
(C.12)n
Note, from this set of four equations, income in region 1 is always higher, which represents the
demand (backward linkage) introduced in Subsection 4.2.2. Correspondingly, the price indices
reduce to
𝐼𝐼1,4 = [1 + 𝑇𝑇𝐹𝐹
1−𝜀𝜀]
1
(1−𝜀𝜀)
(C.13.a)n
𝐼𝐼2,3 = [𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐹𝐹
1−𝜀𝜀]
1
(1−𝜀𝜀)
(C.14.a)n
And for the model with heterogeneous intra-national space:
𝐼𝐼1,4 = [1 + 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷
1−𝜀𝜀]
1
(1−𝜀𝜀)
(C.13.b)n
𝐼𝐼2,3 = [𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐷𝐷𝐷𝐷
1−𝜀𝜀]
1
(1−𝜀𝜀)
(C.14.b)n
The first summand of the price index equations is equal for both spatial layouts (C.13.a) and
(C.13.b). The difference lies in the second summand in the price indices. In the model with
homogenous intra-nation space (C.13.a) and(C.14.a), the cost of living in the peripheral region
is at best equal to that of the agglomerated region, but for transport which satisfy 𝑇𝑇𝐷𝐷> 1 are
always higher than in the agglomerated region, given that 𝜀𝜀> 1 which we assume by default
given our CES utility structure. This is the cost (forward) linkage as described in Subsection
4.2.2. However, in (C.13.b) and (C.14.b), given that 𝑇𝑇𝐷𝐷𝐷𝐷< 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷, we cannot readily determine
whether cost of living is higher or lower in region 1. Moving on, we plug (C.11) through (C.14.a)
into the wage equations:
142 This can be confirmed by guessing 𝑤𝑤1 = 1, working out (4.36) through (4.43) using this value for 𝑊𝑊1 and
seeing that (4.44) is indeed 1. Notice that in our heterogeneous 2+2 case, this also entails assuming an equal
distribution in the foreign country, i.e. full agglomeration as for instance given by 𝜆𝜆4 = 1 together with 𝑊𝑊4 =
1.
201
𝑤𝑤1,4 = °𝑌𝑌1,4(1 + 𝑇𝑇𝐹𝐹
1−𝜀𝜀)
1 + 𝑇𝑇𝐹𝐹
1−𝜀𝜀
+ 𝑌𝑌2,3(𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐹𝐹
1−𝜀𝜀)
𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐹𝐹
1−𝜀𝜀
³
1𝜀𝜀
, 𝑜𝑜𝑜𝑜
𝑤𝑤1,4 = ð(1 + 𝛿𝛿)
2
+ (1 −𝛿𝛿)
2
ñ
1𝜀𝜀
= 1
(C.15.a) n
And similarly, for heterogeneous intra-national space
𝑤𝑤1,4 = °𝑌𝑌1,4(1 + 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷
1−𝜀𝜀)
1 + 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷
1−𝜀𝜀
+ 𝑌𝑌2,3(𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐷𝐷𝐷𝐷
1−𝜀𝜀)
𝑇𝑇𝐷𝐷
1−𝜀𝜀+ 𝑇𝑇𝐷𝐷𝐷𝐷
1−𝜀𝜀
³
1𝜀𝜀
, 𝑜𝑜𝑜𝑜
𝑤𝑤1,4 = ð(1 + 𝛿𝛿)
2
+ (1 −𝛿𝛿)
2
ñ
1𝜀𝜀
= 1
(C.15.b)n
Where we made use of a similar manipulation as in C.1.2, i.e. that 𝐼𝐼−1−𝜀𝜀≡[𝐼𝐼]−1. As such,
wages in the interior are always 1. Now, real wage equations reduce to
𝜔𝜔1,4 =
1
ò(1 + 𝑇𝑇𝐹𝐹
1−𝜀𝜀)
1
1−𝑒𝑒ó
𝛿𝛿
(C.16.a)
𝜔𝜔1,4 =
1
ò(1 + 𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷
1−𝜀𝜀)
1
1−𝑒𝑒ó
𝛿𝛿
(C.16.b)n
Note that, technically, the (real) wage equations for regions 2 and 3 are only implied functions,
as there is no actual manufacturing wage in this spatial configuration given by the absence of
manufacturing workers, i.e. 𝜆𝜆2 = 𝜆𝜆3 = 0. One can think of these implied wages as the maximum
wage that firms moving to this location would be able to pay (Fujita et al. 2001). The derivation
of 𝜔𝜔2,3 follow the same type of manipulations just made (C.15.a) through (C.16.b) and lead to
expressions only dependent on the parameter values 𝛿𝛿, 𝜀𝜀 and transport costs 𝑇𝑇𝐷𝐷, 𝑇𝑇𝐹𝐹, 𝑇𝑇𝐷𝐷𝐷𝐷,
𝑇𝑇𝐷𝐷𝐷𝐷𝐷𝐷.143 We now have all the ingredients for an expression of the real wage differential within
the home or foreign economy, i.e. 𝜔𝜔𝑖𝑖
𝐻𝐻,𝐹𝐹/𝜔𝜔𝑗𝑗
𝐻𝐻,𝐹𝐹. As in C.1.2, to assess the analytical results, we
plot the real wage differential in the home economy against the intra-national transport costs
𝑇𝑇𝐷𝐷 and for our three levels of external transport costs 𝑇𝑇𝐹𝐹. This is done in Figure C.2.
Importantly, note the change in the y-axis; we now express the real wage differential from the
point of view of the peripheral region, i.e. plot 𝜔𝜔2/𝜔𝜔1 to facilitate a comparison to the break
143 Results are not reported here.
202
analysis in C.1.2. More precisely, the range of transport costs 𝑇𝑇𝐷𝐷 for which this type of
equilibrium is sustainable also lies below the constant. Again, we analyze the dynamics from
the point of view of region 2, i.e. assess when full agglomeration in region 1 is unsustainable.
In contrast to C.1.2 we are now also able to discuss peculiarities of the model with
heterogeneous intra-national space analytically.
Panel A
Panel B
Panel C
Panel D
Figure C.2: Internal Transport costs and Sustainable Agglomeration
The results shown in Figure C.2 show similar tendencies as in C.1.2, that is, increased trade
liberalization increases the range of values for which a full agglomeration equilibrium is more
likely (note the shift in the sustain point 𝑇𝑇(𝑆𝑆) to the right in Panel A). By construction, this
mirrors the result in Monfort and Nicolini (2000). Also, by reasons given in Subsection 4.4.5, a
lower level of 𝜀𝜀 cause agglomeration forces to be strengthened, up to the point where a sustain
point does not exist for low transport costs across countries (Panel C). Hence, for a level of
𝑇𝑇𝐹𝐹≤2.00 and 𝜀𝜀= 4, there does not exist a level of internal transport costs for which
203
agglomeration becomes unsustainable. In other words, the existence of external markets renders
the costs of serving domestic markets from a distance negligible and it increasingly pays to
agglomerate given reduced international trade costs. Notice, however, the stark difference to
Panels B and D, i.e. in the case of heterogeneous intra-national space. Here, external trade
liberalization causes a decrease in the range of intra-national transport costs 𝑇𝑇𝐷𝐷 for which
agglomeration in region 1 is sustainable, and for all parameters tested, there exists a level where
agglomeration in region 1 is unsustainable (see Table C.1). Whence the difference? We need to
latch on to the discussion in C.1.2 where we discussed the relative influence of centrifugal and
centripetal forces of a changing 𝑇𝑇𝐹𝐹. For heterogeneous intra-national space, there is now an
additional component which mediates the relative strength of these two forces, namely, the
differential exposure to the external markets, initially shown by Crozet and Koenig 2004b). To
the former (centrifugal forces), while agglomeration tendencies are lowered, one may expect an
increased draw of firms and consumers to the border so as to benefit from better access to new
demand and supply, respectively. To the latter (centripetal forces), there is the possibility that
the dispersion force is further amplified, which pushes economic activity towards region 1 given
that its larger distance to the border provides an increased level of protection from new foreign
competition.
What are the implications for our present results? For all of the parameter configurations
of 𝜀𝜀 and 𝛿𝛿 tested (see Table C.1), we see a falling range of intra-national transport costs 𝑇𝑇𝐷𝐷 for
which agglomeration in region 1 is sustainable.144 Hence, sheltering in the interior regions does
not seem to happen to a larger degree than the draw to the border when international trade
costs are decreased. This goes against Crozet and Koenig (2004b) where a push to the interior
happens at intermediate international trade costs. The difference in these results most likely
stems from the setup of the foreign economy and the moderating force of this. In their model,
the foreign economy is larger than the domestic one, which has arguably larger bearings on the
competition effect just described.145 However, contrary to Behrens et al. (2006), what Crozet
and Koenig (2004b) have not analyzed in their 2 + 1 setup, is the influence of the relative size
of foreign regions on these dynamics. Figure 4.5 in Subsection 4.4.5 provides the main results
of this analysis, where we have seen that the draw to the border may be lower or higher when
foreign economic activity agglomerates in the interior but nonetheless exists for any distribution
of home or foreign activity. Hence, for two equally sized economies, foreign economic inequality
cannot turn around our main results, which is that the draw to the border dominates any
benefit by sheltering from foreign competition. What we can say, however, is that this effect
144 What is also confirmed is that for higher shares of 𝛿𝛿 (consumption share of manufactures) and lower levels
of 𝜀𝜀 (higher product market competition) agglomeration is supported for a wider range of transport costs, i.e.
agglomeration forces are strengthened.
145 Of course, this also depends on the structure of the economies, i.e. whether the two economies are
complementary in their trade or whether one of the countries dominates in either imports or exports. These
effects are analyzed in Brülhart et al. (2004), albeit for a different model set-up concerning the utility function
and thereby not directly comparable to the one in this paper.
204
may be moderated by foreign economic inequality. As such, from the results in Figure 4.5, it
seems as if sheltering in the interior is generally more important when much of the foreign
activity is agglomerated at the border, i.e. in region 3 (a low 𝜆𝜆4). Intuitively plausible, this
effect seems to be more relevant for the case of increased product market competition (compare
Panels B and D).
Table C.1 Sustain and Break points across 𝛿𝛿, 𝜀𝜀 and 𝑇𝑇𝐹𝐹
Symmetric
1.97
2.34
2.47
4.00
3.30
14.62
Asymmetric
-
2.34
-
4.00
-
14.62
Symmetric
1.63
1.81
1.90
2.52
2.30
5.00
Asymmetric
-
1.81
-
2.52
-
5.00
Symmetric
1.46
1.57
1.64
2.00
1.90
3.16
Asymmetric
-
1.57
-
2.00
-
3.16
Symmetric
3.58
Asymmetric
-
2.23
-
3.69
-
12.50
Symmetric
1.83
2.28
2.69
Asymmetric
-
1.78
-
2.45
-
4.78
Symmetric
1.52
1.68
1.79
3.83
2.48
Asymmetric
-
1.56
-
3.83
-
3.11
Symmetric
Asymmetric
-
2.13
-
3.45
-
10.93
Symmetric
3.02
Asymmetric
-
1.73
-
2.37
-
4.49
Symmetric
1.70
2.23
Asymmetric
-
1.54
-
1.94
-
3.01
Sustain and Break Values
5
Notes: The values in this table represent the intra-national iceberg transport costs at which
agglomeration turns "sustainable" [T(S)] and where the symmetric (spreading) equilibrium is
"broken" [T(B)], i.e. at which real wages in the agglomeration exceed those in the periphery and
a migration towards one of the regions leads to real wage gains, respectively. For more details
on the derivation, see Section C.1 of the Appendix.
205
C.1.4 Full Simulations
As anticipated in Section 4.4, this subsection provides the full set of simulations, of which
selected results are presented and discussed in the main text. I thereby provide the three-
dimensional depictions of the simulations, which were discussed as simpler, two-dimensional
illustrations before. Figure C.3 and Figure C.4 plot the plane of real wage differentials 𝜔𝜔1 ⁄ 𝜔𝜔2
spanned by all possible home and foreign spatial configurations, given by relative shares of the
home and foreign workforces 𝜆𝜆1 and 𝜆𝜆4, respectively for a given set of parameter values.
Additionally, I provide the full set of corresponding contour lines in Figure C.5 and Figure C.6
which depict the changing influence of foreign economic inequality for stable and unstable
equilibria. As established in Subsection 4.4.5, I plot the results for all three levels of external
transport costs and additionally, compare results from the main 2+2 setting against the more
general 2+2 setting with homogeneous intra-national space.
206
Panel A
Panel B
Panel C
Panel D
Panel E
Panel F
Figure C.3: Three-dimensional Depiction of Spatial Equilibria (𝜺𝜺= 𝟔𝟔)
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 1.60.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 1.60.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 1.60.
Parameter values: δ = 0.4, ε = 6, Tᵈ = 1.60, Tᶠ = 1.60.
207
Panel A
Panel B
Panel C
Panel D
Panel E
Panel F
Figure C.4: Three-dimensional Depiction of Spatial Equilibria (𝜺𝜺= 𝟒𝟒)
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = ∞.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 2.00.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 1.60.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 1.60.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 1.60.
Parameter values: δ = 0.4, ε = 4, Tᵈ = 1.60, Tᶠ = 1.60.
208
Panel A
Panel B
Panel C
Panel D
Panel E
Panel F
Figure C.5: Contour Lines of Spatial Equilibria (𝜺𝜺= 𝟔𝟔)
209
Panel A
Panel B
Panel C
Panel D
Panel E
Panel F
Figure C.6: Contour Lines of Spatial Equilibria (𝜺𝜺= 𝟒𝟒)
210
C.2 Further Empirical Results and Documentation
211
Figure C.7: East African (Northern and Central) Trade Corridor
Source: Figures taken from Nathan Associates (2011) pages 6 and 31, respectively.
212
Figure C.8: East African Trade, Means of Transport
Source: Table taken from Nathan Associates (2011) page 63.
213
Figure C.9: DiD in Core Regions with varying Intervention Year
Notes: The “Point Estimates” depicted represent difference-in-difference estimates for the “Core Region {0,1}”, including a 95%
confidence interval, for each of the three dependent variables. The estimates are constructed for each (placebo) intervention year,
separately, by expanding the data past the respective cut-off. E.g. for the estimate on the placebo intervention year 1997, the
regressions use data from first available data for all dependent variables and includes survey years up until 2000. For the
intervention year 1998, the data frame is expanded to include the year 2001, for 1999 it includes 2002, and so on. The plot is
restricted to estimates from 1997 onwards, given that data for outcome variables in Placebo Countries before this date is only
sparsely available. See the Table C.16 for a longer range of estimates. for the main outcome variables in placebo countries only
few samples provide the main dependent variables before this date. Again, Wealth Index is available only in Malawi in 1992
Rwanda in1992, and Zambia in 1996. 1997 is the first year for which repeat observations in a country are available. The data
used for the estimations stem from the Women, Men and Household Recodes of the Demographic and Health Surveys (DHS),
including KAP, MIS and AIS rounds. For the “EAC Countries” the total sample includes the Kenya survey rounds of 1988-89,
1993, 1998, 2003, 2008, the Tanzania survey rounds of 1991-92, 1994, 1996, 1999, 2003-04, 2007-08, 2010, and the Uganda survey
rounds of 1988-89, 1995, 1995-96, 2000-01, 2006, 2009, 2011. For the “Placebo Countries” the total sample includes the Ethiopia
survey rounds of 2000,2005, 2011, the Mozambique survey rounds of 1997, 2003, 2009, 2011, the Malawi survey rounds of 1992,
2000, 2004, 2010, the Rwanda DHS survey rounds of 1992,2000, 2005, and the Zambia survey rounds of 1996, 2001-02 and 2007.
Standard errors are clustered at the region level. All dependent variables represent z-scores, i.e. are standardized with a mean of
zero and a standard deviation of one.
214
Figure C.10: DiD in Border Regions with varying Intervention Year
Notes: The results depicted in this table are estimated in the same way as those shown in Figure C.9 such that
most of the respective notes apply here. However, not the following important amendments: The “Point Estimates”
depicted represent difference-in-difference estimates for the Border Region {0,1}, including a 95% confidence
interval, for each of the three dependent variables. See full notes below Figure C.9.
215
Table C.2: Summary Statistics
Panel a)
Mean St.Dev. Min. 1st Quart. Median 3rd Quart. Core Max.
N
Basic Characteristics
Afrobarometer (AFB)
Age
36
14
17
25
33
43
34
101
38,322
Education (Level)
3.3
1.8
0.0
2.0
3.0
4.5
4.2
9
38,556
Female {0,1}
0.5
0.5
0.0
0.0
1.0
1.0
0.5
1.0
38,322
Demographic and Health Survey (DHS)
Age
29
10
15
20
27
36
28
60
332,725
Education (Years)
5.83
4.00
0.00
2.00
7.00
8.00
8.71
26
330,979
Female {0,1}
0.8
0.4
0.0
1.0
1.0
1.0
0.8
1.0
332,725
Kagera Health and Development Survey (KHDS)
Age
22
19
0
7
16
31
23
105
48,075
Education (Years)
5.32
3.30
0.00
3.00
6.00
7.00
6.12
22
30,154
Female {0,1}
0.5
0.5
0.0
0.0
1.0
1.0
0.5
1.0
48,119
Main Covariates
AFB
Core {0,1}
0.13
0.34
0.00
0.00
0.00
0.00
1.00
1.00
38,644
Urban {0,1}
0.26
0.44
0.00
0.00
0.00
1.00
0.73
1.00
38,644
Agglomeration {0,1}
0.28
0.45
0.00
0.00
0.00
1.00
0.64
1.00
38,644
Road Distance to nearest Border Crossing (km)
315
264
0.8
127
236
407
240
1362
38,644
Inv. rel. Distance to nearest Border Crossing [0,1]
0.71
0.22
0.00
0.61
0.76
0.87
0.77
1.00
38,644
DHS
Core {0,1}
0.09
0.29
0.00
0.00
0.00
0.00
1.00
1.00
210,747
Urban {0,1}
0.22
0.42
0.00
0.00
0.00
0.00
0.73
1.00
245,118
Agglomeration {0,1}
0.25
0.43
0.00
0.00
0.00
1.00
0.42
1.00
210,747
Road Distance to nearest Border Crossing (in km)
342
283
0.1
132
262
437
208
1373 205,050
Inv. rel. Distance to nearest Border Crossing [0,1]
0.69
0.23
0.00
0.58
0.74
0.87
0.77
1.00
205,050
KHDS
Core {0,1}
0.21
0.41
0.00
0.00
0.00
0.00
1.00
1.00
9,288
Urban {0,1}
0.22
0.41
0.00
0.00
0.00
0.00
0.98
1.00
9,288
Road Distance to nearest Border Crossing (in km)
148
107
0
80
103
170
119
1134
9,288
Inv. rel. Distance to nearest Border Crossing [0,1]
0.87
0.10
0.00
0.85
0.90
0.93
0.90
1.00
9,288
Geographical Covariates
AFB
Distance to Harbor (in km)
702
322
1
441
738
956
537
1259
38,644
Distance to Navigable River (in km)
305
271
1
73
198
492
398
963
38,644
Distance to Major Lake (in km)
180
180
0
46
117
251
238
700
38,644
Elevation (in m)
1213
507
4
1072
1205
1519
1
3914
38,644
Terrain Ruggedness (standardized)
0.00
1.00
-0.77
-0.56
-0.30
0.13
-0.49 8.63
38,644
Average Monthly Temperature (in Celsius)
24
3
8
22
24
25
24
32
38,644
Average Monthly Rainfall (in mm)
101
27
17
86
102
116
102
214
38,644
Growing Days {0,365}
292
71
14
247
304
358
310
365
38,644
Malaria Ecology Index
7.0
6.4
0.0
1.5
5.6
10.5
8.2
31.1
38,604
DHS
Distance to Harbor (in km)
696
319
0
446
731
927
744
1267 210,747
Distance to Navigable River (in km)
313
269
1
84
207
502
204
961
210,747
Distance to Major Lake (in km)
183
185
0
43
115
254
102
706
210,747
Elevation (in m)
1189
514
2
1059
1193
1472
1
3248 210,747
Terrain Ruggedness (standardized)
0.00
1.00
-0.75
-0.56
-0.30
0.12
-0.46 12.7
210,747
Average Monthly Temperature (in Celsius)
22
3
11
20
22
23
21
30
210,747
Average Monthly Rainfall (in mm)
94
26
15
77
95
109
98
219
210,747
Growing Days {0,365}
286
72
10
229
295
358
329
365
210,747
Malaria Ecology Index
4.9
8.0
0.0
0.0
1.1
5.5
0.1
36.0
171,881
KHDS
Distance to Harbor (in km)
899
159
0
913
919
938
806
1189
9,288
Distance to Navigable River (in km)
70
157
1
19
31
51
139
963
9,288
Distance to Major Lake (in km)
51
104
0
5
17
66
77
665
9,288
Elevation (in m)
1274
250
3
1188
1259
1392
1
4249
9,287
Terrain Ruggedness (standardized)
0.00
1.00
-0.80
-0.73
-0.48
0.35
-0.36 5.85
9,288
Average Monthly Temperature (in Celsius)
21
1
5
20
21
21
21
27
9,287
Average Monthly Rainfall (in mm)
113
31
48
83
112
142
152
168
9,287
Growing Days {0,365}
316
37
154
292
324
353
341
365
9,288
Malaria Ecology Index
4.2
3.2
0.0
2.6
3.9
4.3
6.1
23.0
9,260
Distribution across Sample
Notes: The table is continued on the next page.
216
Distribution across Distance (Quartiles)
Panel b)
Mean St.Dev. Min.
1st
2nd
3rd
4th
Core
N
Consumption
AFB
How often: Gone without (Lived Poverty) [0,4]
1.12
0.95
0
1.05
1.02
1.16
1.25
0.85
36,371
How often: Gone without Food {0,4}
0.98
1.12
0
1.01
0.91
1.00
1.00
0.75
36,335
How often: Gone without Water {0,4}
1.13
1.31
0
0.92
1.00
1.18
1.38
0.86
36,353
How often: Gone without Medical Care {0,4}
1.26
1.23
0
1.23
1.17
1.28
1.37
0.94
36,300
DHS
Wealth Quintile {1,5}
3.14
1.47
1
3.17
3.42
3.03
2.59
4.61
251,684
International Wealth Index (IWI)
23.44
2.71
18
23.37
23.67
23.46
22.89
25.02 332,725
Comparative Wealth Index (CWI)
-0.67
0.85
-2
-0.66
-0.38
-0.60
-0.91
0.46
104,693
KHDS
Annual p.c. Consumpt. in 2010 TZS ('000)
554
522
36
558
551
436
650
747
6,855
Annual p.c. non-Food Consumpt. in 2010 TZS ('000)
204
299
6
199
199
149
261
293
6,921
Value of Occupied Dwelling in 2004 TZS ('000)
650
3527
0
624
1062
477
356
1425
6,076
Value of Durable Assets in 2004 TZS ('000)
112
1060
0
93
164
45
140
287
6,079
Income & Work
AFB
Worked last Year {0,1}
0.55
0.50
0
0.52
0.56
0.56
0.55
0.66
26,880
Employed Work {0,1}
0.23
0.42
0
0.24
0.29
0.23
0.17
0.36
25,775
Occupation Level: Agr.-Worker-Prof. {1,3}
1.66
0.81
1
1.66
1.83
1.65
1.53
2.18
21,482
How often: Gone without Cash Income {0,4}
2.09
1.24
0
2.08
2.02
2.12
2.15
1.75
36,304
DHS
Worked last Year {0,1}
0.76
0.42
0
0.76
0.75
0.79
0.82
0.72
230,419
Employed Work {0,1}
0.19
0.40
0
0.21
0.30
0.18
0.08
0.47
105,528
Occupation Level: Agr.-Worker-Prof. {1,3}
1.51
0.64
1
1.62
1.70
1.51
1.37
2.00
143,027
Paid in Cash {0,1}
0.52
0.50
0
0.52
0.59
0.50
0.42
0.77
136,866
KHDS
Worked last Year {0,1}
0.26
0.44
0
0.26
0.24
0.19
0.33
0.33
20,259
Employed Work {0,1}
0.12
0.33
0
0.10
0.12
0.13
0.14
0.13
32,017
Occupation Level: Agr.-Worker-Prof. {1,3}
1.20
0.49
1
1.22
1.22
1.11
1.25
1.41
20,446
Salaried Work {0,1}
0.01
0.10
0
0.01
0.01
0.01
0.01
0.02
19,465
Monthly Salary in 2004 TZS ('000)
28.8
55.6
0.0
29.0
33.6
20.2
27.8
33.7
2,118
Agglomeration & Migration
AFB
Population Count
1.26
4.34
0
0.81
2.64
1.02
0.55
6.30
38,644
Population Density
1.47
5.05
0
0.96
3.07
1.21
0.65
7.34
38,644
DHS
Population Count
1.18
3.45
0
0.97
1.93
1.49
0.35
5.33
188,990
Population Density
1.39
4.03
0
1.16
2.26
1.75
0.41
6.23
188,990
At region of residence before EAC {0,1}
0.44
0.50
0
0.61
0.68
0.67
0.74
0.74
69,349
KHDS
Population Count
1.01
1.87
0
0.79
0.80
1.01
1.44
747
9,288
Population Density
0.64
2.16
0
0.29
0.61
0.22
1.45
1.89
9,288
At region of residence before EAC {0,1}
0.92
0.28
0
0.90
0.90
0.90
0.86
0.86
38,592
Main reason for Migration: Economic {0,1}
0.09
0.29
0
0.09
0.11
0.05
0.08
0.14
6,946
After moving to curr. Residence: Paid Empl. {0,1}
0.09
0.28
0
0.09
0.07
0.03
0.12
0.14
2,388
Sentiments & Attitudes
AFB
Helps your Country: REC / EAC {1,3}
1.66
0.92
0
1.67
1.65
1.67
1.66
1.63
11,750
Support for Regional Integration {1,5}
3.71
1.51
1
3.87
3.85
3.70
3.38
3.75
6,394
Would like as Neighbor: Immigrant {1,3}
3.38
1.30
1
3.43
3.31
3.29
3.46
3.18
11,708
Ease of Crossing Borders to live and Work {1,4}
2.21
0.93
1
2.35
2.20
2.15
2.13
2.23
4,784
Present vs. Past: Living Standards of People {1,4}
2.43
1.03
1
2.71
2.71
2.66
2.75
2.74
23,602
KHDS
Subjective HH. Wealth: Today (2004) {1,5}
2.6
0.6
1.0
2.7
2.7
2.5
2.6
2.7
3,313
Subjective HH. Wealth: Ten Years ago (1994) {1,5}
2.6
0.7
1.0
2.7
2.7
2.6
2.6
2.7
3,313
Subjective HH Life Satisfaction: Ladder {1,9}
3.8
1.5
1.0
3.9
3.9
3.6
3.8
4.1
3,313
Subjective HH Life Satisfaction: Ladder {1,9}
3.8
1.5
1.0
3.9
3.9
3.6
3.8
4.1
3,313
Notes: The table depicts summary statistics corresponding to the main sample used in the estimations across the paper. See
the manuscript for more information on the specific survey samples and survey years. The last set of data stem from the
Kagera Health and Development Survey (KHDS). The KHDS summary statistics displayed includes repeat observations from
tracked individuals (a maximum of 6 times). Remaining variations in the number of observations sizes stem from differences
in response rates of variables as well as changes in questions asked across surveys. The distribution "across distance" in b)
calculates mean values of the respective variable within quartiles of road distance to nearest Border crossings.
217
Table C.3: Robustness Checks Summarized (Afrobarometer)
Panel a)
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Tables 4.2 and C.9)
Core {0,1} × EAC
-0.261***
-0.013
-0.191***
0.026
0.021
0.041
(0.077)
(0.050)
(0.054)
(0.067)
(0.079)
(0.083)
Core {0,1} × CU
-0.470***
-0.006
0.728**
0.002
0.057
-0.234**
(0.097)
(0.045)
(0.332)
(0.055)
(0.090)
(0.105)
Core {0,1} × CM
-0.338***
-0.076
1.083***
-0.258**
-0.009
0.043
(0.065)
(0.069)
(0.375)
(0.110)
(0.087)
(0.079)
(a) Standard Errors clustered at EA-Level
(see Table C.22)
Core {0,1} × EAC
-0.261**
-0.013
-0.158
0.026
0.021
0.041
(0.126)
(0.041)
(0.135)
(0.044)
(0.084)
(0.159)
Core {0,1} × CU
-0.470***
-0.006
0.769***
0.002
0.057
-0.234
(0.125)
(0.044)
(0.214)
(0.046)
(0.080)
(0.155)
Core {0,1} × CM
-0.338***
-0.076**
1.113***
-0.258***
-0.009
0.043
(0.119)
(0.036)
(0.179)
(0.041)
(0.078)
(0.150)
(0 004)
(0 036)
(0 000)
(0 000)
(0 907)
(0 774)
(b) No Controls
(see Table C.26)
Core {0,1} × EAC
-0.392***
-0.011
-0.188***
0.018
0.029
-0.047
(0.088)
(0.044)
(0.040)
(0.057)
(0.068)
(0.089)
Core {0,1} × CU
-0.607***
0.005
0.738**
0.002
0.103
-0.318***
(0.116)
(0.048)
(0.362)
(0.053)
(0.103)
(0.114)
Core {0,1} × CM
-0.466***
-0.054
1.094***
-0.253***
0.060
-0.061
(0.081)
(0.065)
(0.396)
(0.102)
(0.110)
(0.077)
(c) Extended Geographic Controls
(see Table C.30)
Core {0,1} × EAC
-0.211***
-0.012
-0.216**
0.019
0.021
0.054
(0.080)
(0.051)
(0.086)
(0.069)
(0.076)
(0.081)
Core {0,1} × CU
-0.424***
-0.001
0.694**
0.000
0.069
-0.205**
(0.094)
(0.047)
(0.301)
(0.060)
(0.084)
(0.088)
Core {0,1} × CM
-0.304***
-0.076
1.031***
-0.262**
-0.018
0.046
(0.073)
(0.070)
(0.353)
(0.109)
(0.082)
(0.083)
(d) Including Survey Weights
(see Table C.34)
Core {0,1} × EAC
-0.245***
-0.016
-0.168**
0.016
-0.010
0.039
(0.079)
(0.048)
(0.069)
(0.069)
(0.092)
(0.082)
Core {0,1} × CU
-0.443***
0.001
0.910***
0.014
0.012
-0.187*
(0.084)
(0.054)
(0.351)
(0.063)
(0.097)
(0.098)
Core {0,1} × CM
-0.340***
-0.089
0.916**
-0.268**
-0.002
0.032
(0.067)
(0.078)
(0.388)
(0.114)
(0.096)
(0.082)
(e) Excluding Low-Precision Localities
(see Table C.36)
Core {0,1} × EAC
-0.008
0.036
-0.139***
0.074
-0.011
0.336***
(0.092)
(0.036)
(0.030)
(0.058)
(0.092)
(0.123)
Core {0,1} × CU
-0.238**
0.054
0.757**
0.029
-0.067
0.142
(0.108)
(0.043)
(0.366)
(0.070)
(0.147)
(0.134)
Core {0,1} × CM
-0.081
0.019
1.041***
-0.217**
-0.090
0.383***
(0.078)
(0.050)
(0.213)
(0.089)
(0.107)
(0.110)
Afrobarometer
Dependent Variable
Notes: Table continued on next page.
218
Panel b)
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Tables 4.2 and C.9)
Core {0,1} × EAC
-0.261***
-0.013
-0.191***
0.026
0.021
0.041
(0.077)
(0.050)
(0.054)
(0.067)
(0.079)
(0.083)
Core {0,1} × CU
-0.470***
-0.006
0.728**
0.002
0.057
-0.234**
(0.097)
(0.045)
(0.332)
(0.055)
(0.090)
(0.105)
Core {0,1} × CM
-0.338***
-0.076
1.083***
-0.258**
-0.009
0.043
(0.065)
(0.069)
(0.375)
(0.110)
(0.087)
(0.079)
(f) Agglomerations vs. Core
(see Table C.38)
Agglomeration {0,1} × EAC
0.108
-0.056**
0.022
-0.060*
-0.046
-0.199*
(0.078)
(0.028)
(0.082)
(0.034)
(0.079)
(0.115)
Agglomeration {0,1} × CU
-0.036
-0.006
0.129
-0.011
0.001
-0.350***
(0.092)
(0.031)
(0.102)
(0.035)
(0.075)
(0.115)
Agglomeration {0,1} × CM
0.024***
-0.011***
0.242***
-0.063***
0.042***
-0.334***
(0.080)
(0.027)
(0.150)
(0.045)
(0.071)
(0.117)
(g) Core ≤ 25km
(see Table C.42)
Core ≤ 25km {0,1} × EAC
-0.301***
-0.037
-0.208**
0.035
0.009
-0.021
(0.076)
(0.061)
(0.087)
(0.076)
(0.079)
(0.075)
Core ≤ 25km {0,1} × CU
-0.520***
-0.051
0.728*
-0.040
-0.049
-0.244**
(0.098)
(0.061)
(0.423)
(0.071)
(0.058)
(0.105)
Core ≤ 25km {0,1} × CM
-0.375***
-0.120***
1.195***
-0.342***
-0.101***
0.070***
(0.062)
(0.090)
(0.440)
(0.126)
(0.093)
(0.080)
(h) Core ≤ 10km
(see Table C.46)
Core ≤ 10km {0,1} × EAC
-0.309***
-0.090
0.098
0.007
-0.192***
-0.105
(0.095)
(0.060)
(0.132)
(0.057)
(0.066)
(0.076)
Core ≤ 10km {0,1} × CU
-0.496***
-0.055
0.914*
-0.055
-0.209***
-0.341***
(0.102)
(0.044)
(0.529)
(0.041)
(0.066)
(0.112)
Core ≤ 10km {0,1} × CM
-0.368***
-0.126
1.554***
-0.351***
-0.313***
0.091
(0.068)
(0.079)
(0.597)
(0.097)
(0.114)
(0.082)
(i) Post-CM Development
(see Table C.50)
Core {0,1} × EAC
-0.261***
-0.013
-0.189***
0.026
0.021
0.042
(0.077)
(0.050)
(0.052)
(0.067)
(0.078)
(0.082)
Core {0,1} × CU
-0.470***
-0.006
0.731**
0.003
0.056
-0.233**
(0.097)
(0.045)
(0.335)
(0.055)
(0.091)
(0.105)
Core {0,1} × CM 1[2010-201
-0.344***
-0.044
1.263***
-0.256**
-0.081
0.053
(0.066)
(0.080)
(0.409)
(0.123)
(0.088)
(0.094)
Core {0,1} × 1[t ≥ 2015]
-0.332***
-0.097
0.897**
-0.259**
0.037
0.027
(0.066)
(0.070)
(0.393)
(0.104)
(0.071)
(0.082)
Afrobarometer
Dependent Variable
Notes: This table offers an array of robustness tests on the main results of the paper for the difference-in-difference
effect for individuals living in core agglomerations. As such, the results depicted in this table are estimated in the
same way as those shown in Table 4.2 such that most of the respective notes apply here. The full results (including
border estimates) are given in the table referred to below the description of each test. Binary dependent variables
are estimated through a Linear Probability Model (LPM). The standard errors reported allow for spatial correlation,
i.e. Conley standard errors are used, if not indicated otherwise. ***, **, * represents significance at the 1, 5 and 10
percent level, respectively.
219
Table C.4: Robustness Checks Summarized (DHS)
Panel a)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Tables 4.2 and C.9)
Core {0,1} × EAC
0.433***
0.073*
0.146
0.067*
-0.046
-0.002
(0.141)
(0.040)
(0.278)
(0.038)
(0.062)
(0.026)
Core {0,1} × CU
0.314***
0.037
0.770***
0.014
0.058
0.053
(0.105)
(0.039)
(0.177)
(0.043)
(0.111)
(0.033)
Core {0,1} × CM
0.361**
0.006
0.956***
0.061*
0.007
0.029
(0.142)
(0.041)
(0.111)
(0.031)
(0.088)
(0.034)
(a) Standard Errors clustered at EA-Level
(see Table C.23)
Core {0,1} × EAC
0.435***
0.073*
0.146
0.068***
-0.046
-0.002
(0.110)
(0.040)
(0.201)
(0.021)
(0.058)
(0.038)
Core {0,1} × CU
0.315***
0.037
0.770***
0.014
0.058
0.053
(0.097)
(0.032)
(0.241)
(0.020)
(0.052)
(0.035)
Core {0,1} × CM
0.345***
0.006
0.956***
0.049***
0.007
0.029
(0.091)
(0.029)
(0.185)
(0.017)
(0.047)
(0.032)
(b) No Controls
(see Table C.27)
Core {0,1} × EAC
0.334*
0.077
0.071
0.092
-0.098
-0.065
(0.173)
(0.050)
(0.320)
(0.060)
(0.094)
(0.046)
Core {0,1} × CU
0.386***
0.052
0.708***
0.031
0.060
0.029
(0.149)
(0.042)
(0.170)
(0.062)
(0.136)
(0.049)
Core {0,1} × CM
0.375*
0.018
0.914***
0.077*
-0.024
0.009
(0.224)
(0.047)
(0.104)
(0.044)
(0.116)
(0.043)
(c) Extended Geographic Controls
(see Table C.31)
Core {0,1} × EAC
0.483***
0.066*
0.014
0.042
-0.040
0.016
(0.182)
(0.036)
(0.255)
(0.029)
(0.069)
(0.027)
Core {0,1} × CU
0.383***
0.042
0.721***
-0.003
0.071
0.079**
(0.126)
(0.034)
(0.145)
(0.037)
(0.107)
(0.032)
Core {0,1} × CM
0.403**
0.007
0.886***
0.046*
0.008
0.045
(0.157)
(0.041)
(0.094)
(0.026)
(0.091)
(0.030)
(d) Including Survey Weights
(see Table C.35)
Core {0,1} × EAC
0.433***
0.073*
0.146
0.067*
-0.046
-0.002
(0.141)
(0.040)
(0.278)
(0.038)
(0.062)
(0.026)
Core {0,1} × CU
0.314***
0.037
0.770***
0.014
0.058
0.053
(0.105)
(0.039)
(0.177)
(0.043)
(0.111)
(0.033)
Core {0,1} × CM
0.361**
0.006
0.956***
0.061*
0.007
0.029
(0.142)
(0.041)
(0.111)
(0.031)
(0.088)
(0.034)
(e) Excluding Low-Precision Localities
(see Table C.37)
Core {0,1} × EAC
0.439***
0.072*
-0.130
0.065*
-0.039
-0.003
(0.147)
(0.040)
(0.303)
(0.039)
(0.069)
(0.030)
Core {0,1} × CU
0.326***
0.038
0.734**
0.015
0.063
0.052
(0.107)
(0.038)
(0.293)
(0.043)
(0.113)
(0.035)
Core {0,1} × CM
0.342**
0.006
0.950***
0.035*
0.058
0.042
(0.157)
(0.043)
(0.103)
(0.019)
(0.075)
(0.041)
(f) Agglomerations vs. Core
(see Table C.39)
Agglomeration {0,1} × EAC
0.102
-0.044
0.256
0.015
-0.147**
-0.067
(0.153)
(0.043)
(0.226)
(0.025)
(0.060)
(0.046)
Agglomeration {0,1} × CU
-0.033
-0.052*
0.208*
0.034*
-0.117**
-0.029
(0.108)
(0.027)
(0.111)
(0.019)
(0.050)
(0.041)
Agglomeration {0,1} × CM
-0.239***
-0.041**
0.258
0.002
-0.070**
-0.016
(0.090)
(0.017)
(0.169)
(0.022)
(0.033)
(0.031)
0.0077
0.0167
0.1268
0.9331
0.0314
0.6023
Demographic and Health Survey (DHS)
Dependent Variable
Notes: Table continued on next page.
220
Panel b)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Tables 4.2 and C.9)
Core {0,1} × EAC
0.433***
0.073*
0.146
0.067*
-0.046
-0.002
(0.141)
(0.040)
(0.278)
(0.038)
(0.062)
(0.026)
Core {0,1} × CU
0.314***
0.037
0.770***
0.014
0.058
0.053
(0.105)
(0.039)
(0.177)
(0.043)
(0.111)
(0.033)
Core {0,1} × CM
0.361**
0.006
0.956***
0.061*
0.007
0.029
(0.142)
(0.041)
(0.111)
(0.031)
(0.088)
(0.034)
(g) Core ≤ 25km
(see Table C.43)
Core ≤ 25km {0,1} × EAC
0.411**
0.102**
-0.028
0.061
-0.016
0.020
(0.172)
(0.047)
(0.293)
(0.041)
(0.076)
(0.035)
Core ≤ 25km {0,1} × CU
0.220**
0.060
0.759***
0.011
0.040
0.070**
(0.108)
(0.049)
(0.160)
(0.044)
(0.132)
(0.036)
Core ≤ 25km {0,1} × CU
0.296*
-0.004
1.056***
0.061**
-0.023
0.012
(0.162)
(0.049)
(0.180)
(0.030)
(0.084)
(0.040)
(h) Core ≤ 10km
(see Table C.47)
Core ≤ 10km {0,1} × EAC
0.420**
0.031
-0.110
0.062
-0.083
-0.019
(0.175)
(0.043)
(0.330)
(0.041)
(0.058)
(0.062)
Core ≤ 10km {0,1} × CU
0.156
0.030
0.894***
-0.003
-0.008
0.046
(0.108)
(0.029)
(0.249)
(0.035)
(0.125)
(0.062)
Core ≤ 10km {0,1} × CM
0.270*
-0.024
1.637***
0.052**
-0.087
-0.023
(0.149)
(0.027)
(0.600)
(0.024)
(0.054)
(0.075)
(i) "Post-CM" Estimate
(see Table C.51)
Core {0,1} × EAC
0.422***
0.066*
0.121
0.064*
-0.065
-0.004
(0.131)
(0.036)
(0.257)
(0.038)
(0.049)
(0.023)
Core {0,1} × CU
0.310***
0.035
0.755***
0.012
0.047
0.052
(0.104)
(0.038)
(0.184)
(0.044)
(0.109)
(0.032)
Core {0,1} × CM 1[2010-2014]
0.365***
0.009
1.275***
0.080
-0.061
0.062
(0.114)
(0.037)
(0.217)
(0.050)
(0.072)
(0.054)
Core {0,1} × post-CM 1[t ≥ 2015] 0.349**
0.001
0.636***
0.041
0.034
0.006
(0.168)
(0.042)
(0.094)
(0.025)
(0.081)
(0.036)
(j) Excluding post-EAC Migrants
(see Table C.52)
Core {0,1} × EAC
0.605***
0.066
-0.104
0.110**
-0.037
0.027
(0.145)
(0.045)
(0.266)
(0.044)
(0.064)
(0.046)
Core {0,1} × CU
0.375***
0.046
0.760***
0.037
0.034
0.137***
(0.122)
(0.038)
(0.279)
(0.091)
(0.099)
(0.047)
Core {0,1} × CM
0.256**
-0.021
0.734***
0.062
0.023
0.116**
(0.114)
(0.060)
(0.126)
(0.041)
(0.070)
(0.053)
(k) DHS Sample (excl. AIS, KAP, and MIS)
(see Table C.55)
Core {0,1} × EAC
0.433***
0.073*
-0.133
0.096***
-0.046
0.000
(0.141)
(0.040)
(0.270)
(0.027)
(0.062)
(0.027)
Core {0,1} × CU
0.314***
0.037
0.665**
0.019
0.058
0.073*
(0.105)
(0.039)
(0.267)
(0.059)
(0.111)
(0.042)
Core {0,1} × CM
0.361**
0.006
1.045***
0.076**
0.007
0.029
(0.142)
(0.041)
(0.112)
(0.038)
(0.088)
(0.035)
Demographic and Health Survey (DHS)
Dependent Variable
Notes: This table offers an array of robustness tests on the main results of the paper for the difference-in-difference
effect for individuals living in core agglomerations. As such, the results depicted in this table are estimated in the
same way as those shown in Table 4.2 such that most of the respective notes apply here. The full results (including
border estimates) are given in the table referred to below the description of each test. Binary dependent variables
are estimated through a Linear Probability Model (LPM). The standard errors reported allow for spatial
correlation, i.e. Conley standard errors are used, if not indicated otherwise. ***, **, * represents significance at
the 1, 5 and 10 percent level, respectively.
221
Table C.5: Robustness Checks Summarized (KHDS)
Panel a)
Annual p.c.
Consumpt.
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Table 4.3)
Core {0,1} × EAC 1[2004]
170.320***
706.607***
-0.013
0.026***
0.134***
0.891***
(42.328)
(192.323)
(0.023)
(0.008)
(0.021)
(0.070)
Core {0,1} × CU 1[2010]
275.036***
-0.039
0.069
0.891***
(57.368)
(0.025)
(0.067)
(0.099)
(a) Standard Errors clustered at EA-Level
(see Table C.24)
Core {0,1} × EAC 1[2004]
170.320*
706.607
-0.013
0.026
0.134***
0.891***
(100.162)
(1002.964)
(0.033)
(0.018)
(0.021)
(0.126)
Core {0,1} × CU 1[2010]
275.036**
-0.039
0.069
0.891***
(111.403)
(0.040)
(0.067)
(0.132)
(b) No Controls
(see Table C.28)
Core {0,1} × EAC 1[2004]
122.774***
559.391***
-0.012
0.024***
0.109***
1.112***
(44.950)
(40.735)
(0.027)
(0.008)
(0.021)
(0.296)
Core {0,1} × CU 1[2010]
212.121***
-0.021***
0.035***
1.346***
(71.187)
(0.031)
(0.079)
(0.461)
(c) Extended Geographic Controls
(see Table C.32)
Core {0,1} × EAC 1[2004]
172.061***
713.341***
-0.012
0.026***
0.143***
0.836***
(39.808)
(69.683)
(0.024)
(0.009)
(0.025)
(0.030)
Core {0,1} × CU 1[2010]
269.142***
-0.036
0.080
0.846***
(54.303)
(0.028)
(0.072)
(0.072)
(f) Urbanities vs. Core
(see Table C.40)
Urban {0,1} × EAC 1[2004]
180.332***
672.214***
-0.014
0.020**
0.128***
0.619***
(50.703)
(126.314)
(0.026)
(0.009)
(0.024)
(0.140)
Urban {0,1} × CU 1[2010]
267.907***
-0.055*
0.080
0.578***
(64.455)
(0.029)
(0.070)
(0.171)
(g) Core ≤ 25km
(see Table C.44)
Core ≤ 25km × EAC 1[2004]
170.821***
706.607***
-0.013
0.026***
0.132***
0.930***
(42.732)
(192.323)
(0.023)
(0.008)
(0.020)
(0.102)
Core ≤ 25km {0,1} × CU 1[2010]
276.569***
-0.037
0.074
0.987***
(58.235)
(0.026)
(0.068)
(0.148)
(h) Core ≤ 10km
(see Table C.48)
Core ≤ 10km {0,1} × EAC 1[2004] 166.469***
694.026***
-0.010
0.027***
0.110***
1.004***
(42.819)
(149.664)
(0.023)
(0.009)
(0.015)
(0.135)
Core ≤ 10km {0,1} × CU 1[2010]
228.369***
-0.042*
0.009
1.279***
(58.904)
(0.025)
(0.080)
(0.279)
(j) Excluding post-EAC Migrants
(see Table C.53)
Core {0,1} × EAC 1[2004]
119.373***
772.930***
0.009
0.017*
0.143***
0.930***
(42.378)
(91.216)
(0.022)
(0.010)
(0.026)
(0.092)
Core {0,1} × CU 1[2010]
225.295***
-0.049*
0.141*
0.962***
(46.662)
(0.026)
(0.074)
(0.102)
(m) Logged Dependent Variables
(see Table C.56)
Core {0,1} × EAC 1[2004]
0.216*
0.318
-
-
-
1.185***
(0.114)
(0.349)
(0.025)
Core {0,1} × CU 1[2010]
0.306***
-
-
1.052***
(0.113)
(0.032)
Kagera Health and Development Survey (KHDS)
Dependent Variable
Notes: Table continued on next page.
222
Panel b)
Ann. Food
Consumpt.
(TZS '000)
Ann. non-Food
Consumpt.
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Table C.10)
Core {0,1} × EAC 1[2004]
122.188***
35.341
4013.722***
-0.058
8.109
-1.079
(38.913)
(36.287)
(1314.084)
(0.047)
(25.258)
(1.424)
Core {0,1} × CU 1[2010]
131.046***
133.815***
-0.008
-0.404
(39.032)
(39.345)
(0.045)
(1.459)
(a) Standard Erros clustered at EA-Level
(see Table C.25)
Core {0,1} × EAC 1[2004]
122.188**
35.341
4013.722
-0.058
8.109
-1.079
(49.625)
(57.589)
(5982.927)
(0.075)
(27.118)
(1.648)
Core {0,1} × CU 1[2010]
131.046***
133.815*
-0.008
-0.404
(47.557)
(74.509)
(0.051)
(1.636)
(b) No Controls
(see Table C.29)
Core {0,1} × EAC 1[2004]
95.130***
20.114
3358.445***
-0.004
28.256
-0.435
(33.049)
(28.938)
(574.014)
(0.026)
(21.074)
(1.073)
Core {0,1} × CU 1[2010]
109.186***
96.839***
0.065***
-0.479***
(36.483)
(42.831)
(0.039)
(0.952)
(c) Extended Geographic Controls
(see Table C.33)
Core {0,1} × EAC 1[2004]
120.084***
37.350
3971.477***
-0.064
9.471
-1.209
(35.210)
(36.610)
(577.551)
(0.041)
(25.328)
(1.489)
Core {0,1} × CU 1[2010]
127.440***
130.428***
-0.011
-0.689
(37.672)
(35.080)
(0.047)
(1.486)
(d) Urbanities vs. Core
(see Table C.41)
Urban {0,1} × EAC 1[2004]
120.529***
46.673
3842.175***
-0.078*
8.754
-0.760
(35.593)
(38.592)
(1205.886)
(0.040)
(25.455)
(1.370)
Urban {0,1} × CU 1[2010]
142.787***
114.511***
-0.012
-0.299
(42.518)
(38.933)
(0.045)
(1.513)
(e) Core ≤ 25km
(see Table C.45)
Core ≤ 25km × EAC 1[2004]
122.246***
35.774
4013.722***
-0.058
8.109
-0.995
(38.892)
(36.204)
(1314.084)
(0.047)
(25.258)
(1.439)
Core ≤ 25km {0,1} × CU 1[2010]
131.142***
135.227***
-0.007
-0.351
(39.002)
(39.843)
(0.044)
(1.459)
(f) Core ≤ 10km
(see Table C.49)
Core ≤ 10km {0,1} × EAC 1[2004] 117.794***
36.303
3592.924
-0.058
7.206
-0.380
(34.197)
(45.837)
(2577.006)
(0.042)
(19.772)
(1.427)
Core ≤ 10km {0,1} × CU 1[2010]
116.420***
101.922**
-0.001
-0.116
(36.746)
(42.044)
(0.037)
(1.467)
(j) Excluding post-EAC Migrants
(see Table C.54)
Core {0,1} × EAC 1[2004]
106.040***
0.311
4599.535***
-0.077
9.488
-0.456
(36.417)
(16.952)
(1202.895)
(0.071)
(25.859)
(2.988)
Core {0,1} × CU 1[2010]
140.531***
74.285**
0.009
-0.171
(35.606)
(36.188)
(0.047)
(2.939)
(m) Logged Values
(see Table C.57)
Core {0,1} × EAC 1[2004]
0.271*
0.025
0.547
-0.058
0.269
-1.079
(0.140)
(0.139)
(0.566)
(0.047)
(0.495)
(1.424)
Core {0,1} × CU 1[2010]
0.300**
0.154
-0.008
-0.404
(0.126)
(0.180)
(0.045)
(1.459)
Kagera Health and Development Survey (KHDS)
Dependent Variable
Notes: This table offers an array of robustness tests on the main results of the paper for the difference-in-difference
effect for individuals living in core agglomerations. As such, the results depicted in this table are estimated in the
same way as those shown in Table 4.3 such that most of the respective notes apply here. The full results (including
border estimates) are given in the table referred to below the description of each test. Binary dependent variables
are estimated through a Linear Probability Model (LPM). The standard errors reported allow for spatial
correlation, i.e. Conley standard errors are used, if not indicated otherwise. ***, **, * represents significance at the
1, 5 and 10 percent level, respectively.
223
Table C.6: Robustness Checks Summarized (Afrobarometer)
Panel a)
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Tables 4.2 and C.9)
Border [0,1] × EAC
-0.083
0.050
0.098
0.072
0.017
-0.350
(0.296)
(0.058)
(0.071)
(0.077)
(0.134)
(0.294)
Border [0,1] × CU
0.172
0.060
0.064
0.091
0.040
-0.234
(0.292)
(0.065)
(0.081)
(0.073)
(0.119)
(0.312)
Border [0,1] × CM
0.142
0.098**
0.119
-0.001
0.289***
-0.456
(0.282)
(0.039)
(0.101)
(0.076)
(0.087)
(0.291)
(a) Beeline Distance to Borderpost
(see Table C.58)
(Beeline to) Border [0,1] × EAC
0.096
0.022
0.080
0.012
-0.052
-0.421
(0.295)
(0.054)
(0.062)
(0.068)
(0.131)
(0.325)
(Beeline to) Border [0,1] × CU
0.201
0.097
0.177
0.113
0.127
-0.211
(0.327)
(0.063)
(0.124)
(0.082)
(0.107)
(0.346)
(Beeline to) Border [0,1] × CM
0.221
0.063
0.264
-0.051
0.256***
-0.433
(0.259)
(0.041)
(0.168)
(0.075)
(0.088)
(0.288)
(b) Beeline Distance to Borderline
(see Table C.62)
(Beeline to) Borderline [0,1] × EAC
-0.045
0.062
0.098*
0.078
0.054
-0.402
(0.282)
(0.058)
(0.060)
(0.077)
(0.129)
(0.297)
(Beeline to) Borderline [0,1] × CU
0.144
0.051
0.075
0.072
0.025
-0.235
(0.287)
(0.059)
(0.079)
(0.071)
(0.108)
(0.309)
(Beeline to) Borderline [0,1] × CM
0.213
0.074**
0.131
-0.054
0.262***
-0.432
(0.258)
(0.037)
(0.106)
(0.071)
(0.080)
(0.284)
(c) Logged Distance (+1)
(see Table C.66)
Log(1+BorderDistance) × EAC
0.029
-0.006
-0.018
-0.006
0.015
0.091
(0.071)
(0.014)
(0.022)
(0.019)
(0.031)
(0.069)
Log(1+BorderDistance) × CU
-0.033
-0.023
-0.014
-0.027
-0.015
0.101
(0.073)
(0.018)
(0.032)
(0.023)
(0.031)
(0.074)
Log(1+BorderDistance) × CM
-0.009
-0.027***
-0.032
0.007
-0.068***
0.143**
(0.067)
(0.010)
(0.039)
(0.021)
(0.020)
(0.066)
(d) Border Dummy ≤ 100km
(see Table C.70)
Border ≤ 100km {0,1} × EAC
-0.137
0.002
-0.022
0.011
0.007
-0.125
(0.135)
(0.023)
(0.041)
(0.039)
(0.066)
(0.165)
Border ≤ 100km {0,1} × CU
-0.019
0.034
-0.070
0.075
0.116*
-0.179
(0.156)
(0.035)
(0.050)
(0.051)
(0.065)
(0.169)
Border ≤ 100km {0,1} × CM
-0.032
0.025
-0.035
-0.017
0.153***
-0.201
(0.131)
(0.022)
(0.056)
(0.048)
(0.058)
(0.153)
(e) Border Dummy ≤ 50km
(see Table C.74)
Border ≤ 50km {0,1} × EAC
-0.190
0.029
0.047*
0.009
-0.093
0.078
(0.149)
(0.024)
(0.026)
(0.032)
(0.061)
(0.116)
Border ≤ 50km {0,1} × CU
-0.035
0.095***
0.005
0.050
-0.019
0.039
(0.122)
(0.032)
(0.039)
(0.058)
(0.096)
(0.146)
Border ≤ 50km {0,1} × CM
-0.177
0.077***
0.050*
0.012
0.091
-0.129
(0.145)
(0.025)
(0.029)
(0.058)
(0.056)
(0.127)
(f) Border Dummy ≤ 25km
(see Table C.78)
Border ≤ 25km {0,1} × EAC
-0.478**
-0.020
0.039
-0.034
-0.176
-0.381
(0.225)
(0.048)
(0.039)
(0.054)
(0.151)
(0.288)
Border ≤ 25km {0,1} × CU
-0.029
0.150***
-0.002
0.093
0.007
-0.324
(0.297)
(0.050)
(0.049)
(0.066)
(0.156)
(0.269)
Border ≤ 25km {0,1} × CM
-0.215
0.034
-0.019
-0.097
-0.083
-0.380***
(0.255)
(0.043)
(0.055)
(0.079)
(0.107)
(0.146)
Afrobarometer
Dependent Variable
Notes: This table offers robustness tests on the main results of the paper for the difference-in-difference effect with
treatment intensity increasing towards the border. The results depicted in this table are estimated in the same way as
those shown in Table 4.2 such that most of the respective notes apply here. The full results (including border estimates)
are given in the table referred to below the description of each test.
224
Table C.7: Robustness Checks Summarized (DHS)
Panel a)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Tables 4.2 and C.9)
Border [0,1] × EAC
-0.284
-0.129
0.104
0.114*
-0.184
-0.162
(0.312)
(0.089)
(0.136)
(0.059)
(0.188)
(0.136)
Border [0,1] × CU
-0.372*
-0.138**
0.084
0.118***
-0.291***
-0.271***
(0.224)
(0.054)
(0.089)
(0.043)
(0.106)
(0.083)
Border [0,1] × CM
-0.547***
-0.034
0.060
0.081*
-0.149**
-0.141
(0.202)
(0.040)
(0.095)
(0.046)
(0.075)
(0.092)
(a) Beeline Distance to Borderpost
(see Table C.59)
(Beeline to) Border [0,1] × EAC
-0.160
-0.141
-0.042
0.076
-0.179
-0.131
(0.312)
(0.091)
(0.139)
(0.054)
(0.194)
(0.135)
(Beeline to) Border [0,1] × CU
-0.277
-0.135***
0.067
0.121***
-0.207*
-0.141
(0.265)
(0.045)
(0.143)
(0.043)
(0.118)
(0.101)
(Beeline to) Border [0,1] × CM
-0.550**
-0.066**
-0.027
0.083*
-0.178**
-0.182**
(0.232)
(0.033)
(0.138)
(0.045)
(0.071)
(0.092)
(b) Beeline Distance to Borderline
(see Table C.63)
(Beeline to) Borderline [0,1] × EAC
-0.248
-0.152**
0.026
0.079
-0.168
-0.152
(0.296)
(0.077)
(0.132)
(0.055)
(0.172)
(0.123)
(Beeline to) Borderline [0,1] × CU
-0.368
-0.106**
0.084
0.112**
-0.211*
-0.220***
(0.235)
(0.049)
(0.107)
(0.044)
(0.108)
(0.084)
(Beeline to) Borderline [0,1] × CM -0.559***
-0.024
-0.005
0.063
-0.102
-0.131
(0.199)
(0.036)
(0.109)
(0.045)
(0.064)
(0.083)
(c) Logged Distance (+1)
(see Table C.67)
Log(1+BorderDistance) × EAC
0.029
0.034**
-0.077*
-0.028*
0.018
0.046*
(0.076)
(0.016)
(0.041)
(0.016)
(0.035)
(0.026)
Log(1+BorderDistance) × CU
0.083
0.040***
-0.053*
-0.024**
0.048
0.075***
(0.057)
(0.013)
(0.032)
(0.011)
(0.030)
(0.023)
Log(1+BorderDistance) × CM
0.120**
0.016
-0.066*
-0.015
0.035**
0.048*
(0.050)
(0.010)
(0.039)
(0.012)
(0.018)
(0.025)
(d) Border Dummy ≤ 100km
(see Table C.71)
Border ≤ 100km {0,1} × EAC
-0.421
-0.134***
-0.146*
0.023
-0.181***
-0.204***
(0.261)
(0.035)
(0.084)
(0.038)
(0.063)
(0.025)
Border ≤ 100km {0,1} × CU
-0.345**
-0.099***
-0.051
0.038
-0.144***
-0.177***
(0.167)
(0.023)
(0.058)
(0.023)
(0.048)
(0.034)
Border ≤ 100km {0,1} × CM
-0.379**
-0.065***
-0.056
0.024
-0.133***
-0.159***
(0.150)
(0.015)
(0.056)
(0.025)
(0.034)
(0.049)
(e) Border Dummy ≤ 50km
(see Table C.75)
Border ≤ 50km {0,1} × EAC
0.040
-0.119**
-0.055
0.037
-0.086
-0.093***
(0.214)
(0.050)
(0.078)
(0.041)
(0.061)
(0.033)
Border ≤ 50km {0,1} × CU
-0.048
-0.135***
0.038
0.010
-0.104**
-0.145***
(0.154)
(0.034)
(0.064)
(0.028)
(0.047)
(0.043)
Border ≤ 50km {0,1} × CM
-0.092
-0.093***
0.013
0.020
-0.117***
-0.121*
(0.172)
(0.024)
(0.045)
(0.026)
(0.023)
(0.063)
(f) Border Dummy ≤ 25km
(see Table C.79)
Border ≤ 25km {0,1} × EAC
0.548***
-0.051***
-0.083
0.090*
-0.010
-0.064
(0.121)
(0.018)
(0.088)
(0.049)
(0.072)
(0.044)
Border ≤ 25km {0,1} × CU
0.261***
-0.068***
-0.004
0.039
-0.025
-0.158***
(0.074)
(0.025)
(0.110)
(0.075)
(0.090)
(0.044)
Border ≤ 25km {0,1} × CM
0.244
-0.031
-0.036
0.066
-0.063
-0.088**
(0.167)
(0.026)
(0.092)
(0.043)
(0.062)
(0.038)
0.1443
0.2347
0.6937
0.1218
0.3106
0.0218
Demographic and Health Survey (DHS)
Dependent Variable
Notes: This table offers robustness tests on the main results of the paper for the difference-in-difference effect
with treatment intensity increasing towards the border. The results depicted in this table are estimated in
the same way as those shown in Table 4.2 such that most of the respective notes apply here. The full results
(including border estimates) are given in the table referred to below the description of each test.
225
Table C.8: Robustness Checks Summarized (KHDS)
Panel a)
Annual p.c.
Consumpt.
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Table 4.3)
Border [0,1] × EAC
-853.376
187.152
0.581*
0.071
0.107
-0.848**
(1072.010)
(754.481)
(0.331)
(0.049)
(0.261)
(0.384)
Border [0,1] × CU
-1317.845
0.432
-0.010
-1.631***
(1154.386)
(0.305)
(0.404)
(0.586)
(a) Beeline Distance to Borderpost
(see Table C.60)
(Beeline to) Border [0,1] × EAC
-816.913
69.400
0.563**
0.051
0.078
-0.710**
(1112.207)
(715.743)
(0.279)
(0.044)
(0.238)
(0.312)
(Beeline to) Border [0,1] × CU
-1283.327
0.435
-0.033
-1.567***
(1208.457)
(0.278)
(0.388)
(0.545)
(b) Beeline Distance to Borderline
(see Table C.64)
(Beeline to) Borderline [0,1] × EAC
-757.039
122.074
0.490*
0.041
0.062
-0.570*
(998.849)
(577.684)
(0.261)
(0.044)
(0.234)
(0.318)
(Beeline to) Borderline [0,1] × CU
-1199.598
0.389
-0.043
-1.360***
(1090.390)
(0.255)
(0.374)
(0.500)
(c) Logged Distance (+1)
(see Table C.68)
Log(1+Borderdistance) × EAC
-14.571
-305.802
-0.051
-0.020*
-0.071
-0.404*
(96.504)
(499.702)
(0.043)
(0.011)
(0.052)
(0.225)
Log(1+Borderdistance) × CU
24.227
-0.033
-0.040
-0.330
(115.177)
(0.038)
(0.071)
(0.221)
(d) Border Dummy ≤ 100km
(see Table C.72)
Border ≤ 100km {0,1} × EAC
-43.826
21.219
0.033
0.009
0.017
0.038
(85.443)
(83.145)
(0.042)
(0.010)
(0.045)
(0.063)
Border ≤ 100km {0,1} × CU
-105.265
0.022
0.127
-0.085
(108.035)
(0.041)
(0.089)
(0.102)
(e) Border Dummy ≤ 50km
(see Table C.76)
Border ≤ 50km {0,1} × EAC
49.969
-2.821
-0.004
0.035***
0.047
-0.143*
(292.688)
(77.036)
(0.035)
(0.009)
(0.051)
(0.078)
Border ≤ 50km {0,1} × CU
-13.598
-0.033
-0.022
-0.164*
(306.353)
(0.033)
(0.149)
(0.096)
(f) Border Dummy ≤ 25km
(see Table C.80)
Border ≤ 25km {0,1} × EAC
1046.331***
-242.913
-0.162
0.010
0.188
0.034
(310.431)
(436.378)
(0.122)
(0.030)
(0.210)
(0.066)
Border ≤ 25km {0,1}× CU
1126.525***
-0.037
0.583*
-0.080
(354.396)
(0.059)
(0.312)
(0.128)
0.0015
0.5288
0.0615
0.5315
Kagera Health and Development Survey (KHDS)
Dependent Variable
Notes: Table continued on next page.
226
Panel b)
Ann. Food
Consumpt.
(TZS '000)
Ann. non-Food
Consumpt.
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Baseline Coefficients
(see Table C.10)
Border [0,1] × EAC
-177.749
-670.444
7592.679
-0.201
-96.031
5.949
(237.164)
(1054.004)
(14152.494)
(0.377)
(137.206)
(9.297)
Border [0,1] × CU
-388.178
-911.118
-0.233
6.555
(236.410)
(1116.362)
(0.344)
(9.162)
(a) Beeline Distance to Borderpost
(see Table C.61)
(Beeline to) Border [0,1] × EAC
-111.273
-697.103
4638.227
-0.180
-106.417
3.003
(210.184)
(1104.755)
(12272.383)
(0.339)
(145.479)
(8.924)
(Beeline to) Border [0,1] × CU
-335.069
-924.540
-0.234
3.458
(212.127)
(1172.534)
(0.319)
(8.728)
(b) Beeline Distance to Borderline
(see Table C.65)
(Beeline to) Borderline [0,1] × EAC
-107.710
-638.292
4075.452
-0.135
-92.091
4.749
(207.271)
(981.794)
(11014.623)
(0.327)
(133.728)
(8.546)
(Beeline to) Borderline [0,1] × CU
-323.756
-853.098
-0.203
5.328
(206.881)
(1047.819)
(0.302)
(8.284)
(c) Logged Distance (+1)
(see Table C.69)
Log(1+Borderdistance) × EAC
-22.602
13.234
-3084.228
0.014
9.795
-1.344
(56.181)
(62.018)
(3506.668)
(0.060)
(19.092)
(1.748)
Log(1+Borderdistance) × CU
17.695
8.645
0.025
-1.897
(62.345)
(68.685)
(0.041)
(1.590)
(d) Border Dummy ≤ 100km
(see Table C.73)
Border ≤ 100km {0,1} × EAC
-5.045
-38.379
1207.398
-0.019
-8.716
2.250
(46.765)
(67.330)
(1921.629)
(0.068)
(26.978)
(1.438)
Border ≤ 100km {0,1} × CU
-36.354
-68.217
-0.040
2.278*
(50.035)
(86.151)
(0.064)
(1.312)
(e) Border Dummy ≤ 50km
(see Table C.77)
Border ≤ 50km {0,1} × EAC
0.159
46.895
1521.906
0.052
13.350
-2.628
(186.972)
(107.532)
(1408.312)
(0.089)
(24.030)
(2.832)
Border ≤ 50km {0,1} × CU
-33.556
20.242
0.016
-2.064
(215.623)
(97.817)
(0.058)
(2.874)
(f) Border Dummy ≤ 25km
(see Table C.81)
Border ≤ 25km {0,1} × EAC
536.903***
493.731***
3041.987
-0.033
-487.518
-0.755
(166.385)
(181.678)
(3654.443)
(0.139)
(5036999.421)
(2.051)
Border ≤ 25km {0,1}× CU
663.445***
452.421***
-0.104
0.380
(233.953)
(150.728)
(0.121)
(2.099)
0.0046
0.0027
0.3907
0.8563
Kagera Health and Development Survey (KHDS)
Dependent Variable
Notes: This table offers robustness tests on the main results of the paper for the difference-in-difference effect with
treatment intensity increasing towards the border. The results depicted in this table are estimated in the same
way as those shown in Table 4.3 such that most of the respective notes apply here. The full results (including
border estimates) are given in the table referred to below the description of each test.
227
Table C.9: Further Labor Market Outcomes (AFB and DHS)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[0.55]
[1.66]
[2.09]
[0.76]
[1.51]
[0.52]
Border [0,1] × EAC 1[2002-2004]
0.072
0.017
-0.350
0.114*
-0.184
-0.162
(0.077)
(0.134)
(0.294)
(0.059)
(0.188)
(0.136)
0.347
0.901
0.234
0.054
0.328
0.231
Border [0,1] × CU 1[2005-2009]
0.091
0.040
-0.234
0.118***
-0.291***
-0.271***
(0.073)
(0.119)
(0.312)
(0.043)
(0.106)
(0.083)
0.214
0.734
0.454
0.006
0.006
0.001
Border [0,1] × CM 1[t ≥ 2010]
-0.001
0.289***
-0.456
0.081*
-0.149**
-0.141
(0.076)
(0.087)
(0.291)
(0.046)
(0.075)
(0.092)
0.987***
0.001
0.117***
0.079
0.047
0.125***
Core {0,1} × EAC 1[2002-2004]
0.026
0.021
0.041
0.067*
-0.046
-0.002
(0.067)
(0.079)
(0.083)
(0.038)
(0.062)
(0.026)
0.695
0.786
0.615
0.075
0.454
0.948
Core {0,1} × CU 1[2005-2009]
0.002
0.057
-0.234**
0.014
0.058
0.053
(0.055)
(0.090)
(0.105)
(0.043)
(0.111)
(0.033)
0.965
0.529
0.026
0.750
0.597
0.106
Core {0,1} × CM 1[t ≥ 2010]
-0.258**
-0.009
0.043
0.061*
0.007
0.029
(0.110)
(0.087)
(0.079)
(0.031)
(0.088)
(0.034)
0.0189
0.9170
0.5844
0.0519
0.9383
0.3979
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
26,563
21,232
35,975
169,875
95,717
98,963
R-Squared
0.23
0.25
0.09
0.23
0.26
0.17
R-Squared -Within
0.23
0.25
0.09
0.23
0.26
0.17
Dependent Variable
Afrobarometer
DHS
Income
Notes: The results in each column are produced by a separate regression. The sample mean of the respective
dependent variable is given in brackets above the estimates. Binary dependent variables are estimated through a
Linear Probability Model (LPM). The standard errors reported allow for spatial correlation, i.e. Conley standard
errors are used. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes below
Table 4.2.
228
Table C.10: Further Results (KHDS)
Agglom.
Ann. Food
Consumpt.
(TZS '000)
Ann. non-Food
Consumpt.
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border [0,1] × EAC 1[2004]
-177.749
-670.444
7592.679
-0.201
-96.031
5.949
(237.164)
(1054.004)
(14152.494)
(0.377)
(137.206)
(9.297)
0.454
0.525
0.5920
0.595
0.4843
0.522
Border [0,1] × CU 1[2010]
-388.178
-911.118
-0.233
6.555
(236.410)
(1116.362)
(0.344)
(9.162)
0.101
0.415
0.498***
0.474
Core {0,1} × EAC 1[2004]
122.188***
35.341
4013.722***
-0.058
8.109
-1.079
(38.913)
(36.287)
(1314.084)
(0.047)
(25.258)
(1.424)
0.002
0.330
0.0025
0.220
0.7483
0.449
Core {0,1} × CU 1[2010]
131.046***
133.815***
-0.008
-0.404
(39.032)
(39.345)
(0.045)
(1.459)
0.0008
0.0007
0.8632
0.7819
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.10
0.09
0.07
0.16
0.18
0.09
Dependent Variable
Kagera Health and Development Survey (KHDS)
Consumpt.
Income
Notes: The results in each column are produced by a separate regression. The sample mean of the respective
dependent variable is given in brackets above the estimates. Binary dependent variables are estimated through a
Linear Probability Model (LPM). The standard errors reported allow for spatial correlation, i.e. Conley standard
errors are used. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes below
Table 4.3.
229
Table C.11: Urbanities in Border Regions – Triple Difference Estimates
Consumpt.
Income
Agglom.
Consumpt.
Income
Agglom.
Panel a)
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[3.14]
[0.19]
[0.00]
Border [0,1] × EAC
0.032
0.093**
-0.045
-0.159
-0.030
-0.048
(0.300)
(0.043)
(0.088)
(0.236)
(0.032)
(0.070)
Triple Interaction
0.916*
0.031
0.614
0.499
0.346***
0.492
Border [0,1] × EAC × Urban {0,1}
0.498
-0.147
0.509
-0.356
-0.004
0.649*
(0.368)
(0.092)
(0.448)
(0.313)
(0.071)
(0.343)
Isolated Effect of the EAC on Urbanities in Border Regions
Combined Effect:
Border [0,1] × EAC + Triple Interaction
0.530
-0.054
0.464
-0.515**
-0.035
0.601*
[0.15]
[0.55]
[0.25]
[0.02]
[0.65]
[0.07]
Observations
36,042
25,465
4,156
183,250
71,738
7,692
R-Squared
0.13
0.15
0.29
0.46
0.18
0.34
R-Squared -Within
0.11
0.11
0.27
0.45
0.13
0.33
Panel b)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[0.55]
[1.66]
[1.12]
[0.76]
[1.51]
[0.52]
Border [0,1] × EAC
0.054
0.207**
-0.378
0.074
-0.045
-0.078
(0.064)
(0.099)
(0.321)
(0.047)
(0.047)
(0.086)
Triple Interaction
0.398***
0.037
0.240**
0.114**
0.336
0.362
Border [0,1] × EAC × Urban {0,1}
-0.146
-0.383
0.110
0.058
-0.273*
-0.221**
(0.147)
(0.284)
(0.349)
(0.096)
(0.141)
(0.103)
Isolated Effect of the EAC on Urbanities in Border Regions
Combined Effect:
Border [0,1] × EAC + Triple Interaction
-0.092
-0.177
-0.268
0.131
-0.318**
-0.299***
[0.53]
[0.51]
[0.39]
[0.13]
[0.02]
[0.00]
Observations
26,563
21,232
35,975
169,875
95,717
98,963
R-Squared
0.23
0.29
0.10
0.23
0.32
0.21
R-Squared -Within
0.08
0.26
0.07
0.20
0.24
0.12
Individual Controls
YES
YES
NO/YES
YES
YES
NO/YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Dependent Variable
Afrobarometer
DHS
Afrobarometer
DHS
Income
Notes: This table analyzes the differential effect of Border [0,1] for individuals living in urban regions. Row one
shows the uninteracted effect of Border [0,1], i.e. effect of the EAC for individuals at rural border regions, row two
shows the differential effect for being in an urban area. Row three depicts the combined effect of the two constituent
terms, i.e. the effect of Border [0,1] for individuals in urban regions. The results in each column and panel are
produced by a separate regression. The sample mean of the respective dependent variable is given in brackets
above the estimates. Binary dependent variables are estimated through a Linear Probability Model (LPM). The
standard errors reported allow for spatial correlation, i.e. Conley standard errors are used. ***, **, * represents
significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 4.2.
230
Table C.12: Placebo Test – Disaggregated DiD in contiguous Countries
Consumpt.
Income
Agglom.
Consumpt.
Income
Agglom.
Panel a)
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.25]
[0.24]
[0.00]
[3.17]
[0.17]
[0.00]
Border [0,1] × EAC 1[2002-2004
0.274*
-0.030
-0.199
0.750***
-0.049
-0.467**
(0.161)
(0.084)
(0.222)
(0.161)
(0.073)
(0.205)
Border [0,1] × CU 1[2005-2009]
0.036
0.165**
-0.120
1.349***
0.080
0.181
(0.133)
(0.082)
(0.150)
(0.178)
(0.058)
(0.205)
Border [0,1] × CM 1[t ≥ 2010]
0.179
0.056
0.063
0.740***
0.000
-0.165
(0.143)
(0.076)
(0.164)
(0.140)
(0.037)
(0.120)
Core {0,1} × EAC 1[2002-2004]
-0.114
-0.033
0.580*
-0.757**
-0.245*
-1.551*
(0.145)
(0.042)
(0.310)
(0.331)
(0.135)
(0.844)
Core {0,1} × CU 1[2005-2009]
-0.025
-0.008
0.482*
-0.004
-0.076
-0.411
(0.069)
(0.036)
(0.261)
(0.180)
(0.116)
(0.547)
Core {0,1} × CM 1[t ≥ 2010]
-0.159
-0.056
0.942***
0.214
-0.053
0.123
(0.109)
(0.048)
(0.320)
(0.175)
(0.073)
(0.304)
0.146
0.244
0.003
0.220
0.472
0.686
Observations
28,541
18,994
2,329
200,133
126,587
7,819
R-Squared
0.09
0.19
0.24
0.29
0.14
0.28
R-Squared -Within
0.06
0.15
0.22
0.28
0.14
0.27
Panel b)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[0.49]
[1.62]
[2.12]
[0.67]
[1.48]
[0.44]
Border [0,1] × EAC 1[2002-2004
-0.117
0.001
-0.575***
-0.058
-0.058
-0.007
(0.090)
(0.170)
(0.220)
(0.127)
(0.210)
(0.250)
Border [0,1] × CU 1[2005-2009]
0.107
-0.012
-0.839***
-0.063
0.076
-0.032
(0.081)
(0.156)
(0.149)
(0.073)
(0.108)
(0.126)
Border [0,1] × CM 1[t ≥ 2010]
0.037
0.058
-0.462***
-0.014
-0.081
-0.120
(0.081)
(0.133)
(0.155)
(0.080)
(0.107)
(0.121)
Core {0,1} × EAC 1[2002-2004]
0.034
0.028
-0.365**
0.130**
-0.268**
-0.246
(0.058)
(0.059)
(0.169)
(0.057)
(0.129)
(0.160)
Core {0,1} × CU 1[2005-2009]
0.093**
0.001
-0.448***
0.062
0.013
0.016
(0.041)
(0.076)
(0.044)
(0.060)
(0.073)
(0.096)
Core {0,1} × CM 1[t ≥ 2010]
-0.106
-0.040
-0.320***
0.094**
0.075
0.064
(0.070)
(0.085)
(0.082)
(0.047)
(0.070)
(0.094)
0.130
0.637
0.000
0.044
0.283
0.493
Observations
19,191
12,715
28,274
286,651
187,498
188,349
R-Squared
0.18
0.31
0.11
0.17
0.24
0.17
R-Squared -Within
0.18
0.31
0.11
0.17
0.24
0.17
Individual Controls
YES
YES
NO/YES
YES
YES
NO/YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Income
Dependent Variable
Afrobarometer
DHS
Afrobarometer
DHS
Notes: This table conducts a “placebo” analysis by testing for a spatially differentiated effect across contiguous,
non-EAC countries within the time frame of the EAC’s establishment and expansion. Border [0,1] × EAC 1[2002-
2004] switches to one for individuals sampled from 2002 to and including 2004, CU 1[2005-2009] for individuals
sampled from 2005 and including 2009, and CM 1[t ≥ 2010] for individuals sampled from 2010 onwards. See full
notes in Table 4.4.
231
Table C.13: Region-Based Estimates (Placebo Countries)
Consumpt.
Income
Agglom.
Panel a)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.20]
[0.17]
[0.00]
[0.67]
[1.48]
[0.44]
Border Region {0,1} × EAC 1[t ≥ 2002]
0.230*
-0.004
0.001
-0.022
-0.008
-0.087*
EAC 0[1992-2000]
(0.111)
(0.013)
(0.141)
(0.016)
(0.047)
(0.048)
0.058
0.777***
0.997
0.196**
0.860***
0.094
Core Region {0,1} × EAC 1[t ≥ 2002]
-0.159
-0.056
1.066
0.099**
-0.086
-0.072
EAC 0[1992-2000]
(0.171)
(0.034)
(0.685)
(0.039)
(0.066)
(0.059)
0.3685
0.1247
0.1421
0.0224
0.2162
0.2422
Observations
282,480
167,590
11,391
348,016
225,336
223,908
R-Squared
0.27
0.15
0.37
0.15
0.27
0.20
R-Squared -Within
0.25
0.14
0.35
0.13
0.23
0.13
Panel b)
Border Region {0,1} × pre-EAC 1[1996-2000]
-0.225
-0.004
0.041
-
0.177
0.014
pre-EAC 0[1992-1995]
(0.152)
(0.020)
(0.128)
(0.130)
(0.062)
0.160***
0.847***
0.751***
-
0.193
0.820***
Border Region {0,1} × EAC 1[t ≥ 2002]
0.041
-0.009
0.035
-0.022
0.139
-0.077
EAC 0[1992-1995]
(0.217)
(0.017)
(0.202)
(0.016)
(0.116)
(0.045)
0.855**
0.615***
0.864
0.196*
0.252
0.108
Core Region {0,1} × pre-EAC 1[1996-2000]
0.346
0.088
0.780
-0.107**
0.211**
0.188**
pre-EAC 0[1992-1995]
(0.308)
(0.072)
(0.644)
(0.038)
(0.096)
(0.088)
0.279***
0.239***
0.246
0.014*
0.045
0.050
Core Region {0,1} × EAC 1[t ≥ 2002]
0.107
0.011
1.608*
-0.008
0.072
0.078
EAC 0[1992-1995]
(0.260)
(0.050)
(0.868)
(0.025)
(0.049)
(0.085)
0.686
0.822
0.085
0.755
0.163
0.370
Observations
282,480
167,590
11,391
348,016
225,336
223,908
R-Squared
0.27
0.15
0.38
0.15
0.27
0.20
R-Squared -Within
0.26
0.14
0.35
0.13
0.23
0.13
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
NO
NO
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Income
Dependent Variable
DHS (Region-based)
Notes: This table conducts a “placebo” analysis by testing for a spatially differentiated effect across contiguous,
non-EAC countries within the time frame of the EAC's establishment and expansion. Specifically, the analysis
makes use of the non-GPS survey rounds of the Demographic and Health Surveys (DHS) sampled before 1999 and
additionally conducts “pre-tests” towards the difference-in-differences approach. The data thereby come from the
full sample of Ethiopia, Malawi, Mozambique, Rwanda and Zambia DHS surveys sampled between 1992 and 2019,
making use of AIS, KAP and MIS rounds as well. The sample mean of the respective dependent variable is given
in brackets above the estimates. Border Region {0,1} switches to one for individuals living in a region with a median
road distance to the nearest border crossing of a contiguous EAC country below the 10th percentile of all (within-
country) GPS-border distances in the sample. Core Region {0,1} is a dummy indicating individuals living in the
region which hosts the core agglomeration of their respective country (i.e. Addis Abeba, Kigali, Lilongwe, Lusaka,
Maputo). EAC 1[t ≥ 2002] switches to one for individuals sampled from 2002 onwards. Pre-EAC 1[1996-2000]
switches to one for individuals sampled in survey years between 1996 and including 2000. As such, in panel a), the
reference group of the estimates are comprised of individuals sampled in the full pre-EAC period, i.e. from 1991 to
2000, while in panel b), the reference group is formed by individuals sampled between 1991 and including 1995.
Hence, the DiD estimate on “pre-EAC” in panel b) represents the pre-test. The results in each column and panel
are produced by a separate regression. All regressions include individual-level controls for respondents’ age, gender,
as well as education. The regressions also include country-year fixed effects. Binary dependent variables are
estimated through a Linear Probability Model (LPM). The standard errors reported allow for clustering at the
region level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
232
Table C.14: Disaggregated Region-Based Estimates
Consumpt.
Income
Agglom.
Panel a)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.12]
[0.19]
[0.00]
[0.76]
[1.52]
[0.53]
Border Region {0,1} × EAC 1[2002-2004]
0.113
-0.037
-0.050
-0.091
-0.051
-0.125**
EAC 0[1988-2000]
(0.199)
(0.024)
(0.145)
(0.092)
(0.054)
(0.058)
Border Region {0,1} × CU 1[2005-2009]
0.207
-0.011
-0.215*
-0.120**
0.109
-0.009
CU 0[1988-2000]
(0.218)
(0.029)
(0.125)
(0.054)
(0.077)
(0.070)
Border Region {0,1} × CM 1[t ≥ 2010]
-0.044
-0.035
-0.086
-0.068
-0.063
-0.155**
CM 0[1988-2000]
(0.178)
(0.032)
(0.144)
(0.058)
(0.065)
(0.064)
0.805
0.289**
0.555
0.257***
0.342***
0.021***
Core Region {0,1} × EAC 1[2002-2004]
0.633***
0.114***
3.017***
-0.064
0.101***
-0.005
EAC 0[1988-2000]
(0.116)
(0.022)
(0.517)
(0.065)
(0.032)
(0.075)
Core Region {0,1} × CU 1[2005-2009]
0.399***
0.100***
1.524
-0.094
0.162
0.060**
CU 0[1988-2000]
(0.131)
(0.026)
(1.395)
(0.063)
(0.095)
(0.025)
Core Region {0,1} × CM 1[t ≥ 2010]
0.316**
0.077***
2.195
-0.006
0.062
-0.024
CM 0[1988-2000]
(0.139)
(0.024)
(1.581)
(0.047)
(0.044)
(0.029)
0.030
0.003
0.175
0.894
0.171
0.415
Observations
258,820
104,440
11,841
236,646
140,613
136,163
R-Squared
0.28
0.16
0.55
0.21
0.25
0.18
R-Squared -Within
0.27
0.10
0.53
0.18
0.16
0.08
Panel b)
Border Region {0,1} × pre-EAC 1[1996-2000]
0.098
0.024
0.084
-0.296**
0.014
-0.063
pre-EAC 0[1988-1995]
(0.125)
(0.029)
(0.082)
(0.121)
(0.060)
(0.049)
Border Region {0,1} × EAC 1[2002-2004]
0.181
-0.022
-0.014
-0.314**
-0.041
-0.163***
EAC 0[1988-1995]
(0.220)
(0.029)
(0.132)
(0.150)
(0.070)
(0.059)
Border Region {0,1} × CU 1[2005-2009]
0.230
-0.001
-0.192
-0.307**
0.009
-0.173**
CU 0[1988-1995]
(0.249)
(0.043)
(0.118)
(0.129)
(0.112)
(0.084)
Border Region {0,1} × CM 1[t ≥ 2010]
-0.003
-0.027
-0.044
-0.290**
-0.058
-0.190**
CM 0[1988-1995]
(0.213)
(0.039)
(0.123)
(0.133)
(0.081)
(0.071)
0.988
0.493
0.725
0.037
0.475
0.012
Core Region {0,1} × pre-EAC 1[1996-2000]
0.440*
0.022
0.131
-0.245*
0.099
-0.185***
pre-EAC 0[1988-1995]
(0.223)
(0.026)
(0.262)
(0.121)
(0.102)
(0.052)
Core Region {0,1} × EAC 1[2002-2004]
0.899***
0.127***
3.069***
-0.238
0.158*
-0.096
EAC 0[1988-1995]
(0.214)
(0.019)
(0.582)
(0.143)
(0.082)
(0.082)
Core Region {0,1} × CU 1[2005-2009]
0.674***
0.110***
2.193
-0.250*
0.232*
-0.016
CU 0[1988-1995]
(0.206)
(0.028)
(1.653)
(0.146)
(0.117)
(0.042)
Core Region {0,1} × CM 1[t ≥ 2010]
0.570**
0.087***
2.020
-0.180
0.099
-0.141***
CM 0[1988-1995]
(0.241)
(0.027)
(1.590)
(0.129)
(0.076)
(0.034)
0.024
0.003
0.213
0.174
0.202
0.000
Observations
255,844
102,005
11,702
232,864
137,542
133,083
R-Squared
0.29
0.16
0.54
0.21
0.25
0.17
R-Squared -Within
0.27
0.10
0.53
0.18
0.16
0.08
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
NO
NO
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Income
DHS (Region-based)
Dependent Variable
Notes: This table “disaggregates” the analysis of Table 4.5. See full notes below Table 4.5.
233
Table C.15: Disaggregated Region-Based Estimates (Placebo Countries)
Consumpt.
Income
Agglom.
Panel a)
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.20]
[0.17]
[0.00]
[0.67]
[1.48]
[0.44]
Border Region {0,1} × EAC 1[2002-2004]
0.306**
0.005
0.031
-0.045
-0.022
-0.057
EAC 0[1992-2000]
(0.107)
(0.018)
(0.193)
(0.062)
(0.051)
(0.068)
Border Region {0,1} × CU 1[2005-2009]
0.252
0.010
0.030
-0.023
0.007
-0.051
CU 0[1992-2000]
(0.176)
(0.030)
(0.100)
(0.027)
(0.042)
(0.047)
Border Region {0,1} × CM 1[t ≥ 2010]
0.218
-0.007
-0.012
-0.019
-0.004
-0.102*
CM 0[1992-2000]
(0.150)
(0.012)
(0.156)
(0.026)
(0.063)
(0.056)
0.169***
0.556***
0.942
0.485***
0.954***
0.090***
Core Region {0,1} × EAC 1[2002-2004]
-0.051
-0.049
0.615
0.004
-0.021
0.002
EAC 0[1992-2000]
(0.233)
(0.053)
(0.882)
(0.036)
(0.099)
(0.112)
Core Region {0,1} × CU 1[2005-2009]
0.317
0.087*
1.608**
0.058
0.112
0.110
CU 0[1992-2000]
(0.235)
(0.043)
(0.590)
(0.055)
(0.109)
(0.073)
Core Region {0,1} × CM 1[t ≥ 2010]
-0.295
-0.091**
1.048
0.129***
-0.138*
-0.119**
CM 0[1992-2000]
(0.200)
(0.039)
(0.816)
(0.041)
(0.070)
(0.053)
0.163
0.035
0.220
0.007
0.070
0.043
Observations
282,480
167,590
11,391
348,016
225,336
223,908
R-Squared
0.27
0.16
0.38
0.16
0.27
0.20
R-Squared -Within
0.26
0.15
0.35
0.14
0.23
0.13
Panel b)
Border Region {0,1} × pre-EAC 1[1996-2000]
-0.223
-0.005
0.041
-
0.177
0.014
pre-EAC 0[1992-1995]
(0.152)
(0.019)
(0.128)
(0.130)
(0.062)
# /
Border Region {0,1} × EAC 1[2002-2004]
0.118
-0.002
0.065
-0.045
0.124
-0.048
EAC 0[1992-1995]
(0.205)
(0.016)
(0.227)
(0.062)
(0.088)
(0.062)
Border Region {0,1} × CU 1[2005-2009]
-0.049
-0.043
-0.043
-0.056*
0.076
-0.101***
CU 0[1992-1995]
(0.319)
(0.026)
(0.150)
(0.031)
(0.075)
(0.027)
Border Region {0,1} × CM 1[t ≥ 2010]
-0.005
-0.006
0.040
-0.014
0.149
-0.089
CM 0[1992-1995]
(0.241)
(0.026)
(0.200)
(0.024)
(0.128)
(0.069)
0.983***
0.826***
0.845
0.574***
0.264
0.221
Core Region {0,1} × pre-EAC 1[1996-2000]
0.345
0.088
0.780
-0.098**
0.211**
0.189*
pre-EAC 0[1992-1995]
(0.306)
(0.071)
(0.644)
(0.036)
(0.096)
(0.088)
Core Region {0,1} × EAC 1[2002-2004]
0.215
0.018
1.157
-0.095**
0.137
0.152
EAC 0[1992-1995]
(0.379)
(0.054)
(0.889)
(0.042)
(0.120)
(0.155)
Core Region {0,1} × CU 1[2005-2009]
-0.175
-0.059**
0.417
-0.021
-0.053
-0.014
CU 0[1992-1995]
(0.281)
(0.027)
(0.459)
(0.028)
(0.040)
(0.067)
Core Region {0,1} × CM 1[t ≥ 2010]
0.123
0.034
2.179*
0.033
0.072
0.068
CM 0[1992-1995]
(0.280)
(0.062)
(1.042)
(0.028)
(0.058)
(0.079)
0.666
0.593
0.055
0.259
0.239
0.404
Observations
254,644
150,220
11,391
309,353
202,333
200,813
R-Squared
0.27
0.16
0.43
0.16
0.26
0.20
R-Squared -Within
0.26
0.15
0.40
0.14
0.23
0.12
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
NO
NO
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Dependent Variable
DHS (Region-based)
Income
Notes: This table “disaggregates” the analysis of Table 4.5. See full notes below Table 4.5.
234
Table C.16: Region-Based Estimates with Varying Intervention Year
Consumpt.
Income
Agglom.
Consumpt.
Income
Agglom.
Wealth
Index
(1-5)
Employed
Work
{0,1}
Population
Density
(sdz.)
Wealth
Index
(1-5)
Employed
Work
{0,1}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Border Region {0,1} × EAC 1[t ≥ 1994]
0.617*
-0.001
0.186
-
-
-
Border Region {0,1} × EAC 1[t ≥ 1995]
0.464
0.464
0.328
-
-
-
Border Region {0,1} × EAC 1[t ≥ 1996]
-0.094
0.019
-0.036
-0.359
-0.006
-0.081
Border Region {0,1} × EAC 1[t ≥ 1997]
-0.097
0.024
0.285
-0.160
0.004
0.055
Border Region {0,1} × EAC 1[t ≥ 1998]
-0.042
-0.042
0.298
0.002
0.002
0.116
Border Region {0,1} × EAC 1[t ≥ 1999]
0.208
0.061
0.230
0.011
0.006
0.112
Border Region {0,1} × EAC 1[t ≥ 2000]
0.270**
-0.018
0.382
0.050
0.011
0.148
Border Region {0,1} × EAC 1[t ≥ 2001]
0.167
-0.024
0.143
0.306**
0.004
0.037
Border Region {0,1} × EAC 1[t ≥ 2002]
0.192
-0.027
-0.062
0.318**
0.011
0.021
Border Region {0,1} × EAC 1[t ≥ 2003]
0.199
-0.020
-0.197
0.329**
0.011
0.029
Border Region {0,1} × EAC 1[t ≥ 2004]
0.170
0.019
-0.288
0.275*
-0.005
-0.066
Border Region {0,1} × EAC 1[t ≥ 2005]
0.143
-0.008
-0.264
0.256
0.009
-0.003
Border Region {0,1} × EAC 1[t ≥ 2006]
0.022
-0.026
-0.326*
-0.322
-0.062**
-0.067
Border Region {0,1} × EAC 1[t ≥ 2007]
0.044
-0.026
-0.171
-0.029
-0.042
-0.119
Border Region {0,1} × EAC 1[t ≥ 2008]
-0.150
-0.026
-0.218
-0.077
-0.015
-0.086
Border Region {0,1} × EAC 1[t ≥ 2009]
-0.175
-0.003
-0.213
-0.084
-0.015
-0.125
Border Region {0,1} × EAC 1[t ≥ 2010]
-0.184
-0.004
-0.189
0.003
-0.005
-0.130
Core Region {0,1} × EAC 1[t ≥ 1994]
-0.320
-0.137
0.601***
-
-
-
Core Region {0,1} × EAC 1[t ≥ 1995]
-0.291
-0.131
-1.543**
-
-
-
Core Region {0,1} × EAC 1[t ≥ 1996] 0.811***
0.094***
2.017**
0.928***
0.177***
2.798***
Core Region {0,1} × EAC 1[t ≥ 1997]
0.108
0.061
0.601
0.038
0.040
0.921
Core Region {0,1} × EAC 1[t ≥ 1998]
0.000
0.021
0.601
-0.185
-0.021
0.921
Core Region {0,1} × EAC 1[t ≥ 1999]
-0.089
0.015
-0.813
-0.128
-0.017
0.961
Core Region {0,1} × EAC 1[t ≥ 2000]
0.130
-0.007
0.637
-0.017
-0.002
1.500
Core Region {0,1} × EAC 1[t ≥ 2001] 0.418***
0.017
2.702**
-0.051
-0.046
0.856
Core Region {0,1} × EAC 1[t ≥ 2002] 0.655***
0.116***
4.071***
0.039
0.014
1.189
Core Region {0,1} × EAC 1[t ≥ 2003] 0.607***
0.122***
2.771**
-0.043
0.008
1.224
Core Region {0,1} × EAC 1[t ≥ 2004] 0.466***
0.105***
2.366
-0.205
0.003
0.186
Core Region {0,1} × EAC 1[t ≥ 2005]
0.231*
0.084***
1.480
0.307
0.097**
1.428
Core Region {0,1} × EAC 1[t ≥ 2006]
0.162
0.075**
0.841
0.409
0.092*
2.126*
Core Region {0,1} × EAC 1[t ≥ 2007]
0.149
0.034
3.069***
-0.487**
-0.136**
-0.659
Core Region {0,1} × EAC 1[t ≥ 2008]
-0.147
0.033
0.231
-0.283
-0.105***
0.922
Core Region {0,1} × EAC 1[t ≥ 2009]
-0.116
0.032*
0.373
-0.320
-0.105***
0.714
Core Region {0,1} × EAC 1[t ≥ 2010]
-0.120
0.039**
0.572
-0.340**
-0.096***
0.562
Regression Sample
DHS (Region-based) - EAC
DHS (Region-based) - Placebo
Notes: This table makes tests for varying (placebo) intervention years well before and after the EAC’s establishment by making
use of the non-GPS survey rounds of the Demographic and Health Surveys (DHS) sampled before 1999. For the “EAC Countries”
tested in columns (1) through (3), the total sample includes the Kenya survey rounds of 1988-89, 1993, 1998, 2003, 2008, Tanzania
1991-92, 1994 (KAP), 1996, 1999, 2003-04 (AIS), 2004-2005, 2007-08 (AIS), 2010, 2011-12 (AIS), and Uganda 1988-89, 1995,
1995-96, 2000-01, 2006, 2009 (MIS), 2011, 2011 (AIS). For the “Placebo Countries” tested in columns (4) through (6), the total
sample includes the Ethiopia survey rounds of 2000, 2005, 2011, Malawi 1992, 1996 (KAP) 2000, 2004, 2010, 2012 (MIS),
Mozambique 1997, 2003, 2009 (AIS), 2011, Rwanda DHS 1992, 2000, 2005, and Zambia 1996 (KAP), 2001-02 and 2007. The
estimates are constructed for each (placebo) intervention year, separately, by expanding the data past the respective cut-off. E.g.
for the estimate on the placebo intervention year 1994, the regressions use data from first available data for all dependent variables
(as early as 1988 depending on the variable) and includes survey years up until 1997. For the intervention year 1995, the data
frame is expanded to include the year 1998, for 1996 it includes 1999, and so on. Border Region {0,1} switches to one for
individuals living in a region with a mean road distance to EAC border crossings below the 10th percentile of all (within-country)
GPS-border distances in the sample. Core Region {0,1} is a dummy indicating individuals living in the region which hosts the
core agglomeration of their respective country (i.e. Nairobi, Dar es Salaam and Kampala). The standard errors reported allow
for clustering at the region level. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively.
235
Table C.17: Border Heterogeneity (AFB and DHS)
Consumpt.
Income
Agglom.
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
Last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
[3.14]
[0.19]
[0.00]
[0.76]
[1.51]
[0.52]
TZA-UGA ≤ 50km {0,1} × EAC 1[t ≥ 2002]
0.321***
0.113*
0.058
0.227***
0.208***
-0.225***
0.256
-0.115**
-0.078*
0.022
-0.178***
-0.147***
(0.043)
(0.060)
(0.048)
(0.038)
(0.035)
(0.057)
(0.164)
(0.057)
(0.046)
(0.017)
(0.031)
(0.037)
0.000
0.060*
0.226***
0.000
0.000
0.000
0.118
0.043
0.091
0.177***
0.000
0.000
TZA-KEN ≤ 50km {0,1} × EAC 1[t ≥ 2002]
-0.273***
0.035
0.006
-0.002
-0.049
0.042
-0.512***
-0.142***
0.056
0.031
-0.114**
-0.016
(0.077)
(0.037)
(0.066)
(0.054)
(0.060)
(0.069)
(0.063)
(0.021)
(0.069)
(0.065)
(0.048)
(0.065)
0.0004
0.348***
0.933***
0.966***
0.415***
0.5427
0.000
0.000
0.421***
0.630***
0.016
0.809***
KEN-UGA ≤ 50km {0,1} × EAC 1[t ≥ 2002]
-
0.078***
0.043***
-0.051**
-0.052
-
0.020
-0.035*
0.058***
0.004
-0.032
-0.238***
(0.022)
(0.015)
(0.025)
(0.050)
(0.064)
(0.019)
(0.009)
(0.023)
(0.039)
(0.037)
0.000
0.004
0.046
0.294***
0.758**
0.076***
0.000
0.856***
0.421***
0.000
Core {0,1} × EAC 1[t ≥ 2002]
-0.361***
-0.036
0.767**
-0.136
0.026
-0.059
0.358**
0.016
0.812***
0.052
-0.002
0.015
(0.074)
(0.062)
(0.336)
(0.103)
(0.065)
(0.082)
(0.151)
(0.044)
(0.121)
(0.036)
(0.099)
(0.030)
0.0000
0.5640
0.0226
0.1900
0.6936
0.4738
0.0181
0.7231
0.0000
0.1505
0.9818
0.6218
Individual Controls
YES
YES
NO
YES
YES
YES
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.13
0.14
0.25
0.23
0.25
0.09
0.34
0.16
0.29
0.23
0.26
0.17
R-Squared -Within
0.10
0.10
0.22
0.07
0.22
0.06
0.32
0.11
0.28
0.20
0.17
0.09
Dependent Variable
Afrobarometer
DHS
Income
Income
Notes: This table analyzes the heterogeneity across the three Border regions. TZA-UGA {0,1}, TZA-KEN {0,1} and KEN-UGA {0,1} thereby switch to one for individuals living
within 50km road distance to the respective border crossing (within either country). The results in each column and panel are produced by a separate regression. The sample
mean of the respective dependent variable is given in brackets above the estimates. Binary dependent variables are estimated through a Linear Probability Model (LPM). The
standard errors reported allow for spatial correlation, i.e. Conley standard errors are used. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full
notes below Table 4.1.
236
Table C.18: Opinion Polling (AFB and KHDS)
Panel a)
Support for:
Regional
Integration
{1,5}
Ease of:
Crossing
Borders
{1,4}
Helps your
Country:
REC / EAC
{1,3}
Helps your
Country:
African Union
{1,3}
Like as Neighbor:
Immigr./For.
Worker
{1,5}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Sample Mean of Dep. Var.
[2.70]
[2.71]
[3.71]
[2.21]
[1.66]
[1.62]
[3.38]
Border [0,1]
0.041
-1.006***
0.087
0.368***
-0.109
-0.115
-0.245
(0.130)
(0.245)
(0.249)
(0.083)
(0.095)
(0.092)
(0.160)
Core {0,1}
-0.019
0.172**
0.000
-0.039
-0.033
-0.004
-0.198***
(0.050)
(0.073)
(0.059)
(0.039)
(0.038)
(0.027)
(0.057)
Border [0,1] × EAC 1[t ≥ 2002]
-0.356**
0.745***
-
-
-
-
-
(0.146)
(0.255)
Core {0,1} × EAC 1[t ≥ 2002]
-0.050
-0.209*
-
-
-
-
-
(0.130)
(0.112)
0.698
0.062
0.998
0.315
0.380
0.878
0.001
Observations
36,213
23,370
6,362
4,766
11,687
13,069
11,647
R-Squared
0.08
0.08
0.10
0.06
0.03
0.03
0.02
R-Squared -Within
0.01
0.02
0.01
0.02
0.00
0.00
0.01
DiD
Panel b)
Main Reason
for Migrating:
Economic
{0,1}
After Migr.
to Res.:
Paid Empl.
{0,1}
After Migr
to Res.: Paid
form. Empl.
{0,1}
Activity in
Residence:
Working
{0,1}
HH. Wealth
vs. 10 years
ago (1994)
{1,5}
HH. Wealth
today
(2004)
{1,5}
HH. Life
Satisfaction
(2004)
{1,9}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Sample Mean of Dep. Var.
[0.20]
[0.14]
[0.07]
[0.83]
[2.62]
[2.62]
[3.79]
Border [0,1]
-0.631
-0.564**
-0.408**
0.547*
0.274
0.237
0.783
(5.759)
(0.241)
(0.208)
(0.327)
(0.278)
(0.265)
(0.819)
Core {0,1}
0.674
0.061
-0.069
-0.051
0.160
0.292**
0.444
(0.927)
(0.073)
(0.072)
(0.103)
(0.188)
(0.143)
(0.415)
0.469
0.4021
0.3357
0.6201
0.3954
0.0407
0.2853
Border [0,1] * EAC 1[ 2004]
0.855
-
-
-
-
-
-
(3.719)
Border [0,1] * CU 1[2010]
1.143
-
-
-
-
-
-
(3.944)
0.773
Core {0,1} * EAC 1[2004]
-0.519
-
-
-
-
-
-
(0.901)
Core {0,1} * CU 1[2010]
-0.392
-
-
-
-
-
-
(0.600)
0.514
Observations
1,347
900
900
774
2,833
2,833
2,833
Observations - Fixed Effects
1,202
2
2
2
2
2
2
R-Squared
0.89
0.16
0.22
0.03
0.02
0.08
0.10
R-Squared -Within
0.096
0.155
0.223
0.024
0.021
0.078
0.10
Individual Controls
YES
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO/YES
NO
NO
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
YES
Dependent Variable
Contemporary Opinion
Afrobarometer
Pres. vs.
Past: Living
Standards
{1,4}
Pres. vs.
Past: Life
Satsfaction
{1,4}
DiD
Contemporary Opinion
Kagera Health and Development Survey (KHDS)
Notes: This table analyzes opinions and sentiments of survey respondents. As these data are not available for all tested
survey rounds, a difference-in-differences (DiD) estimate can only be conducted for columns (1) and (2) in panel a) as
well as column (1) in panel b). The results in each column and panel are produced by a separate regression. All regressions
include individual-level controls for respondents’ age, gender, as well as education, and also include the geographic controls
average monthly temperature, average monthly rainfall, elevation, ruggedness, and the number of growing days. The
regressions also include country-year fixed effects. Binary dependent variables are estimated through a Linear Probability
Model (LPM). The standard errors reported allow for spatial correlation, i.e. Conley standard errors are used. ***, **, *
represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 4.1 and Table 4.3,
respectively.
237
Table C.19: Border Price Pass-Through (KHDS)
Maize
(TZS)
Millet
(TZS)
Tea/Coffee
(TZS)
Sugar
(TZS)
(1)
(2)
(3)
(4)
Sample Mean of Dep. Var.
[207.25]
[122.83]
[24.60]
[52.50]
Border [0,1] × EAC 1[ 2004]
-0.019
-1.066*
0.308
-0.048
(0.090)
(0.563)
(0.928)
(0.094)
0.834***
0.0642
0.740***
0.611***
Core {0,1} × EAC 1[ 2004]
-0.036***
-
-0.080*
-0.021***
(0.006)
(0.047)
(0.004)
0.0000
-
0.0908
0.0000
Individual Controls
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
Household Fixed Effects
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
Observations
3,267
193
948
1,722
Observations - Fixed Effects
1,990
127
554
899
R-Squared
0.99
0.99
0.99
0.99
R-Squared -Within
0.01
0.05
0.05
0.00
Unit Price of Item in Tanzanian Shilling
Dependent Variable
KHDS
Notes: This table tests for a differential price-pass through of traded, homogenous, staple
goods, following the establishment of the EAC. The results in each column are produced by
a separate regression. Data come from the Kagera Health and Development Surveys (KHDS)
consumption components which askes respondents for seasonal prices of goods. These data
are collected in four waves across 1991-1994, as well as in 2004. All regressions include
country-year fixed effects. The results in each column and panel are produced by a separate
regression. Binary dependent variables are estimated through a Linear Probability Model
(LPM). The standard errors reported allow for spatial correlation, i.e. Conley standard errors
are used. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See
Table 4.3 for full table notes.
238
Table C.20: Decomposed Indices (Afrobarometer)
Freq. gone
without:
[Water]
{0,4}
Freq. gone
without:
[Food]
{0,4}
Freq. g. without:
[Medicine/
Med. Care]
{0,4}
Occupation:
[Agrarian]
{0,1}
Occupation:
[Worker]
{0,1}
Occupation:
[Professional]
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.13]
[0.98]
[1.26]
[0.55]
[0.23]
[0.21]
Border [0,1] × EAC 1[2002-2004]
-0.620*
0.120
0.261
-0.029
0.041
-0.012
(0.341)
(0.379)
(0.354)
(0.083)
(0.053)
(0.059)
0.069
0.752
0.460
0.728
0.439
0.836
Border [0,1] × CU 1[2005-2009]
-0.268
0.117
0.667*
-0.044
0.048
-0.004
(0.336)
(0.341)
(0.344)
(0.077)
(0.059)
(0.054)
0.425
0.730
0.053
0.565
0.416
0.943
Border [0,1] × CM 1[t ≥ 2010]
-0.302
0.260
0.466
-0.243***
0.196***
0.047
(0.333)
(0.301)
(0.332)
(0.056)
(0.048)
(0.042)
0.365
0.388
0.161
0.000
0.000
0.272
Core {0,1} × EAC 1[2002-2004]
-0.167
-0.070
-0.546***
0.006
-0.033
0.027
(0.157)
(0.106)
(0.086)
(0.066)
(0.059)
(0.022)
0.289
0.505
0.000
0.929
0.571
0.205
Core {0,1} × CU 1[2005-2009]
-0.563***
-0.253**
-0.591***
-0.061
0.065*
-0.004
(0.100)
(0.116)
(0.121)
(0.062)
(0.035)
(0.030)
0.000
0.029
0.000
0.324
0.068
0.901
Core {0,1} × CM 1[t ≥ 2010]
-0.452***
-0.082
-0.480***
-0.057
0.124**
-0.066**
(0.082)
(0.080)
(0.086)
(0.066)
(0.055)
(0.030)
0.000
0.304
0.000
0.388
0.026
0.028
Observations
36,024
36,007
35,973
21,232
21,232
21,232
R-Squared
0.08
0.10
0.10
0.25
0.13
0.20
R-Squared -Within
0.06
0.09
0.07
0.20
0.07
0.18
Dependent Variable
Afrobarometer
Lived Poverty
Income
Notes: This table analyzes the components of the two composite indices used throughout the paper. The results in
each column and panel are produced by a separate regression. All regressions include individual-level controls for
respondents’ age, gender, as well as education, and also include the geographic controls average monthly
temperature, average monthly rainfall, elevation, ruggedness, and the number of growing days. The regressions also
include country-year fixed effects. Binary dependent variables are estimated through a Linear Probability Model
(LPM). The standard errors reported allow for spatial correlation, i.e. Conley standard errors are used. ***, **, *
represents significance at the 1, 5 and 10 percent level, respectively. See full notes below Table 4.1.
239
Table C.21: Decomposed Indices (DHS and KHDS)
International
Wealth Index
Comparative
Wealth Index
Occupation:
[Agrarian]
{0,1}
Occupation:
[Worker]
{0,1}
Occupation:
[Professional]
{0,1}
Occupation:
[Agrarian]
{0,1}
Occupation:
[Worker]
{0,1}
Occupation:
[Professional]
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Sample Mean of Dep. Var.
[23.40]
-[0.67]
[0.57]
[0.35]
[0.08]
[0.84]
[0.12]
[0.04]
Border [0,1] × EAC 1[2002-2004]
0.842
0.069
0.125
-0.067
-0.059*
-0.030
-0.048
0.077
(0.762)
(0.171)
(0.185)
(0.186)
(0.030)
(0.224)
(0.210)
(0.077)
0.269
0.686
0.498
0.721
0.051
0.8945
0.8202
0.3119
Border [0,1] × CU 1[2005-2009]
-0.153
-0.123
0.249***
-0.206**
-0.043
-
-
-
(0.481)
(0.119)
(0.093)
(0.093)
(0.035)
0.750
0.302
0.008
0.027
0.229
0.692
0.388
0.525
Border [0,1] × CM 1[t ≥ 2010]
-0.185
-0.105
0.149**
-0.149**
0.000
-0.113***
0.236***
-0.123
(0.483)
(0.127)
(0.069)
(0.069)
(0.019)
(0.286)
(0.273)
(0.193)
0.702
0.407
0.032
0.031
0.996
0.000
0.000
0.119
Core {0,1} × EAC 1[2002-2004]
1.869***
0.724***
0.056
-0.066
0.010
-0.157*
0.180**
-0.023
(0.375)
(0.159)
(0.058)
(0.055)
(0.011)
(0.021)
(0.023)
(0.015)
0.000
0.000
0.331
0.234
0.357
0.0862
0.0227
0.1967
Core {0,1} × CU 1[2005-2009]
0.644***
0.320***
-0.025
-0.008
0.033
-
-
-
(0.228)
(0.093)
(0.119)
(0.131)
(0.021)
0.005
0.001
0.831
0.954
0.120
Core {0,1} × CM 1[t ≥ 2010]
0.571***
0.273***
0.025
-0.056
0.032*
-0.099***
0.128***
-0.029***
(0.111)
(0.094)
(0.073)
(0.060)
(0.017)
(0.057)
(0.056)
(0.023)
0.000
0.004
0.737
0.353
0.066
Observations
183,253
73,985
95,717
95,717
95,717
15,685
15,685
15,685
Observations - Fixed Effects
-
-
-
-
-
6,253
6,253
6,253
R-Squared
0.17
0.46
0.23
0.13
0.15
0.79
0.72
0.80
R-Squared -Within
0.116
0.437
0.145
0.076
0.100
0.116
0.06
0.08
Individual Controls
YES
YES
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
NO
NO
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
YES
YES
Dependent Variable
DHS
KHDS
Consumption
Income
Notes: This table analyzes the components of the two composite indices used throughout the paper. The results in each column and panel are produced
by a separate regression. All regressions include individual-level controls for respondents’ age, gender, as well as education, and also include the geographic
controls average monthly temperature, average monthly rainfall, elevation, ruggedness, and the number of growing days. The regressions also include
country-year fixed effects. Binary dependent variables are estimated through a Linear Probability Model (LPM). The standard errors reported allow for
spatial correlation, i.e. Conley standard errors are used. ***, **, * represents significance at the 1, 5 and 10 percent level, respectively. See full notes
below Table 4.2.
240
Table C.22: Robustness Check: Clustered Standard Errors (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.66]
[2.09]
Border [0,1] × EAC 1[2002-2004]
-0.083
0.050
0.083
0.072
0.017
-0.350*
(0.170)
(0.045)
(0.058)
(0.057)
(0.115)
(0.194)
0.625***
0.275***
0.157***
0.201**
0.885***
0.071
Border [0,1] × CU 1[2005-2009]
0.172
0.060
0.045
0.091
0.040
-0.234
(0.162)
(0.050)
(0.061)
(0.061)
(0.110)
(0.188)
0.288**
0.234**
0.458***
0.135***
0.714**
0.214
Border [0,1] × CM 1[t ≥ 2010]
0.142
0.098***
0.099
-0.001
0.289***
-0.456***
(0.147)
(0.037)
(0.063)
(0.049)
(0.092)
(0.176)
(0.336)
(0.008)
(0.115)
(0.979)
(0.002)
(0.010)
Core {0,1} × EAC 1[2002-2004]
-0.261**
-0.013
-0.158
0.026
0.021
0.041
(0.126)
(0.041)
(0.135)
(0.044)
(0.084)
(0.159)
0.038
0.755***
0.244
0.548***
0.798***
0.794***
Core {0,1} × CU 1[2005-2009]
-0.470***
-0.006
0.769***
0.002
0.057
-0.234
(0.125)
(0.044)
(0.214)
(0.046)
(0.080)
(0.155)
0.000
0.895***
0.000
0.958***
0.478***
0.131***
Core {0,1} × CM 1[t ≥ 2010]
-0.338***
-0.076**
1.113***
-0.258***
-0.009
0.043
(0.119)
(0.036)
(0.179)
(0.041)
(0.078)
(0.150)
0.0044
0.0363
0.0000
0.0000
0.9071
0.7740
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,104
26,563
21,232
35,975
R-Squared
0.12
0.13
0.28
0.23
0.25
0.09
R-Squared -Within
0.10
0.10
0.25
0.23
0.25
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The standard errors are
clustered at the enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively.
241
Table C.23: Robustness Check: Clustered Standard Errors (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.19]
[0.00]
[0.76]
[1.51]
[0.52]
Border [0,1] × EAC 1[2002-2004]
-0.294
-0.129*
0.104
0.110***
-0.184
-0.162*
(0.296)
(0.070)
(0.094)
(0.041)
(0.134)
(0.097)
0.321
0.065
0.271***
0.007
0.168
0.093
Border [0,1] × CU 1[2005-2009]
-0.377
-0.138***
0.084
0.115***
-0.291***
-0.271***
(0.274)
(0.036)
(0.085)
(0.037)
(0.088)
(0.059)
0.168
0.000
0.322***
0.002
0.001
0.000
Border [0,1] × CM 1[t ≥ 2010]
-0.505**
-0.034
0.060
0.065*
-0.149**
-0.141***
(0.257)
(0.031)
(0.076)
(0.034)
(0.073)
(0.051)
0.049
0.265***
0.424***
0.057
0.043
0.005***
Core {0,1} × EAC 1[2002-2004]
0.435***
0.073*
0.146
0.068***
-0.046
-0.002
(0.110)
(0.040)
(0.201)
(0.021)
(0.058)
(0.038)
0.000
0.067*
0.466
0.001
0.425***
0.965***
Core {0,1} × CU 1[2005-2009]
0.315***
0.037
0.770***
0.014
0.058
0.053
(0.097)
(0.032)
(0.241)
(0.020)
(0.052)
(0.035)
0.001
0.243***
0.001
0.487***
0.265***
0.127***
Core {0,1} × CM 1[t ≥ 2010]
0.345***
0.006
0.956***
0.049***
0.007
0.029
(0.091)
(0.029)
(0.185)
(0.017)
(0.047)
(0.032)
0.0001
0.8329
0.0000
0.0046
0.8838
0.3652
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
204,991
71,738
7,692
191,616
95,717
98,963
R-Squared
0.34
0.16
0.30
0.22
0.26
0.17
R-Squared -Within
0.33
0.11
0.29
0.22
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that
most of the respective notes apply here. However, note the following important amendments: The standard errors
are clustered at the enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively.
242
Table C.24: Robustness Check: Clustered Standard Errors (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border [0,1] × EAC 1[2004]
-853.376
187.152
0.581**
0.071
0.107
-0.848***
(1085.855)
(651.437)
(0.239)
(0.050)
(0.261)
(0.288)
0.432
0.7740
0.015
0.1587
0.681
0.003
Border [0,1] × CU 1[2010]
-1317.845
0.432*
-0.010
-1.631***
(1156.327)
(0.246)
(0.404)
(0.289)
0.255
0.079***
0.981***
0.000
Core {0,1} × EAC 1[2004]
170.320*
706.607
-0.013
0.026
0.134***
0.891***
(100.162)
(1002.964)
(0.033)
(0.018)
(0.021)
(0.126)
0.089
0.4814
0.687
0.1458
0.000
0.000
Core {0,1} × CU 1[2010]
275.036**
-0.039
0.069
0.891***
(111.403)
(0.040)
(0.067)
(0.132)
0.0137
0.3307
0.3013
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.43
R-Squared -Within
0.12
0.09
0.04
0.01
0.14
0.41
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: The standard errors are
clustered at the enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively.
243
Table C.25: Robustness Check: Cluster Standard Errors (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border [0,1] × EAC 1[2004]
-177.749
-670.444
7592.679
-0.201
-96.031
5.949
(262.369)
(1034.002)
(12613.087)
(0.393)
(135.964)
(9.453)
0.498
0.517
0.5474
0.609
0.4809
0.529
Border [0,1] × CU 1[2010]
-388.178
-911.118
-0.233
6.555
(267.416)
(1093.951)
(0.297)
(8.908)
0.147
0.405
0.432***
0.462
Core {0,1} × EAC 1[2004]
122.188**
35.341
4013.722
-0.058
8.109
-1.079
(49.625)
(57.589)
(5982.927)
(0.075)
(27.118)
(1.648)
0.014
0.540
0.5026
0.436
0.7653
0.512
Core {0,1} × CU 1[2010]
131.046***
133.815*
-0.008
-0.404
(47.557)
(74.509)
(0.051)
(1.636)
0.0059
0.0727
0.8796
0.8050
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.10
0.09
0.07
0.16
0.18
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: The standard errors are
clustered at the enumeration area level. ***, **, * represents significance at the 1, 5 and 10 percent level,
respectively.
244
Table C.26: Robustness Check: No Controls (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
Border [0,1] × EAC 1[2002-2004]
-0.420
0.060
0.047
0.103
0.013
-0.524*
(0.340)
(0.075)
(0.038)
(0.083)
(0.186)
(0.305)
0.216**
0.420***
0.213***
0.214***
0.946***
0.086
Border [0,1] × CU 1[2005-2009]
-0.088
0.046
0.023
0.080
0.029
-0.341
(0.335)
(0.074)
(0.060)
(0.083)
(0.143)
(0.311)
0.793***
0.538***
0.698***
0.335***
0.839***
0.274
Border [0,1] × CM 1[t ≥ 2010]
-0.151***
0.110
0.065***
0.024***
0.319
-0.599
(0.307)
(0.042)
(0.072)
(0.087)
(0.114)
(0.297)
(0.002)
(0.301)
(0.000)
(0.002)
(0.842)
(0.837)
Core {0,1} × EAC 1[2002-2004]
-0.392***
-0.011
-0.188***
0.018
0.029
-0.047
(0.088)
(0.044)
(0.040)
(0.057)
(0.068)
(0.089)
0.000
0.801***
0.000
0.756***
0.669***
0.597*
Core {0,1} × CU 1[2005-2009]
-0.607***
0.005
0.738**
0.002
0.103
-0.318***
(0.116)
(0.048)
(0.362)
(0.053)
(0.103)
(0.114)
0.000
0.914***
0.042
0.973***
0.321***
0.005
Core {0,1} × CM 1[t ≥ 2010]
-0.466***
-0.054
1.094***
-0.253***
0.060
-0.061
(0.081)
(0.065)
(0.396)
(0.102)
(0.110)
(0.077)
0.0023
0.3014
0.0000
0.0017
0.8416
0.8367
Individual Controls
NO
NO
NO
NO
NO
NO
Geographic Controls
NO
NO
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,371
25,775
4,156
26,880
21,482
36,304
R-Squared
0.04
0.06
0.27
0.18
0.09
0.05
R-Squared -Within
0.02
0.02
0.24
0.18
0.09
0.05
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: All individual- and
geographic controls are removed from the right side of the regression equation.
245
Table C.27: Robustness Check: No Controls (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.11]
[0.18]
[0.00]
[0.78]
[1.51]
[0.52]
Border [0,1] × EAC 1[2002-2004]
-0.275
-0.162
0.096
0.169**
-0.257
-0.180
(0.390)
(0.134)
(0.150)
(0.074)
(0.232)
(0.187)
0.480***
0.228**
0.522***
0.023
0.268
0.336
Border [0,1] × CU 1[2005-2009]
-0.104
-0.098*
0.077
0.175***
-0.282**
-0.220**
(0.307)
(0.059)
(0.086)
(0.056)
(0.128)
(0.103)
0.734**
0.095***
0.375***
0.002
0.027
0.032
Border [0,1] × CM 1[t ≥ 2010]
-0.318
-0.036
0.039
0.133**
-0.216**
-0.152
(0.246)
(0.044)
(0.095)
(0.052)
(0.085)
(0.094)
0.195
0.416***
0.684***
0.010
0.011
0.106
Core {0,1} × EAC 1[2002-2004]
0.334*
0.077
0.071
0.092
-0.098
-0.065
(0.173)
(0.050)
(0.320)
(0.060)
(0.094)
(0.046)
0.053
0.122**
0.826
0.123***
0.296***
0.161***
Core {0,1} × CU 1[2005-2009]
0.386***
0.052
0.708***
0.031
0.060
0.029
(0.149)
(0.042)
(0.170)
(0.062)
(0.136)
(0.049)
0.010
0.213***
0.000
0.614***
0.660***
0.558***
Core {0,1} × CM 1[t ≥ 2010]
0.375*
0.018
0.914***
0.077*
-0.024
0.009
(0.224)
(0.047)
(0.104)
(0.044)
(0.116)
(0.043)
0.0946
0.7067
0.0000
0.0762
0.8380
0.8383
Individual Controls
NO
NO
NO
NO
NO
NO
Geographic Controls
NO
NO
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,306
71,757
7,692
169,898
95,740
98,986
R-Squared
0.17
0.11
0.29
0.04
0.17
0.13
R-Squared -Within
0.16
0.06
0.28
0.04
0.17
0.13
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that
most of the respective notes apply here. However, note the following important amendments: All individual- and
geographic controls are removed from the right side of the regression equation.
246
Table C.28: Robustness Check: No Controls (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border [0,1] × EAC 1[2004]
-244.619
231.719
0.481**
0.065*
0.194
-1.516**
(739.987)
(230.704)
(0.231)
(0.035)
(0.245)
(0.658)
0.741
0.3153
0.037
0.0644
0.429
0.021
Border [0,1] × CU 1[2010]
-613.663***
0.449***
0.219***
-1.946***
(871.268)
(0.246)
(0.414)
(0.731)
Core {0,1} × EAC 1[2004]
122.774***
559.391***
-0.012
0.024***
0.109***
1.112***
(44.950)
(40.735)
(0.027)
(0.008)
(0.021)
(0.296)
0.006
0.0000
0.670
0.0019
0.000
0.000
Core {0,1} × CU 1[2010]
212.121***
-0.021***
0.035***
1.346***
(71.187)
(0.031)
(0.079)
(0.461)
Individual Controls
NO
NO
NO
NO
NO
NO
Geographic Controls
NO
NO
NO
NO
NO
NO
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
6,527
5,856
31,263
19,119
20,120
2,293
Observations - Fixed Effects
4,424
2,956
14,530
8,503
7,570
12
R-Squared
0.86
0.93
0.60
0.69
0.80
0.36
R-Squared -Within
0.02
0.05
0.01
0.00
0.08
0.34
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: All individual- and geographic controls
are removed from the right side of the regression equation.
247
Table C.29: Robustness Check: No Controls (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border [0,1] × EAC 1[2004]
54.492
-291.081
3979.185
0.005
-194.034
5.802
(208.728)
(669.004)
(3760.092)
(0.312)
(319.076)
(7.924)
0.794
0.664
0.2900
0.987
0.5433
0.464
Border [0,1] × CU 1[2010]
-145.003***
-453.312***
-0.070***
7.066***
(255.255)
(750.695)
(0.252)
(7.395)
Core {0,1} × EAC 1[2004]
95.130***
20.114
3358.445***
-0.004
28.256
-0.435
(33.049)
(28.938)
(574.014)
(0.026)
(21.074)
(1.073)
0.004
0.487
0.0000
0.862
0.1805
0.685
Core {0,1} × CU 1[2010]
109.186***
96.839***
0.065***
-0.479***
(36.483)
(42.831)
(0.039)
(0.952)
Individual Controls
NO
NO
NO
NO
NO
YES
Geographic Controls
NO
NO
NO
NO
NO
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
6,530
6,583
5,855
19,657
2,036
7,429
Observations - Fixed Effects
4,424
4,451
2,956
13,170
1,374
4,900
R-Squared
0.87
0.82
0.50
0.80
0.83
0.90
R-Squared -Within
0.01
0.01
0.02
0.00
0.03
0.00
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: All individual- and
geographic controls are removed from the right side of the regression equation.
248
Table C.30: Robustness Test: Extended Set of Geographic Controls (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.66]
[2.09]
Border [0,1] × EAC 1[2002-2004]
0.097
0.048
0.039
0.032
0.054
-0.750***
(0.363)
(0.068)
(0.074)
(0.090)
(0.152)
(0.274)
0.790***
0.485***
0.596***
0.724***
0.721***
0.006
Border [0,1] × CU 1[2005-2009]
0.281
0.097
0.065
0.105
0.128
-0.548*
(0.359)
(0.074)
(0.076)
(0.085)
(0.137)
(0.288)
0.434*
0.192
0.392***
0.218***
0.352
0.057
Border [0,1] × CM 1[t ≥ 2010]
0.228
0.124***
0.141
-0.037
0.331***
-0.841**
(0.349)
(0.044)
(0.099)
(0.090)
(0.107)
(0.242)
(0.514)
(0.005)
(0.153)
(0.780)
(0.005)
(0.044)
Core {0,1} × EAC 1[2002-2004]
-0.211***
-0.012
-0.216**
0.019
0.021
0.054
(0.080)
(0.051)
(0.086)
(0.069)
(0.076)
(0.081)
0.009
0.806***
0.012
0.783***
0.778***
0.502**
Core {0,1} × CU 1[2005-2009]
-0.424***
-0.001
0.694**
0.000
0.069
-0.205**
(0.094)
(0.047)
(0.301)
(0.060)
(0.084)
(0.088)
0.000
0.980***
0.021
0.996***
0.411***
0.019
Core {0,1} × CM 1[t ≥ 2010]
-0.304***
-0.076
1.031***
-0.262**
-0.018
0.046
(0.073)
(0.070)
(0.353)
(0.109)
(0.082)
(0.083)
0.0000
0.2779
0.0035
0.0126
0.5813
0.4340
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,002
25,441
4,151
26,539
21,209
35,935
R-Squared
0.13
0.14
0.29
0.24
0.26
0.10
R-Squared -Within
0.11
0.10
0.26
0.24
0.26
0.10
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I add the extended set of
geographic controls anticipated in the manuscript, namely locations’ Absolute Latitude, Malaria Ecology, distance
to Navigable Rivers, Major Lakes and Major Harbors and add the dummy non-EAC ≤ 100km {0,1} as well as the
interaction of it with all period dummies to net out effects potentially stemming from a change (loss) in market
access at non-EAC borders.
249
Table C.31: Robustness Check: Extended Set of Geographic Controls (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
Border [0,1] × EAC 1[2002-2004]
-0.098
-0.127
-0.015
0.072
-0.156
-0.071
(0.333)
(0.080)
(0.138)
(0.060)
(0.177)
(0.163)
0.768***
0.113
0.915***
0.228***
0.377
0.663***
Border [0,1] × CU 1[2005-2009]
-0.079
-0.101**
0.061
0.081**
-0.303***
-0.206**
(0.289)
(0.052)
(0.108)
(0.040)
(0.096)
(0.088)
0.783***
0.050**
0.572***
0.040*
0.002
0.019
Border [0,1] × CM 1[t ≥ 2010]
-0.257
-0.022
0.018
0.021
-0.134*
-0.049
(0.267)
(0.041)
(0.100)
(0.047)
(0.070)
(0.090)
0.336
0.589***
0.861***
0.657***
0.054
0.585***
Core {0,1} × EAC 1[2002-2004]
0.483***
0.066*
0.014
0.042
-0.040
0.016
(0.182)
(0.036)
(0.255)
(0.029)
(0.069)
(0.027)
0.008
0.065
0.955
0.148***
0.562***
0.560***
Core {0,1} × CU 1[2005-2009]
0.383***
0.042
0.721***
-0.003
0.071
0.079**
(0.126)
(0.034)
(0.145)
(0.037)
(0.107)
(0.032)
0.002
0.223***
0.000
0.942***
0.508***
0.014
Core {0,1} × CM 1[t ≥ 2010]
0.403**
0.007
0.886***
0.046*
0.008
0.045
(0.157)
(0.041)
(0.094)
(0.026)
(0.091)
(0.030)
0.0101
0.8661
0.0000
0.0847
0.9291
0.1304
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.35
0.16
0.30
0.23
0.26
0.18
R-Squared -Within
0.34
0.11
0.29
0.23
0.26
0.18
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I add the extended set of
geographic controls anticipated in the manuscript, namely locations’ Absolute Latitude, Malaria Ecology, distance
to Navigable Rivers, Major Lakes and Major Harbors and add the dummy non-EAC ≤ 100km {0,1} as well as the
interaction of it with all period dummies to net out effects potentially stemming from a change (loss) in market
access at non-EAC borders.
250
Table C.32: Robustness Check: Extended Set of Geographic Controls (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border [0,1] × EAC 1[2004]
-854.048
35.596
0.578*
0.075
0.102
0.314
(1034.553)
(643.196)
(0.318)
(0.052)
(0.255)
(0.655)
0.409
0.9559
0.069
0.1525
0.688
0.632
Border [0,1] × CU 1[2010]
-1283.907
0.454
0.118
-0.188
(1086.753)
(0.302)
(0.410)
(0.583)
0.238
0.133***
0.774***
0.747
Core {0,1} × EAC 1[2004]
172.061***
713.341***
-0.012
0.026***
0.143***
0.836***
(39.808)
(69.683)
(0.024)
(0.009)
(0.025)
(0.030)
0.000
0.0000
0.626
0.0027
0.000
0.000
Core {0,1} × CU 1[2010]
269.142***
-0.036
0.080
0.846***
(54.303)
(0.028)
(0.072)
(0.072)
0.0000
0.1907
0.2682
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,467
2,678
24,941
14,241
15,664
2,265
Observations - Fixed Effects
3,801
2,346
12,732
6,980
6,247
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.45
R-Squared -Within
0.14
0.10
0.04
0.01
0.14
0.42
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: I add the extended set of geographic
controls anticipated in the manuscript, namely locations’ Absolute Latitude, Malaria Ecology, distance to
Navigable Rivers, Major Lakes and Major Harbors and add the dummy non-EAC ≤ 100km {0,1} as well as the
interaction of it with all period dummies to net out effects potentially stemming from a change (loss) in market
access at non-EAC borders.
251
Table C.33: Robustness Check: Extended Set of Geographic Controls (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border [0,1] × EAC 1[2004]
-174.383
-673.897
8162.184
-0.165
-86.860
5.782
(249.275)
(1024.382)
(14988.404)
(0.370)
(144.913)
(9.909)
0.484
0.511
0.5864
0.655
0.5492
0.560
Border [0,1] × CU 1[2010]
-378.751***
-886.309***
-0.157***
5.605***
(261.737)
(1058.009)
(0.329)
(9.731)
Core {0,1} × EAC 1[2004]
120.084***
37.350
3971.477***
-0.064
9.471
-1.209
(35.210)
(36.610)
(577.551)
(0.041)
(25.328)
(1.489)
0.001
0.308
0.0000
0.122
0.7086
0.417
Core {0,1} × CU 1[2010]
127.440***
130.428***
-0.011***
-0.689***
(37.672)
(35.080)
(0.047)
(1.486)
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,469
5,499
2,678
16,303
1,775
5,380
Observations - Fixed Effects
3,802
3,815
2,346
11,581
1,186
3,754
R-Squared
0.89
0.83
0.69
0.86
0.95
0.89
R-Squared -Within
0.10
0.11
0.07
0.16
0.23
0.10
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: I add the extended set of geographic
controls anticipated in the manuscript, namely locations’ Absolute Latitude, Malaria Ecology, distance to
Navigable Rivers, Major Lakes and Major Harbors and add the dummy non-EAC ≤ 100km {0,1} as well as the
interaction of it with all period dummies to net out effects potentially stemming from a change (loss) in market
access at non-EAC borders.
252
Table C.34: Robustness Check: Using Survey Weights (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.66]
[2.09]
Border [0,1] × EAC 1[2002-2004]
-0.032
0.078
0.093*
0.113
0.141
-0.288
(0.315)
(0.059)
(0.057)
(0.075)
(0.125)
(0.300)
0.920***
0.187**
0.098
0.129
0.260***
0.336*
Border [0,1] × CU 1[2005-2009]
0.263
0.085
0.065
0.110
0.064
-0.179
(0.293)
(0.068)
(0.077)
(0.076)
(0.129)
(0.319)
0.369**
0.213**
0.399***
0.147***
0.619**
0.575
Border [0,1] × CM 1[t ≥ 2010]
0.152
0.103**
0.117
0.013
0.284***
-0.434
(0.288)
(0.040)
(0.111)
(0.074)
(0.095)
(0.294)
(0.597)
(0.011)
(0.293)
(0.859)
(0.003)
(0.141)
Core {0,1} × EAC 1[2002-2004]
-0.245***
-0.016
-0.168**
0.016
-0.010
0.039
(0.079)
(0.048)
(0.069)
(0.069)
(0.092)
(0.082)
0.002
0.746***
0.014
0.811***
0.913***
0.638***
Core {0,1} × CU 1[2005-2009]
-0.443***
0.001
0.910***
0.014
0.012
-0.187*
(0.084)
(0.054)
(0.351)
(0.063)
(0.097)
(0.098)
0.000
0.986***
0.010
0.829***
0.902***
0.057*
Core {0,1} × CM 1[t ≥ 2010]
-0.340***
-0.089
0.916**
-0.268**
-0.002
0.032
(0.067)
(0.078)
(0.388)
(0.114)
(0.096)
(0.082)
0.0000
0.2510
0.0183
0.0192
0.9836
0.6925
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.27
0.24
0.25
0.10
R-Squared -Within
0.10
0.10
0.24
0.24
0.25
0.10
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I produce a weighted
regression, employing the survey sampling weights provided by the Afrobarometer and accounting for the pooling
across countries and years by standardizing the weights for each country-survey round pair.
253
Table C.35: Robustness Check: Using Survey Weights (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
Border [0,1] × EAC 1[2002-2004]
-0.284
-0.129
0.104
0.114*
-0.184
-0.162
(0.312)
(0.089)
(0.136)
(0.059)
(0.188)
(0.136)
0.362
0.144
0.447***
0.054
0.328
0.231
Border [0,1] × CU 1[2005-2009]
-0.372*
-0.138**
0.084
0.118***
-0.291***
-0.271***
(0.224)
(0.054)
(0.089)
(0.043)
(0.106)
(0.083)
0.096
0.011
0.348***
0.006
0.006
0.001
Border [0,1] × CM 1[t ≥ 2010]
-0.547***
-0.034
0.060
0.081*
-0.149**
-0.141
(0.202)
(0.040)
(0.095)
(0.046)
(0.075)
(0.092)
0.007
0.384***
0.523***
0.079
0.047
0.125***
Core {0,1} × EAC 1[2002-2004]
0.433***
0.073*
0.146
0.067*
-0.046
-0.002
(0.141)
(0.040)
(0.278)
(0.038)
(0.062)
(0.026)
0.002
0.071*
0.599
0.075***
0.454***
0.948***
Core {0,1} × CU 1[2005-2009]
0.314***
0.037
0.770***
0.014
0.058
0.053
(0.105)
(0.039)
(0.177)
(0.043)
(0.111)
(0.033)
0.003
0.334***
0.000
0.750***
0.597***
0.106***
Core {0,1} × CM 1[t ≥ 2010]
0.361**
0.006
0.956***
0.061*
0.007
0.029
(0.142)
(0.041)
(0.111)
(0.031)
(0.088)
(0.034)
0.0109
0.8791
0.0000
0.0519
0.9383
0.3979
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,294
169,875
95,717
98,963
R-Squared
0.34
0.16
0.26
0.22
0.28
0.19
R-Squared -Within
0.34
0.10
0.24
0.22
0.28
0.19
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I produce a weighted
regression, employing the survey sampling weights provided by the DHS and accounting for the pooling across
countries and years by standardizing the weights for each country-survey round pair.
254
Table C.36: Robustness Check: Excluding Low-Precision Localities (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.09]
[0.22]
[0.00]
[0.63]
[1.75]
[2.03]
Border [0,1] × EAC 1[2002-2004]
0.288
0.088
0.137
0.139
0.006
0.127
(0.448)
(0.091)
(0.092)
(0.123)
(0.202)
(0.399)
0.521
0.337***
0.137
0.256**
0.978***
0.751**
Border [0,1] × CU 1[2005-2009]
0.574
0.106
0.096
0.109
0.067
0.297
(0.406)
(0.097)
(0.141)
(0.110)
(0.187)
(0.381)
0.158
0.277*
0.497***
0.321***
0.723**
0.437***
Border [0,1] × CM 1[t ≥ 2010]
0.282
0.142***
0.170
0.063
0.337**
-0.102
(0.373)
(0.052)
(0.117)
(0.083)
(0.145)
(0.380)
(0.448)
(0.006)
(0.144)
(0.447)
(0.020)
(0.789)
Core {0,1} × EAC 1[2002-2004]
-0.008
0.036
-0.139***
0.074
-0.011
0.336***
(0.092)
(0.036)
(0.030)
(0.058)
(0.092)
(0.123)
0.935***
0.316***
0.000
0.205***
0.904***
0.006
Core {0,1} × CU 1[2005-2009]
-0.238**
0.054
0.757**
0.029
-0.067
0.142
(0.108)
(0.043)
(0.366)
(0.070)
(0.147)
(0.134)
0.027
0.214***
0.039
0.685***
0.647***
0.289
Core {0,1} × CM 1[t ≥ 2010]
-0.081
0.019
1.041***
-0.217**
-0.090
0.383***
(0.078)
(0.050)
(0.213)
(0.089)
(0.107)
(0.110)
0.3033
0.6986
0.0000
0.0149
0.4017
0.0005
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
15,507
12,116
1,926
13,074
10,345
15,479
R-Squared
0.13
0.14
0.31
0.23
0.27
0.09
R-Squared -Within
0.10
0.09
0.27
0.23
0.27
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I drop “low-precision”
localities from the regression. This is implemented by dropping all observations for which the AidData precision
code is above 2.
255
Table C.37: Robustness Check: Excluding Low-Precision Localities (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.09]
[0.16]
[0.00]
[0.77]
[1.48]
[0.45]
Border [0,1] × EAC 1[2002-2004]
-0.259
-0.103
0.519*
0.123*
-0.186
-0.174
(0.314)
(0.084)
(0.306)
(0.063)
(0.227)
(0.146)
0.409
0.220*
0.090
0.052
0.412
0.234
Border [0,1] × CU 1[2005-2009]
-0.402*
-0.126**
0.217
0.135***
-0.308***
-0.284***
(0.226)
(0.054)
(0.138)
(0.046)
(0.113)
(0.081)
0.076
0.019
0.117***
0.003
0.006
0.000
Border [0,1] × CM 1[t ≥ 2010]
-0.446**
-0.034
0.000
0.036
-0.143*
-0.142
(0.192)
(0.039)
(0.077)
(0.043)
(0.074)
(0.093)
0.020
0.389***
1.000***
0.410***
0.051
0.126***
Core {0,1} × EAC 1[2002-2004]
0.439***
0.072*
-0.130
0.065*
-0.039
-0.003
(0.147)
(0.040)
(0.303)
(0.039)
(0.069)
(0.030)
0.003
0.076**
0.667
0.097***
0.565***
0.920***
Core {0,1} × CU 1[2005-2009]
0.326***
0.038
0.734**
0.015
0.063
0.052
(0.107)
(0.038)
(0.293)
(0.043)
(0.113)
(0.035)
0.002
0.321***
0.012
0.735***
0.577***
0.138***
Core {0,1} × CM 1[t ≥ 2010]
0.342**
0.006
0.950***
0.035*
0.058
0.042
(0.157)
(0.043)
(0.103)
(0.019)
(0.075)
(0.041)
0.0291
0.8917
0.0000
0.0689
0.4378
0.3066
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
149,062
61,998
3,224
140,076
75,568
78,587
R-Squared
0.33
0.17
0.31
0.21
0.28
0.17
R-Squared -Within
0.32
0.12
0.29
0.21
0.28
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I drop "low-precision"
localities from the regression. This is implemented by dropping all observations for which coordinates are not
generated from a GPS receiver used by the fieldworker.
256
Table C.38: Robustness Check: Agglomerations vs. Core (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.66]
[2.09]
Border [0,1] × EAC 1[2002-2004]
-0.121
0.061
0.098
0.084
0.029
-0.253
(0.293)
(0.059)
(0.067)
(0.074)
(0.131)
(0.287)
0.681***
0.295***
0.146***
0.257***
0.824***
0.377***
Border [0,1] × CU 1[2005-2009]
0.170
0.062
0.027
0.093
0.044
-0.102
(0.290)
(0.065)
(0.069)
(0.074)
(0.120)
(0.307)
0.559***
0.344***
0.693***
0.206***
0.713***
0.738**
Border [0,1] × CM 1[t ≥ 2010]
0.132***
0.093***
0.014***
0.016***
0.265***
-0.319***
(0.280)
(0.042)
(0.087)
(0.077)
(0.090)
(0.286)
Agglomeration {0,1} × EAC 1[2002-2004]
0.108
-0.056**
0.022
-0.060*
-0.046
-0.199*
(0.078)
(0.028)
(0.082)
(0.034)
(0.079)
(0.115)
0.165***
0.042***
0.792***
0.078***
0.561***
0.083
Agglomeration {0,1} × CU 1[2005-2009]
-0.036
-0.006
0.129
-0.011
0.001
-0.350***
(0.092)
(0.031)
(0.102)
(0.035)
(0.075)
(0.115)
0.694***
0.850***
0.206
0.746***
0.988***
0.002
Agglomeration {0,1} × CM 1[t ≥ 2010]
0.024***
-0.011***
0.242***
-0.063***
0.042***
-0.334***
(0.080)
(0.027)
(0.150)
(0.045)
(0.071)
(0.117)
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.29
0.23
0.26
0.10
R-Squared -Within
0.10
0.10
0.26
0.23
0.26
0.10
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The Core {0,1} dummy is
replaced with Agglomeration {0,1} which switches to 1 for households living within 50km of an “urban center”
demarcated as such in the year 2000 (see manuscript).
257
Table C.39: Robustness Check: Agglomerations vs. Core (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.74]
[1.48]
[0.49]
Border [0,1] × EAC 1[2002-2004]
-0.571*
-0.171**
-0.063
0.125**
-0.219
-0.188
(0.310)
(0.083)
(0.129)
(0.056)
(0.190)
(0.136)
0.066
0.041
0.626***
0.026
0.250
0.167
Border [0,1] × CU 1[2005-2009]
-0.533**
-0.152***
-0.006
0.117***
-0.317***
-0.289***
(0.237)
(0.045)
(0.106)
(0.042)
(0.106)
(0.084)
0.024
0.001
0.952***
0.005
0.003
0.001
Border [0,1] × CM 1[t ≥ 2010]
-0.605***
-0.036***
-0.020***
0.092***
-0.168***
-0.155***
(0.211)
(0.031)
(0.105)
(0.045)
(0.067)
(0.091)
Agglomeration {0,1} × EAC 1[2002-2004]
0.102
-0.044
0.256
0.015
-0.147**
-0.067
(0.153)
(0.043)
(0.226)
(0.025)
(0.060)
(0.046)
0.506***
0.308***
0.256
0.560***
0.015
0.145***
Agglomeration {0,1} × CU 1[2005-2009]
-0.033
-0.052*
0.208*
0.034*
-0.117**
-0.029
(0.108)
(0.027)
(0.111)
(0.019)
(0.050)
(0.041)
0.760***
0.055
0.062
0.077***
0.019
0.480***
Agglomeration {0,1} × CM 1[t ≥ 2010]
-0.239***
-0.041***
0.258***
0.002***
-0.070***
-0.016***
(0.090)
(0.017)
(0.169)
(0.022)
(0.033)
(0.031)
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.34
0.16
0.21
0.23
0.26
0.17
R-Squared -Within
0.33
0.11
0.20
0.23
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that
most of the respective notes apply here. However, note the following important amendments: The Core {0,1}
dummy is replaced with Agglomeration {0,1} which switches to 1 for households living within 50km of an “urban
center” demarcated as such in the year 2000 (see manuscript).
258
Table C.40: Robustness Check: Agglomerations vs. Core (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border [0,1] × EAC 1[2004]
-1039.542
131.910
0.510
0.080
0.098
-0.980**
(1026.698)
(700.554)
(0.324)
(0.052)
(0.251)
(0.465)
0.311
0.8508
0.116
0.1285
0.697
0.035
Border [0,1] × CU 1[2010]
-1160.586
0.347
0.050
-1.387**
(1116.591)
(0.315)
(0.397)
(0.602)
0.299
0.270***
0.900***
0.021
Agglomeration {0,1} × EAC 1[2004]
191.119***
702.325***
-0.006
0.025***
0.134***
0.811***
(65.909)
(125.914)
(0.023)
(0.009)
(0.021)
(0.027)
0.004
0.0000
0.776
0.0072
0.000
0.000
Agglomeration {0,1} × CU 1[2010]
259.145***
-0.036
0.076
0.754***
(77.785)
(0.023)
(0.072)
(0.051)
0.0009
0.1197
0.2891
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,657
2,767
25,372
14,367
15,817
2,377
Observations - Fixed Effects
3,932
2,427
12,987
7,058
6,309
12
R-Squared
0.87
0.97
0.64
0.72
0.82
0.40
R-Squared -Within
0.12
0.09
0.04
0.01
0.14
0.38
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: The Core {0,1} dummy is replaced
with Agglomeration {0,1} which switches to 1 for households living within 50km of an “urban center” demarcated
as such in the year 2000 (see manuscript).
259
Table C.41: Robustness Check: Agglomerations vs. Core (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border [0,1] × EAC 1[2004]
-335.740
-720.571
7606.797
-0.108
-100.772
8.349
(269.364)
(986.817)
(14133.842)
(0.344)
(137.806)
(8.808)
0.213
0.465
0.5908
0.753
0.4649
0.343
Border [0,1] × CU 1[2010]
-316.002
-838.609
-0.216
8.743
(277.875)
(1070.563)
(0.329)
(8.815)
0.256
0.434
0.511***
0.321
Agglomeration {0,1} × EAC 1[2004] 131.930***
48.610
4019.615***
-0.076
8.229
-1.135
(47.167)
(36.410)
(1307.509)
(0.057)
(25.101)
(1.375)
0.005
0.182
0.0023
0.180
0.7431
0.409
Agglomeration {0,1} × CU 1[2010]
136.575***
113.736***
-0.005
-0.870
(49.804)
(43.356)
(0.043)
(1.335)
0.0062
0.0088
0.9126
0.5146
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,659
5,693
2,767
16,687
1,813
5,550
Observations - Fixed Effects
3,933
3,947
2,427
11,851
1,220
3,830
R-Squared
0.87
0.83
0.71
0.85
0.95
0.89
R-Squared -Within
0.11
0.09
0.07
0.15
0.18
0.08
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: The Core {0,1} dummy is replaced
with Agglomeration {0,1} which switches to 1 for households living within 50km of an “urban center” demarcated
as such in the year 2000 (see manuscript).
260
Table C.42: Robustness Check: Core ≤ 25km (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
Border [0,1] × EAC 1[2002-2004]
-0.082
0.048
0.076
0.068
0.008
-0.341
(0.295)
(0.059)
(0.057)
(0.077)
(0.135)
(0.295)
0.781***
0.414***
0.178***
0.378***
0.951***
0.248
Border [0,1] × CU 1[2005-2009]
0.170
0.059
0.042
0.088
0.033
-0.231
(0.292)
(0.065)
(0.070)
(0.073)
(0.118)
(0.313)
0.560***
0.360***
0.547***
0.226***
0.778***
0.460
Border [0,1] × CM 1[t ≥ 2010]
0.139***
0.097***
0.077***
-0.002***
0.285***
-0.453***
(0.282)
(0.039)
(0.079)
(0.075)
(0.086)
(0.292)
Core ≤ 25km {0,1} × EAC
-0.301***
-0.037
-0.208**
0.035
0.009
-0.021
(0.076)
(0.061)
(0.087)
(0.076)
(0.079)
(0.075)
0.000
0.551***
0.017
0.644***
0.906***
0.774***
Core ≤ 25km {0,1} × CU
-0.520***
-0.051
0.728*
-0.040
-0.049
-0.244**
(0.098)
(0.061)
(0.423)
(0.071)
(0.058)
(0.105)
0.000
0.404***
0.086
0.570*
0.405***
0.020
Core ≤ 25km {0,1} × CM
-0.375***
-0.120***
1.195***
-0.342***
-0.101***
0.070***
(0.062)
(0.090)
(0.440)
(0.126)
(0.093)
(0.080)
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.34
0.23
0.26
0.10
R-Squared -Within
0.10
0.10
0.32
0.23
0.26
0.10
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The spatial cut-off criteria
for living in core agglomerations is reduced to 25km.
261
Table C.43: Robustness Check: Core ≤ 25km (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.11]
[0.18]
[0.00]
[0.78]
[1.51]
[0.52]
Border [0,1] × EAC 1[2002-2004]
-0.291
-0.166**
0.088
0.117**
-0.239
-0.194
(0.315)
(0.082)
(0.123)
(0.059)
(0.167)
(0.129)
0.356
0.042
0.477***
0.047
0.152
0.132
Border [0,1] × CU 1[2005-2009]
-0.363
-0.144***
0.078
0.119***
-0.293***
-0.274***
(0.226)
(0.053)
(0.071)
(0.042)
(0.106)
(0.083)
0.109
0.007
0.271***
0.005
0.006
0.001
Border [0,1] × CM 1[t ≥ 2010]
-0.546*
-0.036
0.048***
0.082**
-0.153
-0.142
(0.209)
(0.039)
(0.076)
(0.045)
(0.073)
(0.090)
0.071
0.373***
0.006
0.031
0.105***
0.764***
Core ≤ 25km {0,1} × EAC
0.411**
0.102**
-0.028
0.061
-0.016
0.020
(0.172)
(0.047)
(0.293)
(0.041)
(0.076)
(0.035)
0.017
0.030
0.923
0.138***
0.831***
0.569***
Core ≤ 25km {0,1} × CU
0.220**
0.060
0.759***
0.011
0.040
0.070**
(0.108)
(0.049)
(0.160)
(0.044)
(0.132)
(0.036)
0.042
0.219***
0.000
0.809***
0.763***
0.048***
Core ≤ 25km {0,1} × CM
0.296*
-0.004
1.056***
0.061**
-0.023
0.012
(0.162)
(0.049)
(0.180)
(0.030)
(0.084)
(0.040)
0.0706
0.3729
0.0064
0.0311
0.1048
0.7638
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.34
0.17
0.37
0.23
0.26
0.17
R-Squared -Within
0.33
0.11
0.36
0.23
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The spatial cut-off criteria
for living in core agglomerations is reduced to 25km.
262
Table C.44: Robustness Check: Core ≤ 25km (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border [0,1] × EAC 1[2004]
-854.539
187.152
0.580*
0.070
0.107
-0.920**
(1072.198)
(754.481)
(0.331)
(0.049)
(0.261)
(0.390)
0.426
0.8043
0.079
0.1555
0.681
0.018
Border [0,1] × CU 1[2010]
-1319.662
0.432
-0.002
-1.721***
(1154.706)
(0.305)
(0.403)
(0.585)
0.253
0.157***
0.996***
0.003
Core ≤ 25km {0,1} × EAC 1[2004] 170.821***
706.607***
-0.013
0.026***
0.132***
0.930***
(42.732)
(192.323)
(0.023)
(0.008)
(0.020)
(0.102)
0.000
0.0003
0.583
0.0022
0.000
0.000
Core ≤ 25km {0,1} × CU 1[2010]
276.569***
-0.037
0.074
0.987***
(58.235)
(0.026)
(0.068)
(0.148)
0.0000
0.1514
0.2740
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.44
R-Squared -Within
0.12
0.09
0.04
0.01
0.14
0.42
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: The spatial cut-off criteria for living
in core agglomerations is reduced to 25km.
263
Table C.45: Robustness Check: Core ≤ 25km (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border [0,1] × EAC 1[2004]
-177.656
-671.700
7592.679
-0.201
-96.031
5.806
(237.098)
(1053.982)
(14152.494)
(0.377)
(137.206)
(9.285)
0.454
0.524
0.5920
0.594
0.4843
0.532
Border [0,1] × CU 1[2010]
-388.050
-913.067
-0.234
6.415
(236.433)
(1116.349)
(0.344)
(9.148)
0.101
0.414
0.497***
0.483
Core ≤ 25km {0,1} × EAC 1[2004]
122.246***
35.774
4013.722***
-0.058
8.109
-0.995
(38.892)
(36.204)
(1314.084)
(0.047)
(25.258)
(1.439)
0.002
0.323
0.0025
0.221
0.7483
0.489
Core ≤ 25km {0,1} × CU 1[2010]
131.142***
135.227***
-0.007
-0.351
(39.002)
(39.843)
(0.044)
(1.459)
0.0008
0.0007
0.8694
0.8097
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.10
0.09
0.07
0.16
0.18
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: The spatial cut-off criteria for living
in core agglomerations is reduced to 25km.
264
Table C.46: Robustness Check: Core ≤ 10km (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
Border [0,1] × EAC 1[2002-2004]
-0.089
0.051
0.091
0.075
0.017
-0.342
(0.294)
(0.059)
(0.065)
(0.078)
(0.136)
(0.295)
0.763***
0.387***
0.158***
0.339***
0.900***
0.248
Border [0,1] × CU 1[2005-2009]
0.151
0.060
0.061
0.090
0.039
-0.237
(0.291)
(0.065)
(0.074)
(0.073)
(0.118)
(0.313)
0.605***
0.357***
0.414***
0.220***
0.739**
0.450
Border [0,1] × CM 1[t ≥ 2010]
0.125
0.096**
0.147
-0.018
0.306***
-0.461
(0.281)
(0.040)
(0.102)
(0.076)
(0.087)
(0.292)
(0.656)
(0.016)
(0.150)
(0.817)
(0.000)
(0.115)
Core ≤ 10km {0,1} × EAC
-0.309***
-0.090
0.098
0.007
-0.192***
-0.105
(0.095)
(0.060)
(0.132)
(0.057)
(0.066)
(0.076)
0.001
0.134**
0.458
0.902***
0.004
0.166
Core ≤ 10km {0,1} × CU
-0.496***
-0.055
0.914*
-0.055
-0.209***
-0.341***
(0.102)
(0.044)
(0.529)
(0.041)
(0.066)
(0.112)
0.000
0.209*
0.084
0.176
0.001
0.002
Core ≤ 10km {0,1} × CM
-0.368***
-0.126
1.554***
-0.351***
-0.313***
0.091
(0.068)
(0.079)
(0.597)
(0.097)
(0.114)
(0.082)
0.0000
0.1079
0.0093
0.0003
0.0061
0.2681
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.36
0.23
0.25
0.09
R-Squared -Within
0.10
0.10
0.34
0.23
0.25
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The spatial cut-off criteria
for living in core agglomerations is reduced to 10km.
265
Table C.47: Robustness Check: Core ≤ 10km (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.11]
[0.18]
[0.00]
[0.78]
[1.51]
[0.52]
Border [0,1] × EAC 1[2002-2004]
-0.235
-0.138
0.134
0.116**
-0.212
-0.172
(0.314)
(0.095)
(0.113)
(0.059)
(0.184)
(0.142)
0.455
0.145
0.235*
0.048
0.248
0.226
Border [0,1] × CU 1[2005-2009]
-0.326
-0.132**
0.129
0.118***
-0.278**
-0.265***
(0.232)
(0.055)
(0.081)
(0.041)
(0.113)
(0.087)
0.161
0.016
0.113
0.004
0.014
0.002
Border [0,1] × CM 1[t ≥ 2010]
-0.492**
-0.027
0.119*
0.082*
-0.132*
-0.129
(0.212)
(0.039)
(0.071)
(0.044)
(0.073)
(0.091)
0.020
0.494***
0.095
0.065
0.072
0.158***
Core ≤ 10km {0,1} × EAC
0.420**
0.031
-0.110
0.062
-0.083
-0.019
(0.175)
(0.043)
(0.330)
(0.041)
(0.058)
(0.062)
0.017
0.466***
0.739
0.135***
0.154***
0.756***
Core ≤ 10km {0,1} × CU
0.156
0.030
0.894***
-0.003
-0.008
0.046
(0.108)
(0.029)
(0.249)
(0.035)
(0.125)
(0.062)
0.150
0.308***
0.000
0.924***
0.950***
0.458***
Core ≤ 10km {0,1} × CM
0.270*
-0.024
1.637***
0.052**
-0.087
-0.023
(0.149)
(0.027)
(0.600)
(0.024)
(0.054)
(0.075)
0.0706
0.3729
0.0064
0.0311
0.1048
0.7638
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.32
0.16
0.41
0.23
0.26
0.16
R-Squared -Within
0.31
0.11
0.41
0.23
0.26
0.16
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The spatial cut-off criteria
for living in core agglomerations is reduced to 10km.
266
Table C.48: Robustness Check: Core ≤ 10km (KHDS 1/2)
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border [0,1] × EAC 1[2004]
-885.490
187.211
0.575*
0.067
0.134
-1.111**
(1142.344)
(767.899)
(0.321)
(0.051)
(0.280)
(0.474)
0.438
0.8075
0.074
0.1874
0.632
0.019
Border [0,1] × CU 1[2010]
-1299.896
0.440
0.083
-2.064***
(1227.192)
(0.303)
(0.485)
(0.651)
0.290
0.147***
0.864***
0.002
Core ≤ 10km {0,1} × EAC 1[2004]
166.469***
694.026***
-0.010
0.027***
0.110***
1.004***
(42.819)
(149.664)
(0.023)
(0.009)
(0.015)
(0.135)
0.000
0.0000
0.664
0.0016
0.000
0.000
Core ≤ 10km {0,1} × CU 1[2010]
228.369***
-0.042*
0.009
1.279***
(58.904)
(0.025)
(0.080)
(0.279)
0.0001
0.0944
0.9111
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
NO
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.50
R-Squared -Within
0.11
0.09
0.04
0.01
0.13
0.48
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: The spatial cut-off criteria for living
in core agglomerations is reduced to 10km.
267
Table C.49: Robustness Check: Core ≤ 10km (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[1885042.91]
[7.81]
Border [0,1] × EAC 1[2004]
-184.828
-696.169
8211.182
-0.215
-83.871
4.575
(238.949)
(1110.470)
(15101.789)
(0.378)
(133.064)
(10.346)
0.439
0.531
0.5870
0.571
0.5287
0.658
Border [0,1] × CU 1[2010]
-375.958
-904.913
-0.275
5.905
(259.850)
(1165.691)
(0.339)
(10.004)
0.148
0.438
0.417***
0.555
Core ≤ 10km {0,1} × EAC 1[2004] 117.794***
36.303
3592.924
-0.058
7.206
-0.380
(34.197)
(45.837)
(2577.006)
(0.042)
(19.772)
(1.427)
0.001
0.428
0.1642
0.161
0.7156
0.790
Core ≤ 10km {0,1} × CU 1[2010]
116.420***
101.922**
-0.001
-0.116
(36.746)
(42.044)
(0.037)
(1.467)
0.0016
0.0154
0.9855
0.9372
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,786
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,194
3,763
R-Squared
0.88
0.83
0.69
0.86
1.00
0.89
R-Squared -Within
0.09
0.08
0.06
0.15
0.17
0.08
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: The spatial cut-off criteria for living
in core agglomerations is reduced to 10km.
268
Table C.50: Post-CM Estimate (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.66]
[2.09]
Border {0,1} × EAC
-0.080
0.049
0.098
0.072
0.017
-0.346
(0.295)
(0.058)
(0.069)
(0.077)
(0.134)
(0.295)
0.788***
0.401***
0.160**
0.347***
0.898***
0.241
Border {0,1} × CU
0.176
0.059
0.064
0.091
0.041
-0.229
(0.292)
(0.065)
(0.081)
(0.073)
(0.119)
(0.313)
0.548***
0.359***
0.427***
0.212***
0.733*
0.464
Border {0,1} × CM 1[2010-2014]
0.039
0.120**
0.084
-0.076
0.379***
-0.611**
(0.293)
(0.051)
(0.118)
(0.103)
(0.111)
(0.274)
0.893***
0.018
0.478***
0.459***
0.001
0.026
Border {0,1} × post-CM 1[t ≥ 2015]
0.277
0.087**
0.166
0.036
0.244***
-0.252
(0.284)
(0.039)
(0.108)
(0.078)
(0.093)
(0.334)
0.330
0.025**
0.123
0.640***
0.009
0.450***
Core {0,1} × EAC
-0.261***
-0.013
-0.189***
0.026
0.021
0.042
(0.077)
(0.050)
(0.052)
(0.067)
(0.078)
(0.082)
0.001
0.800***
0.000
0.693***
0.791***
0.610***
Core {0,1} × CU
-0.470***
-0.006
0.731**
0.003
0.056
-0.233**
(0.097)
(0.045)
(0.335)
(0.055)
(0.091)
(0.105)
0.000
0.899***
0.029
0.962***
0.539***
0.026
Core {0,1} × CM 1[2010-2014]
-0.344***
-0.044
1.263***
-0.256**
-0.081
0.053
(0.066)
(0.080)
(0.409)
(0.123)
(0.088)
(0.094)
0.000
0.585***
0.002
0.038
0.360***
0.571***
Core {0,1} × post-CM 1[t ≥ 2015]
-0.332***
-0.097
0.897**
-0.259**
0.037
0.027
(0.066)
(0.070)
(0.393)
(0.104)
(0.071)
(0.082)
0.0000
0.1688
0.0224
0.0129
0.6064
0.7386
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.28
0.23
0.26
0.10
R-Squared -Within
0.10
0.10
0.25
0.23
0.26
0.10
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The CM period dummy is
split into a two dummies, i.e. timeframes, namely CM 1[2010-2014] and a post-CM time period post-CM 1[t ≥ 2015]
to provide a test on the hypothesized transitory shift to a welfare-equalizing equilibrium.
269
Table C.51: Post-CM Estimate (DHS)
Agglom.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
Border {0,1} × EAC
-0.271
-0.114
0.124
0.118**
-0.158
-0.154
(0.304)
(0.086)
(0.133)
(0.060)
(0.186)
(0.131)
0.373
0.186
0.350***
0.048
0.395
0.240
Border {0,1} × CU
-0.363*
-0.135**
0.095
0.121***
-0.284***
-0.270***
(0.220)
(0.054)
(0.085)
(0.043)
(0.107)
(0.083)
0.099
0.011
0.266***
0.005
0.008
0.001
Border {0,1} × CM 1[2010-2014]
-0.632***
-0.049
0.075
0.125**
-0.191**
-0.108
(0.226)
(0.032)
(0.110)
(0.055)
(0.090)
(0.097)
0.005
0.132***
0.494***
0.024
0.034
0.263*
Border {0,1} × post-CM 1[t ≥ 2015]
-0.443**
-0.023
0.066
0.026
-0.113
-0.158
(0.210)
(0.047)
(0.082)
(0.056)
(0.096)
(0.104)
0.035
0.630***
0.419***
0.641***
0.235***
0.128***
Core {0,1} × EAC
0.422***
0.066*
0.121
0.064*
-0.065
-0.004
(0.131)
(0.036)
(0.257)
(0.038)
(0.049)
(0.023)
0.001
0.066*
0.639
0.088***
0.185***
0.850***
Core {0,1} × CU
0.310***
0.035
0.755***
0.012
0.047
0.052
(0.104)
(0.038)
(0.184)
(0.044)
(0.109)
(0.032)
0.003
0.357***
0.000
0.777***
0.667***
0.104*
Core {0,1} × CM 1[2010-2014]
0.365***
0.009
1.275***
0.080
-0.061
0.062
(0.114)
(0.037)
(0.217)
(0.050)
(0.072)
(0.054)
0.001
0.808***
0.000
0.105***
0.396***
0.254***
Core {0,1} × post-CM 1[t ≥ 2015]
0.349**
0.001
0.636***
0.041
0.034
0.006
(0.168)
(0.042)
(0.094)
(0.025)
(0.081)
(0.036)
0.0373
0.9849
0.0000
0.1035
0.6751
0.8698
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.34
0.16
0.30
0.23
0.26
0.17
R-Squared -Within
0.33
0.11
0.29
0.23
0.26
0.17
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: The CM period dummy is
split into a two dummies, i.e. timeframes, namely CM 1[2010-2014] and a post-CM time period post-CM 1[t ≥ 2015]
to provide a test on the hypothesized transitory shift to a welfare-equalizing equilibrium.
270
Table C.52: Robustness Check: Excluding “Post-EAC Migrants” (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.30]
[0.14]
[0.00]
[0.70]
[1.39]
[0.50]
Border [0,1] × EAC 1[2002-2004]
-1.316***
-0.017
0.530*
0.522***
-0.090
-0.250
(0.402)
(0.073)
(0.293)
(0.126)
(0.222)
(0.184)
0.001
0.815***
0.071
0.000
0.684*
0.173
Border [0,1] × CU 1[2005-2009]
-0.694**
-0.155***
0.178
0.328***
-0.389***
-0.402***
(0.296)
(0.044)
(0.141)
(0.081)
(0.133)
(0.084)
0.019
0.000
0.207***
0.000
0.003
0.000
Border [0,1] × CM 1[t ≥ 2010]
-0.355
0.001
0.041
0.090
-0.159
-0.124
(0.254)
(0.049)
(0.074)
(0.072)
(0.104)
(0.129)
0.162
0.987***
0.585***
0.209*
0.124***
0.336***
Core {0,1} × EAC 1[2002-2004]
0.605***
0.066
-0.104
0.110**
-0.037
0.027
(0.145)
(0.045)
(0.266)
(0.044)
(0.064)
(0.046)
0.000
0.143***
0.695
0.013
0.565***
0.554***
Core {0,1} × CU 1[2005-2009]
0.375***
0.046
0.760***
0.037
0.034
0.137***
(0.122)
(0.038)
(0.279)
(0.091)
(0.099)
(0.047)
0.002
0.228***
0.006
0.683***
0.730***
0.004
Core {0,1} × CM 1[t ≥ 2010]
0.256**
-0.021
0.734***
0.062
0.023
0.116**
(0.114)
(0.060)
(0.126)
(0.041)
(0.070)
(0.053)
0.0247
0.7236
0.0000
0.1317
0.7413
0.0302
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
33,944
24,740
2,706
34,053
24,715
24,823
R-Squared
0.36
0.14
0.29
0.23
0.27
0.19
R-Squared -Within
0.34
0.10
0.27
0.23
0.27
0.19
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I exclude the individuals
in the sample which did not live in the survey location at least three years before the establishment of the EAC
(before 1999), i.e. excludes “post-EAC Migrants”.
271
Table C.53: Robustness Check: Excluding “Post-EAC Migrants” (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[491.72]
[114.29]
[0.10]
[0.01]
[1.17]
[0.00]
Border [0,1] × EAC 1[2004]
20.962
195.460
0.492
-0.001
-0.036
-0.838**
(341.477)
(644.348)
(0.312)
(0.076)
(0.213)
(0.375)
0.951
0.7618
0.115
0.9930
0.867
0.026
Border [0,1] × CU 1[2010]
-206.044
0.591**
-0.684
-1.762***
(331.479)
(0.279)
(0.577)
(0.614)
0.534
0.034***
0.236*
0.004
Core {0,1} × EAC 1[2004]
119.373***
772.930***
0.009
0.017*
0.143***
0.930***
(42.378)
(91.216)
(0.022)
(0.010)
(0.026)
(0.092)
0.005
0.0000
0.704
0.0920
0.000
0.000
Core {0,1} × CU 1[2010]
225.295***
-0.049*
0.141*
0.962***
(46.662)
(0.026)
(0.074)
(0.102)
0.0000
0.0567
0.0556
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
4,480
2,392
22,531
13,600
14,418
2,010
Observations - Fixed Effects
3,155
2,086
12,318
6,718
5,990
12
R-Squared
0.90
0.97
0.68
0.74
0.84
0.45
R-Squared -Within
0.16
0.09
0.03
0.02
0.07
0.43
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: I exclude the individuals in the sample
which did not live in the survey location at least three years before the establishment of the EAC (before 1999),
i.e. excludes “post-EAC Migrants”.
272
Table C.54: Robustness Test: Excluding “Post-EAC Migrants” (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[316.66]
[174.90]
[670.91]
[0.27]
[26.19]
[10.84]
Border [0,1] × EAC 1[2004]
-132.988
157.347
7986.289
0.000
-141.718
-0.544
(261.808)
(148.747)
(14898.392)
(0.352)
(187.157)
(18.589)
0.612
0.290
0.5923
0.999
0.4493
0.977
Border [0,1] × CU 1[2010]
-299.371
99.518
-0.074
1.058
(259.523)
(163.821)
(0.337)
(19.004)
0.249
0.544*
0.827***
0.956**
Core {0,1} × EAC 1[2004]
106.040***
0.311
4599.535***
-0.077
9.488
-0.456
(36.417)
(16.952)
(1202.895)
(0.071)
(25.859)
(2.988)
0.004
0.985
0.0002
0.281
0.7138
0.879
Core {0,1} × CU 1[2010]
140.531***
74.285**
0.009
-0.171
(35.606)
(36.188)
(0.047)
(2.939)
0.0001
0.0403
0.8421
0.9537
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
4,482
4,495
2,392
14,166
1,586
2,748
Observations - Fixed Effects
3,157
3,162
2,086
10,659
1,037
2,277
R-Squared
0.88
0.88
0.68
0.89
0.94
0.96
R-Squared -Within
0.11
0.13
0.08
0.17
0.11
0.24
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: I exclude the individuals in the sample
which did not live in the survey location at least three years before the establishment of the EAC (before 1999),
i.e. excludes “post-EAC Migrants”.
273
Table C.55: Robustness Test: DHS only (excluding AIS, MIS, KAP)(DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.19]
[0.00]
[0.75]
[1.51]
[0.53]
Border [0,1] × EAC 1[2002-2004]
-0.284
-0.129
0.365
0.413***
-0.184
-0.163
(0.312)
(0.089)
(0.284)
(0.113)
(0.188)
(0.136)
0.362
0.144
0.199
0.000
0.328
0.230
Border [0,1] × CU 1[2005-2009]
-0.372*
-0.138**
0.140
0.161***
-0.291***
-0.269***
(0.224)
(0.054)
(0.140)
(0.054)
(0.106)
(0.097)
0.096
0.011
0.318***
0.003
0.006
0.006
Border [0,1] × CM 1[t ≥ 2010]
-0.547***
-0.034***
0.084***
0.098***
-0.149***
-0.141***
(0.202)
(0.040)
(0.105)
(0.049)
(0.075)
(0.092)
Core {0,1} × EAC 1[2002-2004]
0.433***
0.073*
-0.133
0.096***
-0.046
0.000
(0.141)
(0.040)
(0.270)
(0.027)
(0.062)
(0.027)
0.002
0.071*
0.621
0.000
0.454***
0.989***
Core {0,1} × CU 1[2005-2009]
0.314***
0.037
0.665**
0.019
0.058
0.073*
(0.105)
(0.039)
(0.267)
(0.059)
(0.111)
(0.042)
0.003
0.334***
0.013
0.742***
0.597***
0.080***
Core {0,1} × CM 1[t ≥ 2010]
0.361***
0.006***
1.045***
0.076***
0.007***
0.029***
(0.142)
(0.041)
(0.112)
(0.038)
(0.088)
(0.035)
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
4,843
125,539
95,717
96,277
R-Squared
0.34
0.16
0.29
0.22
0.26
0.17
R-Squared -Within
0.33
0.11
0.28
0.22
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: Excludes DHS’ special
survey rounds including the AIS, KAP and MIS surveys.
274
Table C.56: Robustness Check: Logged Dependent Variables (KHDS 1/2)
Agglom.
Log(1+
Annual p.c.
Consumption
(TZS '000)
Log(1+
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Log(1+
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[6.06]
[112.23]
-
-
-
[0.60]
Border [0,1] * EAC 1[2004]
0.008
-0.130
-
-
-
-1.313***
(0.932)
(3.355)
(0.354)
0.993
0.9691
0.000
Border [0,1] * CU 1[2010]
-0.835
-
-
-
-
-2.066***
(0.869)
(0.576)
0.336
0.000
Core {0,1} * EAC 1[2004]
0.216*
0.318
-
-
-
1.185***
(0.114)
(0.349)
(0.025)
0.058
0.3630
0.000
Core {0,1} * CU 1[2010]
0.306***
-
-
-
-
1.052***
(0.113)
(0.032)
0.0070
0.0000
Individual Controls
YES
YES
-
-
-
NO
Geographic Controls
YES
YES
-
-
-
YES
Individual Fixed Effects
NO
NO
-
-
-
NO
Household Fixed Effects
YES
YES
-
-
-
NO
Country-Year Fixed Effect
YES
YES
-
-
-
YES
Observations
5,492
2,695
-
-
-
2,292
Observations - Fixed Effect
3,816
2,363
-
-
-
12
R-Squared
0.90
0.95
-
-
-
0.97
R-Squared -Within
0.14
0.08
-
-
-
0.48
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: Relevant dependent variables are
"logged", i.e. are transformed in the way log(1+Y).
275
Table C.57: Robustness Check: Logged Dependent Variables (KHDS 2/2)
Agglom.
Log(1+
Ann. Food
Consumption
(TZS '000)
Log(1+
Ann. non-Food
Consumption
(TZS '000)
Log(1+
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Log(1+
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
[5.62]
[4.89]
[4.18]
-
[2.79]
-
Border [0,1] × EAC 1[2004]
-0.179
0.293
1.416
-
-4.955
-
(1.032)
(1.305)
(4.798)
(4.548)
0.862
0.823
0.7681
0.2764
Border [0,1] × CU 1[2010]
-1.131
-0.099
-
-
-
-
(0.978)
(1.314)
0.248
0.940***
Core {0,1} × EAC 1[2004]
0.271*
0.025
0.547
-
0.269
-
(0.140)
(0.139)
(0.566)
(0.495)
0.053
0.857
0.3344
0.5862
Core {0,1} × CU 1[2010]
0.300**
0.154
-
-
-
-
(0.126)
(0.180)
0.0170
0.3926
Individual Controls
YES
YES
YES
-
YES
-
Geographic Controls
YES
YES
YES
-
YES
-
Individual Fixed Effects
NO
NO
NO
-
YES
-
Household Fixed Effects
YES
YES
YES
-
NO
-
Country-Year Fixed Effects
YES
YES
YES
-
YES
-
Observations
5,494
5,524
2,695
-
1,782
-
Observations - Fixed Effects
3,817
3,830
2,363
-
1,190
-
R-Squared
0.86
0.90
0.96
-
0.89
-
R-Squared -Within
0.09
0.13
0.14
-
0.04
-
Dependent Variable
KHDS
Consumpt.
Income
Sample Mean of Dep. Var.
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: Relevant dependent variables are
"logged", i.e. are transformed in the way log(1+Y).
276
Table C.58: Robustness Check: Beeline Distance to Border Crossings (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
(Beeline to) Border [0,1] × EAC
0.096
0.022
0.080
0.012
-0.052
-0.421
(0.295)
(0.054)
(0.062)
(0.068)
(0.131)
(0.325)
0.744***
0.684***
0.196
0.859***
0.689***
0.195
(Beeline to) Border [0,1] × CU
0.201
0.097
0.177
0.113
0.127
-0.211
(0.327)
(0.063)
(0.124)
(0.082)
(0.107)
(0.346)
0.539**
0.125**
0.153
0.170***
0.236
0.541
(Beeline to) Border [0,1] × CM
0.221
0.063
0.264
-0.051
0.256***
-0.433
(0.259)
(0.041)
(0.168)
(0.075)
(0.088)
(0.288)
(0.394)
(0.122)
(0.116)
(0.500)
(0.003)
(0.133)
Core {0,1} × EAC 1[2002-2004]
-0.266***
-0.009
-0.173***
0.031
0.017
0.068
(0.077)
(0.050)
(0.049)
(0.068)
(0.081)
(0.088)
0.000
0.866***
0.000
0.646***
0.832***
0.441**
Core {0,1} × CU 1[2005-2009]
-0.450***
0.006
0.750**
0.016
0.070
-0.178
(0.107)
(0.050)
(0.321)
(0.062)
(0.097)
(0.110)
0.000
0.910***
0.019
0.793***
0.475***
0.104*
Core {0,1} × CM 1[t ≥ 2010]
-0.331***
-0.066
1.098***
-0.255**
0.016
0.062
(0.066)
(0.073)
(0.357)
(0.108)
(0.095)
(0.079)
0.0000
0.3643
0.0021
0.0188
0.8647
0.4306
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.28
0.23
0.25
0.10
R-Squared -Within
0.10
0.10
0.25
0.23
0.25
0.10
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that
most of the respective notes apply here. However, note the following important amendments: I replace road
distances to the border with "beeline" (as the crow flies) distances.
277
Table C.59: Robustness Test: Beeline Distance to Border Crossings (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
(Beeline to) Border [0,1] × EAC
-0.160
-0.141
-0.042
0.076
-0.179
-0.131
(0.312)
(0.091)
(0.139)
(0.054)
(0.194)
(0.135)
0.607**
0.119
0.764***
0.158
0.357*
0.330**
(Beeline to) Border [0,1] × CU
-0.277
-0.135***
0.067
0.121***
-0.207*
-0.141
(0.265)
(0.045)
(0.143)
(0.043)
(0.118)
(0.101)
0.297
0.003
0.638***
0.005
0.079
0.161
(Beeline to) Border [0,1] × CM
-0.550**
-0.066**
-0.027
0.083*
-0.178**
-0.182**
(0.232)
(0.033)
(0.138)
(0.045)
(0.071)
(0.092)
0.018
0.045
0.844***
0.063
0.013
0.048**
Core {0,1} × EAC 1[2002-2004]
0.381**
0.058*
0.103
0.080**
-0.073
-0.024
(0.155)
(0.031)
(0.266)
(0.040)
(0.047)
(0.029)
0.014
0.058***
0.700
0.044
0.122***
0.403***
Core {0,1} × CU 1[2005-2009]
0.271**
0.019
0.767***
0.030
0.023
0.025
(0.122)
(0.031)
(0.178)
(0.039)
(0.098)
(0.034)
0.026
0.540***
0.000
0.437***
0.817***
0.457***
Core {0,1} × CM 1[t ≥ 2010]
0.292**
0.000
0.934***
0.072**
-0.018
0.006
(0.138)
(0.036)
(0.120)
(0.031)
(0.078)
(0.040)
0.0338
0.9943
0.0000
0.0194
0.8232
0.8729
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,364
71,753
7,696
169,967
95,736
98,982
R-Squared
0.34
0.16
0.30
0.23
0.26
0.17
R-Squared -Within
0.33
0.11
0.29
0.23
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace road distances to
the border with "beeline" (as the crow flies) distances.
278
Table C.60: Robustness Check: Beeline Distance to Border Crossings (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
(Beeline to) Border [0,1] × EAC
-816.913
69.400
0.563**
0.051
0.078
-0.710**
(1112.207)
(715.743)
(0.279)
(0.044)
(0.238)
(0.312)
0.463
0.9228
0.044
0.2414
0.742
0.023
(Beeline to) Border [0,1] × CU
-1283.327
0.435
-0.033
-1.567***
(1208.457)
(0.278)
(0.388)
(0.545)
0.288
0.118***
0.932***
0.004
Core {0,1} × EAC 1[2004]
154.777***
717.253***
-0.004
0.027***
0.137***
0.857***
(19.851)
(132.421)
(0.020)
(0.008)
(0.018)
(0.058)
0.000
0.0000
0.843
0.0003
0.000
0.000
Core {0,1} × CU 1[2010]
254.967***
-0.032
0.074
0.864***
(41.741)
(0.022)
(0.065)
(0.099)
0.0000
0.1411
0.2545
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.43
R-Squared -Within
0.12
0.09
0.04
0.01
0.14
0.41
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: I replace road distances to the border
with "beeline" (as the crow flies) distances.
279
Table C.61: Robustness Check: Beeline Distance to Border Crossings (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
(Beeline to) Border [0,1] × EAC
-111.273
-697.103
4638.227
-0.180
-106.417
3.003
(210.184)
(1104.755)
(12272.383)
(0.339)
(145.479)
(8.924)
0.597
0.528
0.7057
0.594
0.4648
0.737
(Beeline to) Border [0,1] × CU
-335.069
-924.540
-0.234
3.458
(212.127)
(1172.534)
(0.319)
(8.728)
0.114
0.431
0.463***
0.692
Core {0,1} × EAC 1[2004]
116.354***
25.677
4250.588***
-0.061
6.974
-0.861
(35.359)
(17.989)
(1163.358)
(0.044)
(23.486)
(1.397)
0.001
0.154
0.0003
0.167
0.7666
0.538
Core {0,1} × CU 1[2010]
123.575***
121.374***
-0.010
-0.192
(36.280)
(18.776)
(0.041)
(1.418)
0.0007
0.0000
0.8090
0.8926
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.10
0.10
0.07
0.16
0.18
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: I replace road distances to the border
with "beeline" (as the crow flies) distances.
280
Table C.62: Robustness Check: Beeline Distance to Border Line (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
(Beeline to) Borderline [0,1] × EAC
-0.045
0.062
0.098*
0.078
0.054
-0.402
(0.282)
(0.058)
(0.060)
(0.077)
(0.129)
(0.297)
0.874***
0.285***
0.100
0.313***
0.672***
0.175
(Beeline to) Borderline [0,1] × CU
0.144
0.051
0.075
0.072
0.025
-0.235
(0.287)
(0.059)
(0.079)
(0.071)
(0.108)
(0.309)
0.615***
0.389***
0.342***
0.315***
0.818***
0.448
(Beeline to) Borderline [0,1] × CM
0.213
0.074**
0.131
-0.054
0.262***
-0.432
(0.258)
(0.037)
(0.106)
(0.071)
(0.080)
(0.284)
(0.407)
(0.047)
(0.216)
(0.446)
(0.001)
(0.129)
Core {0,1} × EAC 1[2002-2004]
-0.270***
-0.011
-0.184***
0.029
0.022
0.064
(0.077)
(0.051)
(0.051)
(0.068)
(0.080)
(0.083)
0.000
0.833***
0.000
0.672***
0.779***
0.437**
Core {0,1} × CU 1[2005-2009]
-0.466***
-0.003
0.733**
0.007
0.059
-0.204**
(0.094)
(0.046)
(0.332)
(0.057)
(0.091)
(0.103)
0.000
0.946***
0.027
0.906***
0.519***
0.048
Core {0,1} × CM 1[t ≥ 2010]
-0.340***
-0.071
1.090***
-0.256**
0.001
0.060
(0.064)
(0.071)
(0.374)
(0.108)
(0.090)
(0.079)
0.0000
0.3167
0.0035
0.0174
0.9902
0.4465
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.27
0.23
0.25
0.09
R-Squared -Within
0.10
0.10
0.25
0.23
0.25
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace road distances to
the border with shortest "beeline" (as the crow flies) distance to the entire borderline.
281
Table C.63: Robustness Check: Beeline Distance to Border Line (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
(Beeline to) Borderline [0,1] × EAC
-0.248
-0.152**
0.026
0.079
-0.168
-0.152
(0.296)
(0.077)
(0.132)
(0.055)
(0.172)
(0.123)
0.404
0.048
0.842***
0.154
0.329
0.214
(Beeline to) Borderline [0,1] × CU
-0.368
-0.106**
0.084
0.112**
-0.211*
-0.220***
(0.235)
(0.049)
(0.107)
(0.044)
(0.108)
(0.084)
0.117
0.029
0.430***
0.010
0.051
0.009
(Beeline to) Borderline [0,1] × CM
-0.559***
-0.024
-0.005
0.063
-0.102
-0.131
(0.199)
(0.036)
(0.109)
(0.045)
(0.064)
(0.083)
0.005
0.499***
0.962***
0.162**
0.111**
0.117***
Core {0,1} × EAC 1[2002-2004]
0.418***
0.069*
0.142
0.073*
-0.055
-0.009
(0.142)
(0.036)
(0.272)
(0.039)
(0.056)
(0.025)
0.003
0.053
0.603
0.058***
0.324***
0.718***
Core {0,1} × CU 1[2005-2009]
0.301***
0.033
0.778***
0.019
0.044
0.038
(0.108)
(0.035)
(0.180)
(0.042)
(0.105)
(0.032)
0.005
0.352***
0.000
0.648***
0.677***
0.232***
Core {0,1} × CM 1[t ≥ 2010]
0.338**
0.007
0.958***
0.065**
0.001
0.023
(0.134)
(0.041)
(0.116)
(0.031)
(0.086)
(0.035)
0.0118
0.8615
0.0000
0.0384
0.9929
0.5118
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,364
71,753
7,696
169,967
95,736
98,982
R-Squared
0.34
0.16
0.30
0.23
0.26
0.17
R-Squared -Within
0.33
0.11
0.29
0.23
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace road distances to
the border with shortest "beeline" (as the crow flies) distance to the entire borderline.
282
Table C.64: Robustness Check: Beeline Distance to Border Line (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
(Beeline to) Borderline [0,1] × EAC
-757.039
122.074
0.490*
0.041
0.062
-0.570*
(998.849)
(577.684)
(0.261)
(0.044)
(0.234)
(0.318)
0.449
0.8328
0.061
0.3524
0.790
0.074
(Beeline to) Borderline [0,1] × CU
-1199.598
0.389
-0.043
-1.360***
(1090.390)
(0.255)
(0.374)
(0.500)
0.271
0.126***
0.908***
0.007
Core {0,1} × EAC 1[2004]
162.273***
716.850***
-0.008
0.027***
0.137***
0.861***
(33.694)
(120.817)
(0.022)
(0.008)
(0.019)
(0.062)
0.000
0.0000
0.712
0.0008
0.000
0.000
Core {0,1} × CU 1[2010]
267.119***
-0.036
0.076
0.872***
(51.500)
(0.023)
(0.067)
(0.102)
0.0000
0.1198
0.2582
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.43
R-Squared -Within
0.12
0.09
0.04
0.01
0.14
0.41
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace road distances to
the border with shortest "beeline" (as the crow flies) distance to the entire borderline.
283
Table C.65: Robustness Check: Beeline Distance to Border Line (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
(Beeline to) Borderline [0,1] × EAC
-107.710
-638.292
4075.452
-0.135
-92.091
4.749
(207.271)
(981.794)
(11014.623)
(0.327)
(133.728)
(8.546)
0.603
0.516
0.7116
0.679
0.4913
0.579
(Beeline to) Borderline [0,1] × CU
-323.756
-853.098
-0.203
5.328
(206.881)
(1047.819)
(0.302)
(8.284)
0.118
0.416
0.503***
0.520
Core {0,1} × EAC 1[2004]
117.191***
31.893
4207.051***
-0.061
8.250
-1.011
(37.145)
(27.424)
(1186.314)
(0.048)
(24.091)
(1.419)
0.002
0.245
0.0005
0.199
0.7321
0.476
Core {0,1} × CU 1[2010]
126.877***
129.708***
-0.009
-0.351
(38.103)
(29.727)
(0.043)
(1.441)
0.0009
0.0000
0.8434
0.8074
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.10
0.09
0.07
0.16
0.18
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace road distances to
the border with shortest "beeline" (as the crow flies) distance to the entire borderline.
284
Table C.66: Robustness Check: Logged Distance (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.66]
[2.09]
Log(BorderDistance) × EAC
0.029
-0.006
-0.018
-0.006
0.015
0.091
(0.071)
(0.014)
(0.022)
(0.019)
(0.031)
(0.069)
0.680***
0.670***
0.415***
0.768***
0.623***
0.189*
Log(BorderDistance) × CU
-0.033
-0.023
-0.014
-0.027
-0.015
0.101
(0.073)
(0.018)
(0.032)
(0.023)
(0.031)
(0.074)
0.646***
0.196***
0.672***
0.247***
0.631***
0.171
Log(BorderDistance) × CM
-0.009
-0.027***
-0.032
0.007
-0.068***
0.143**
(0.067)
(0.010)
(0.039)
(0.021)
(0.020)
(0.066)
(0.892)
(0.006)
(0.414)
(0.750)
(0.001)
(0.031)
Log(CoreDistance) × EAC
-0.022
0.026*
0.031
0.020
0.054*
-0.043
(0.054)
(0.014)
(0.024)
(0.020)
(0.029)
(0.033)
0.679***
0.061***
0.193
0.324***
0.061*
0.186***
Log(CoreDistance) × CU
0.067
0.016
-0.174**
0.022
0.037**
0.063
(0.050)
(0.012)
(0.074)
(0.014)
(0.018)
(0.040)
0.181***
0.163***
0.019
0.130
0.039
0.116***
Log(CoreDistance) × CM
0.009
0.030*
-0.302***
0.088***
0.034
-0.025
(0.054)
(0.016)
(0.093)
(0.025)
(0.026)
(0.030)
0.8658
0.0688
0.0012
0.0004
0.1960
0.4074
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.12
0.13
0.30
0.23
0.25
0.10
R-Squared -Within
0.10
0.10
0.28
0.23
0.25
0.10
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I log-transform the road
distances to the border in the way (1+BorderDistance).
285
Table C.67: Robustness Check: Logged Distance (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
Log(BorderDistance) × EAC
0.029
0.034**
-0.077*
-0.028*
0.018
0.046*
(0.076)
(0.016)
(0.041)
(0.016)
(0.035)
(0.026)
0.700***
0.037
0.057
0.086***
0.614***
0.079
Log(BorderDistance) × CU
0.083
0.040***
-0.053*
-0.024**
0.048
0.075***
(0.057)
(0.013)
(0.032)
(0.011)
(0.030)
(0.023)
0.142
0.002
0.099
0.027*
0.109***
0.001
Log(BorderDistance) × CM
0.120**
0.016
-0.066*
-0.015
0.035**
0.048*
(0.050)
(0.010)
(0.039)
(0.012)
(0.018)
(0.025)
0.016***
0.115***
0.085
0.219***
0.044
0.057***
Log(CoreDistance) × EAC
0.004
-0.004
0.057
-0.023**
0.038**
0.008
(0.042)
(0.006)
(0.043)
(0.011)
(0.017)
(0.017)
0.915***
0.516***
0.184
0.033***
0.021
0.617***
Log(CoreDistance) × CU
-0.017
-0.001
-0.179***
-0.011
0.023
-0.010
(0.027)
(0.008)
(0.055)
(0.011)
(0.025)
(0.011)
0.542***
0.909***
0.001
0.311***
0.347***
0.362***
Log(CoreDistance) × CM
-0.011
0.007
-0.209***
-0.021**
0.027**
-0.005
(0.027)
(0.007)
(0.057)
(0.009)
(0.013)
(0.018)
0.6917
0.2949
0.0003
0.0218
0.0335
0.7668
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.34
0.16
0.31
0.23
0.25
0.17
R-Squared -Within
0.33
0.11
0.30
0.23
0.25
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I log-transform the road
distances to the border in the way (1+BorderDistance).
286
Table C.68: Robustness Check: Logged Distance (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Log(BorderDistance) × EAC
-14.571
-305.802
-0.051
-0.020*
-0.071
-0.404*
(96.504)
(499.702)
(0.043)
(0.011)
(0.052)
(0.225)
0.880
0.5410
0.231
0.0756
0.170
0.072
Log(BorderDistance) × CU
24.227
-0.033
-0.040
-0.330
(115.177)
(0.038)
(0.071)
(0.221)
0.833
0.376
0.574
0.136
Log(CoreDistance) × EAC
-2817.106
-3141.332
0.399
-0.074
-0.723
-4.660**
(3095.358)
(6222.287)
(0.529)
(0.137)
(0.697)
(2.237)
0.363
0.6140
0.450
0.5911
0.299
0.037
Log(CoreDistance) × CU
-2850.940
0.390
-0.711
-4.669**
(3096.469)
(0.527)
(0.696)
(2.230)
0.3573
0.4588
0.3068
0.0364
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,475
2,678
24,726
14,018
15,433
2,288
Observations - Fixed Effects
3,800
2,347
12,660
6,899
6,161
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.49
R-Squared -Within
0.12
0.05
0.03
0.01
0.14
0.47
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following amendments: I log-transform the road distances to
the border in the way (1+BorderDistance).
287
Table C.69: Robustness Check: Logged Distance (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Log(BorderDistance) × EAC
-22.602
13.234
-3084.228
0.014
9.795
-1.344
(56.181)
(62.018)
(3506.668)
(0.060)
(19.092)
(1.748)
0.688
0.831
0.3798
0.815
0.6081
0.442
Log(BorderDistance) × CU
17.695
8.645
0.025
-1.897
(62.345)
(68.685)
(0.041)
(1.590)
0.777
0.900
0.539***
0.233
Log(CoreDistance) × EAC
-519.714
-2241.741
-28757.050
-0.206
49.406
-22.758
(727.313)
(2833.684)
(45558.582)
(0.658)
(300.652)
(21.510)
0.475
0.429
0.5284
0.755
0.8695
0.290
Log(CoreDistance) × CU
-483.284
-2312.587
-0.168
-23.282
(727.883)
(2835.007)
(0.657)
(21.488)
0.5068
0.4148
0.7976
0.2787
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,477
5,507
2,678
16,245
1,767
5,354
Observations - Fixed Effects
3,801
3,814
2,347
11,553
1,180
3,712
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.10
0.10
0.06
0.16
0.18
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that
most of the respective notes apply here. However, note the following amendments: I log-transform the road
distances to the border in the way (1+BorderDistance).
288
Table C.70: Robustness Check: Border ≤ 100km (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
Border ≤ 100km {0,1} × EAC
-0.137
0.002
-0.022
0.011
0.007
-0.125
(0.135)
(0.023)
(0.041)
(0.039)
(0.066)
(0.165)
0.310***
0.938***
0.591***
0.772***
0.914***
0.449**
Border ≤ 100km {0,1} × CU
-0.019
0.034
-0.070
0.075
0.116*
-0.179
(0.156)
(0.035)
(0.050)
(0.051)
(0.065)
(0.169)
0.901***
0.331***
0.165***
0.141***
0.073
0.289
Border ≤ 100km {0,1} × CM
-0.032
0.025
-0.035
-0.017
0.153***
-0.201
(0.131)
(0.022)
(0.056)
(0.048)
(0.058)
(0.153)
(0.804)
(0.268)
(0.528)
(0.718)
(0.009)
(0.190)
Core {0,1} × EAC
-0.292***
-0.010
-0.189***
0.033
0.020
0.010
(0.082)
(0.049)
(0.051)
(0.066)
(0.078)
(0.091)
0.000
0.830***
0.000
0.617***
0.794***
0.910***
Core {0,1} × CU
-0.457***
0.005
0.722**
0.024
0.085
-0.268**
(0.110)
(0.044)
(0.332)
(0.055)
(0.089)
(0.117)
0.000
0.917***
0.030
0.668**
0.341***
0.021**
Core {0,1} × CM
-0.331***
-0.065
1.083***
-0.261**
0.037
-0.011
(0.070)
(0.070)
(0.377)
(0.110)
(0.086)
(0.082)
0.0000
0.3554
0.0040
0.0182
0.6701
0.8953
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.13
0.14
0.27
0.23
0.25
0.09
R-Squared -Within
0.11
0.10
0.25
0.23
0.25
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 100km to the border (using road
distances).
289
Table C.71: Robustness Check: Border ≤ 100km (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
Border ≤ 100km {0,1} × EAC
-0.421
-0.134***
-0.146*
0.023
-0.181***
-0.204***
(0.261)
(0.035)
(0.084)
(0.038)
(0.063)
(0.025)
0.107
0.000
0.082
0.548***
0.004
0.000
Border ≤ 100km {0,1} × CU
-0.345**
-0.099***
-0.051
0.038
-0.144***
-0.177***
(0.167)
(0.023)
(0.058)
(0.023)
(0.048)
(0.034)
0.039
0.000
0.379***
0.108***
0.003
0.000
Border ≤ 100km {0,1} × CM
-0.379**
-0.065***
-0.056
0.024
-0.133***
-0.159***
(0.150)
(0.015)
(0.056)
(0.025)
(0.034)
(0.049)
0.012
0.000
0.319**
0.342***
0.000
0.001
Core {0,1} × EAC
0.360**
0.041
0.138
0.075*
-0.090
-0.047**
(0.157)
(0.041)
(0.275)
(0.039)
(0.060)
(0.020)
0.021
0.315***
0.616
0.053**
0.129***
0.019***
Core {0,1} × CU
0.249*
0.011
0.775***
0.026
0.017
0.006
(0.128)
(0.040)
(0.181)
(0.045)
(0.113)
(0.032)
0.053
0.791***
0.000
0.563***
0.880***
0.849***
Core {0,1} × CM
0.272
-0.006
0.957***
0.068**
-0.024
-0.005
(0.169)
(0.041)
(0.116)
(0.033)
(0.089)
(0.027)
0.1089
0.8918
0.0000
0.0407
0.7890
0.8656
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.34
0.16
0.29
0.23
0.26
0.17
R-Squared -Within
0.33
0.11
0.28
0.23
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 100km to the border (using road
distances).
290
Table C.72: Robustness Check: Border ≤ 100km (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border ≤ 100km {0,1} × EAC
-43.826
21.219
0.033
0.009
0.017
0.038
(85.443)
(83.145)
(0.042)
(0.010)
(0.045)
(0.063)
0.608
0.7987
0.428
0.3685
0.710
0.543
Border ≤ 100km {0,1} × CU
-105.265
0.022
0.127
-0.085
(108.035)
(0.041)
(0.089)
(0.102)
0.330
0.598***
0.154
0.401
Core {0,1} × EAC
147.835***
688.652***
-0.001
0.024**
0.127***
0.809***
(53.216)
(86.586)
(0.028)
(0.010)
(0.010)
(0.069)
0.006
0.0000
0.966
0.0161
0.000
0.000
Core {0,1} × CU
264.581***
-0.027
-0.002
0.844***
(74.478)
(0.030)
(0.082)
(0.118)
0.0004
0.3608
0.9845
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.43
R-Squared -Within
0.12
0.09
0.03
0.01
0.14
0.40
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 100km to the border (using road
distances).
291
Table C.73: Robustness Check: Border ≤ 100km (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border ≤ 100km {0,1} × EAC
-5.045
-38.379
1207.398
-0.019
-8.716
2.250
(46.765)
(67.330)
(1921.629)
(0.068)
(26.978)
(1.438)
0.914
0.569
0.5303
0.774
0.7468
0.118
Border ≤ 100km {0,1} × CU
-36.354
-68.217
-0.040
2.278*
(50.035)
(86.151)
(0.064)
(1.312)
0.468
0.429
0.530***
0.083
Core {0,1} × EAC
114.786**
20.077
3679.881***
-0.059
10.154
-2.264
(50.173)
(27.673)
(1389.983)
(0.066)
(21.925)
(1.573)
0.022
0.468
0.0085
0.368
0.6434
0.150
Core {0,1} × CU
131.199**
123.391***
0.002
-1.594
(50.987)
(39.155)
(0.057)
(1.520)
0.0102
0.0017
0.9743
0.2946
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.95
0.89
R-Squared -Within
0.10
0.09
0.07
0.15
0.20
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 100km to the border (using road
distances).
292
Table C.74: Robustness Check: Border ≤ 50km (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
Border ≤ 50km {0,1} × EAC
-0.190
0.029
0.047*
0.009
-0.093
0.078
(0.149)
(0.024)
(0.026)
(0.032)
(0.061)
(0.116)
0.201***
0.221**
0.076***
0.786***
0.126***
0.499***
Border ≤ 50km {0,1} × CU
-0.035
0.095***
0.005
0.050
-0.019
0.039
(0.122)
(0.032)
(0.039)
(0.058)
(0.096)
(0.146)
0.774***
0.004
0.906***
0.393***
0.842***
0.786***
Border ≤ 50km {0,1} × CM
-0.177
0.077***
0.050*
0.012
0.091
-0.129
(0.145)
(0.025)
(0.029)
(0.058)
(0.056)
(0.127)
(0.223)
(0.002)
(0.086)
(0.841)
(0.105)
(0.308)
Core {0,1} × EAC
-0.275***
-0.009
-0.185***
0.030
0.015
0.026
(0.080)
(0.048)
(0.051)
(0.065)
(0.078)
(0.090)
0.001
0.849***
0.000
0.642***
0.847***
0.777***
Core {0,1} × CU
-0.457***
0.004
0.734**
0.010
0.059
-0.244**
(0.103)
(0.044)
(0.332)
(0.054)
(0.088)
(0.113)
0.000
0.934***
0.027
0.857***
0.505***
0.030***
Core {0,1} × CM
-0.339***
-0.065
1.092***
-0.259**
0.017
0.005
(0.068)
(0.068)
(0.376)
(0.109)
(0.087)
(0.085)
0.0000
0.3423
0.0037
0.0173
0.8477
0.9497
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.13
0.14
0.27
0.23
0.25
0.09
R-Squared -Within
0.11
0.10
0.25
0.23
0.25
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 50km to the border (using road
distances).
293
Table C.75: Robustness Check: Border ≤ 50km (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
Border ≤ 50km {0,1} × EAC
0.040
-0.119**
-0.055
0.037
-0.086
-0.093***
(0.214)
(0.050)
(0.078)
(0.041)
(0.061)
(0.033)
0.852***
0.017
0.478***
0.359***
0.157
0.004
Border ≤ 50km {0,1} × CU
-0.048
-0.135***
0.038
0.010
-0.104**
-0.145***
(0.154)
(0.034)
(0.064)
(0.028)
(0.047)
(0.043)
0.755***
0.000
0.554***
0.718***
0.026
0.001
Border ≤ 50km {0,1} × CM
-0.092
-0.093***
0.013
0.020
-0.117***
-0.121*
(0.172)
(0.024)
(0.045)
(0.026)
(0.023)
(0.063)
0.592
0.000
0.771***
0.430***
0.000
0.053***
Core {0,1} × EAC
0.452***
0.063
0.163
0.073*
-0.057
-0.011
(0.161)
(0.043)
(0.277)
(0.037)
(0.068)
(0.027)
0.005
0.143***
0.556
0.052***
0.402***
0.678***
Core {0,1} × CU
0.316**
0.022
0.786***
0.019
0.037
0.031
(0.125)
(0.042)
(0.182)
(0.046)
(0.116)
(0.036)
0.011
0.603***
0.000
0.673***
0.752***
0.394***
Core {0,1} × CM
0.345**
-0.001
0.968***
0.064**
-0.006
0.016
(0.164)
(0.043)
(0.115)
(0.032)
(0.093)
(0.032)
0.0349
0.9817
0.0000
0.0495
0.9457
0.6105
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.34
0.16
0.29
0.23
0.26
0.17
R-Squared -Within
0.32
0.11
0.28
0.23
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 50km to the border (using road
distances).
294
Table C.76: Robustness Check: Border ≤ 50km (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[553.78]
[112.23]
[0.12]
[0.01]
[1.20]
[0.00]
Border ≤ 50km {0,1} × EAC
49.969
-2.821
-0.004
0.035***
0.047
-0.143*
(292.688)
(77.036)
(0.035)
(0.009)
(0.051)
(0.078)
0.864
0.9708
0.905
0.0001
0.359
0.068
Border ≤ 50km {0,1} × CU
-13.598
-0.033
-0.022
-0.164*
(306.353)
(0.033)
(0.149)
(0.096)
0.965
0.322***
0.883***
0.087
Core {0,1} × EAC
118.770***
706.221***
0.020
0.032***
0.145***
0.808***
(14.381)
(27.153)
(0.023)
(0.006)
(0.016)
(0.023)
0.000
0.0000
0.386
0.0000
0.000
0.000
Core {0,1} × CU
196.942***
-0.013
0.079
0.799***
(59.110)
(0.025)
(0.063)
(0.117)
0.0009
0.5932
0.2104
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,492
2,695
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.43
R-Squared -Within
0.11
0.09
0.03
0.01
0.14
0.40
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 50km to the border (using road
distances).
295
Table C.77: Robustness Test: Border ≤ 50km (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.89]
[204.38]
[649.97]
[0.26]
[28.83]
[7.81]
Border ≤ 50km {0,1} × EAC
0.159
46.895
1521.906
0.052
13.350
-2.628
(186.972)
(107.532)
(1408.312)
(0.089)
(24.030)
(2.832)
0.999
0.663
0.2807
0.554
0.5787
0.354
Border ≤ 50km {0,1} × CU
-33.556
20.242
0.016
-2.064
(215.623)
(97.817)
(0.058)
(2.874)
0.876
0.836
0.782***
0.473
Core {0,1} × EAC
111.672***
-5.570
4526.138***
-0.069*
2.537
-0.820
(30.388)
(28.461)
(968.869)
(0.038)
(24.289)
(1.357)
0.000
0.845
0.0000
0.069
0.9168
0.545
Core {0,1} × CU
108.651***
78.868**
-0.021
-0.138
(37.305)
(32.295)
(0.037)
(1.365)
0.0036
0.0147
0.5660
0.9194
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
5,494
5,524
2,695
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.09
0.09
0.07
0.15
0.17
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 50km to the border (using road
distances).
296
Table C.78: Robustness Check: Border ≤ 25km (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
Border ≤ 25km {0,1} × EAC
-0.478**
-0.020
0.039
-0.034
-0.176
-0.381
(0.225)
(0.048)
(0.039)
(0.054)
(0.151)
(0.288)
0.034
0.669***
0.320***
0.529***
0.245***
0.186
Border ≤ 25km {0,1} × CU
-0.029
0.150***
-0.002
0.093
0.007
-0.324
(0.297)
(0.050)
(0.049)
(0.066)
(0.156)
(0.269)
0.922***
0.003
0.961***
0.156
0.963***
0.228
Border ≤ 25km {0,1} × CM
-0.215
0.034
-0.019
-0.097
-0.083
-0.380***
(0.255)
(0.043)
(0.055)
(0.079)
(0.107)
(0.146)
(0.400)
(0.434)
(0.726)
(0.217)
(0.437)
(0.009)
Core {0,1} × EAC
-0.275***
-0.011
-0.187***
0.030
0.019
0.010
(0.078)
(0.048)
(0.051)
(0.065)
(0.078)
(0.087)
0.000
0.823***
0.000
0.649***
0.810***
0.911***
Core {0,1} × CU
-0.459***
0.000
0.734**
0.008
0.061
-0.255**
(0.099)
(0.044)
(0.332)
(0.054)
(0.087)
(0.109)
0.000
0.998***
0.027
0.878***
0.487***
0.020***
Core {0,1} × CM
-0.333***
-0.069
1.088***
-0.261**
0.008
0.005
(0.066)
(0.068)
(0.375)
(0.109)
(0.087)
(0.082)
0.0000
0.3125
0.0037
0.0166
0.9229
0.9489
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.13
0.14
0.27
0.23
0.25
0.09
R-Squared -Within
0.11
0.10
0.25
0.23
0.25
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 25km to the border (using road
distances).
297
Table C.79: Robustness Check: Border ≤ 25km (DHS)
Consumpt.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.14]
[0.18]
[0.00]
[0.76]
[1.51]
[0.52]
Border ≤ 25km {0,1} × EAC
0.548***
-0.051***
-0.083
0.090*
-0.010
-0.064
(0.121)
(0.018)
(0.088)
(0.049)
(0.072)
(0.044)
0.000
0.004
0.342***
0.065*
0.891***
0.144
Border ≤ 25km {0,1} × CU
0.261***
-0.068***
-0.004
0.039
-0.025
-0.158***
(0.074)
(0.025)
(0.110)
(0.075)
(0.090)
(0.044)
0.000
0.006
0.970***
0.600***
0.781***
0.000
Border ≤ 25km {0,1} × CM
0.244***
-0.031***
-0.036***
0.066***
-0.063***
-0.088***
(0.167)
(0.026)
(0.092)
(0.043)
(0.062)
(0.038)
Core {0,1} × EAC
0.460***
0.070
0.166
0.072*
-0.051
-0.007
(0.155)
(0.043)
(0.276)
(0.037)
(0.068)
(0.029)
0.003
0.102***
0.548
0.051***
0.451***
0.802***
Core {0,1} × CU
0.324***
0.029
0.784***
0.019
0.044
0.037
(0.122)
(0.042)
(0.182)
(0.046)
(0.115)
(0.037)
0.008
0.493***
0.000
0.673***
0.704***
0.325***
Core {0,1} × CM
0.356***
0.005***
0.967***
0.064***
0.001***
0.023***
(0.158)
(0.043)
(0.115)
(0.032)
(0.092)
(0.034)
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.34
0.16
0.29
0.23
0.26
0.17
R-Squared -Within
0.32
0.11
0.28
0.23
0.26
0.17
Dependent Variable
DHS
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 25km to the border (using road
distances).
298
Table C.80: Robustness Check: Border ≤ 25km (KHDS 1/2)
Agglom.
Annual p.c.
Consumption
(TZS '000)
Value of
dur. Assets
(TZS '000)
Employed
Work
{0,1}
Salaried
Work
{0,1}
Occupation
Level
{1,3}
Population
Density
(sdz.)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[552.31]
[112.51]
[0.12]
[0.01]
[1.20]
[0.00]
Border ≤25km {0,1} × EAC
1046.331***
-242.913
-0.162
0.010
0.188
0.034
(310.431)
(436.378)
(0.122)
(0.030)
(0.210)
(0.066)
0.001
0.5778
0.186
0.7404
0.372
0.608
Border ≤ 25km {0,1} × CU
1126.525***
-0.037
0.583*
-0.080
(354.396)
(0.059)
(0.312)
(0.128)
0.001
0.529***
0.062
0.532
Core {0,1} × EAC
115.804***
705.777***
0.020
0.030***
0.143***
0.816***
(3.903)
(9.572)
(0.022)
(0.006)
(0.015)
(0.028)
0.000
0.0000
0.370
0.0000
0.000
0.000
Core {0,1} × CU
198.590***
-0.012
0.080
0.803***
(31.553)
(0.024)
(0.060)
(0.118)
0.0000
0.6285
0.1800
0.0000
Individual Controls
YES
YES
YES
YES
YES
NO
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
YES
YES
YES
NO
Household Fixed Effects
YES
YES
NO
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
35,813
17,674
24,972
14,254
15,685
2,292
Observations - Fixed Effects
3,816
2,363
12,747
6,988
6,253
12
R-Squared
0.88
0.97
0.64
0.72
0.82
0.43
R-Squared -Within
0.12
0.09
0.03
0.01
0.14
0.40
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 25km to the border (using road
distances).
299
Table C.81: Robustness Check: Border ≤ 25km (KHDS 2/2)
Agglom.
Ann. Food
Consumption
(TZS '000)
Ann. non-Food
Consumption
(TZS '000)
Value of
Dwelling
(TZS '000)
Worked
Last Year
{0,1}
Monthly
Salary
(TZS '000)
Time lived
at Location
(Years)
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[349.53]
[203.27]
[660.74]
[0.26]
[28.83]
[7.81]
Border ≤25km {0,1} × EAC 536.903***
493.731***
3041.987
-0.033
-487.518
-0.755
(166.385)
(181.678)
(3654.443)
(0.139)
(5036999.421)
(2.051)
0.001
0.007
0.4052
0.810
0.9999
0.713
Border ≤ 25km {0,1} × CU
663.445***
452.421***
-0.104
0.380
(233.953)
(150.728)
(0.121)
(2.099)
0.005
0.003
0.391***
0.856
Core {0,1} × EAC
107.507***
-4.390
4497.546***
-0.068*
1.907
-0.734
(17.149)
(14.237)
(325.789)
(0.038)
(23.209)
(1.355)
0.000
0.758
0.0000
0.074
0.9346
0.588
Core {0,1} × CU
109.736***
79.415***
-0.020
-0.044
(21.871)
(14.376)
(0.036)
(1.366)
0.0000
0.0000
0.5804
0.9740
Individual Controls
YES
YES
YES
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Individual Fixed Effects
NO
NO
NO
YES
YES
YES
Household Fixed Effects
YES
YES
YES
NO
NO
NO
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
35,826
36,025
17,673
16,330
1,782
5,411
Observations - Fixed Effects
3,817
3,830
2,363
11,599
1,190
3,763
R-Squared
0.88
0.83
0.69
0.86
0.94
0.89
R-Squared -Within
0.10
0.09
0.07
0.15
0.17
0.09
Dependent Variable
KHDS
Consumpt.
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.3 such that most
of the respective notes apply here. However, note the following important amendments: I replace the continuous
distance variable with a dichotomous dummy switching to 1 for individuals within 25km to the border (using road
distances).
300
Table C.82: Robustness Check: Flexible Distance Specification (Afrobarometer)
Consumpt.
Income
Agglom.
Lived
Poverty
[0,4]
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Cash Inc.
Deprivation
{0,4}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[1.12]
[0.23]
[0.00]
[0.55]
[1.76]
[2.09]
Cont. Distance to Border × EAC (in '00 km)
0.010
0.013
-0.008
0.016
0.035
0.092
(0.066)
(0.014)
(0.017)
(0.018)
(0.029)
(0.068)
Sq. Cont. Distance to Border × EAC (in '00
0.000
-0.002*
-0.001
-0.003*
-0.004*
-0.004
(0.005)
(0.001)
(0.002)
(0.001)
(0.002)
(0.005)
0.980***
0.055***
0.678***
0.066***
0.052***
0.424**
Cont. Distance to Border × CM (in '00 km)
0.006
0.002
-0.015
-0.009
-0.014
0.182**
(0.071)
(0.017)
(0.039)
(0.020)
(0.027)
(0.072)
Sq. Cont. Distance to Border × CU (in '00
-0.004
-0.001
0.002
0.000
0.001
-0.016***
(0.006)
(0.001)
(0.003)
(0.002)
(0.002)
(0.006)
0.547***
0.485***
0.581***
0.933***
0.712***
0.007
Cont. Distance to Border × CM (in '00 km)
-0.004
-0.013
-0.045
-0.001
-0.034
0.165***
(0.061)
(0.010)
(0.055)
(0.019)
(0.022)
(0.062)
Sq. Cont. Distance to Border × CM (in '00
-0.001
0.000
0.004
0.000
0.000
-0.011**
(0.005)
(0.001)
(0.004)
(0.001)
(0.002)
(0.005)
0.902***
0.867***
0.353***
0.960***
0.826***
0.028
Cont. Distance to Core × EAC (in '00 km)
-0.048
0.026
0.084**
0.024
0.056
-0.065
(0.101)
(0.018)
(0.036)
(0.027)
(0.046)
(0.065)
Sq. Cont. Distance to Core × EAC (in '00
0.001
0.000
-0.008**
0.001
0.001
0.004
(0.010)
(0.002)
(0.003)
(0.003)
(0.005)
(0.007)
0.939***
0.836***
0.018
0.672***
0.875***
0.636***
Cont. Distance to Core × CU (in '00 km)
0.126
0.013
-0.205***
0.019
0.015
0.135*
(0.096)
(0.018)
(0.078)
(0.021)
(0.033)
(0.078)
Sq. Cont. Distance to Core × CU (in '00 km)
-0.012
0.000
0.015**
0.000
0.002
-0.011
(0.009)
(0.002)
(0.007)
(0.002)
(0.003)
(0.007)
0.179***
0.859***
0.031
0.913***
0.459***
0.118***
Cont. Distance to Core × CM (in '00 km)
0.031
0.030*
-0.318***
0.069**
0.034
0.005
(0.096)
(0.018)
(0.110)
(0.031)
(0.038)
(0.054)
Sq. Cont. Distance to Core × CM (in '00 km)
-0.006
-0.001
0.023**
-0.003
-0.001
-0.002
(0.008)
(0.001)
(0.009)
(0.002)
(0.003)
(0.005)
0.9390
0.8355
0.0177
0.6725
0.8754
0.6362
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
36,042
25,465
4,156
26,563
21,232
35,975
R-Squared
0.13
0.14
0.20
0.23
0.25
0.09
R-Squared -Within
0.11
0.10
0.17
0.23
0.25
0.09
Dependent Variable
Afrobarometer
Income
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I use the continuous (road)
distances to both border crossings as well as core cities and also add their squared values as a separate regressor
interacted with the treatment time periods.
301
Table C.83: Robustness Test: Flexible Distance Specification (DHS)
Consumpt.
Income
Agglom.
Wealth
Index
{1,5}
Employed
Work
{0,1}
Population
Density
(sdz.)
Worked
last Year
{0,1}
Occupation
Level
{1,3}
Paid in
Cash
{0,1}
(1)
(2)
(3)
(4)
(5)
(6)
Sample Mean of Dep. Var.
[3.11]
[0.18]
[0.00]
[0.78]
[1.51]
[0.52]
Cont. Distance to Border × EAC (in '00 km)
0.008
0.055**
-0.193***
-0.027*
-0.026
0.068*
(0.080)
(0.026)
(0.066)
(0.016)
(0.049)
(0.040)
Sq. Cont. Distance to Border × EAC (in '00
-0.004
-0.011***
0.014***
0.002*
-0.001
-0.011**
(0.006)
(0.003)
(0.005)
(0.001)
(0.005)
(0.005)
0.569***
0.000
0.004
0.079***
0.911***
0.024
Cont. Distance to Border × CM (in '00 km)
0.060
0.015
-0.110**
-0.014
-0.038
0.053**
(0.058)
(0.018)
(0.053)
(0.011)
(0.038)
(0.026)
Sq. Cont. Distance to Border × CU (in '00 km)
-0.006
-0.001
0.009**
0.001
0.007
-0.004
(0.005)
(0.002)
(0.004)
(0.001)
(0.004)
(0.003)
0.286***
0.508***
0.033
0.410***
0.114***
0.209***
Cont. Distance to Border × CM (in '00 km)
0.099*
-0.001
-0.149**
-0.007
0.014
0.050**
(0.054)
(0.011)
(0.059)
(0.011)
(0.021)
(0.023)
Sq. Cont. Distance to Border × CM (in '00
-0.008*
0.000
0.012***
0.001
-0.001
-0.004**
(0.005)
(0.001)
(0.004)
(0.001)
(0.002)
(0.002)
0.099***
0.880***
0.009
0.551***
0.536***
0.014
Cont. Distance to Core × EAC (in '00 km)
-0.015
-0.138***
-0.033
-0.050***
-0.078
-0.063
(0.103)
(0.024)
(0.073)
(0.014)
(0.054)
(0.038)
Sq. Cont. Distance to Core × EAC (in '00 km)
0.003
0.035***
0.004
0.005***
0.027**
0.022**
(0.009)
(0.005)
(0.007)
(0.001)
(0.011)
(0.009)
0.776***
0.000
0.547**
0.001
0.011
0.010
Cont. Distance to Core × CU (in '00 km)
-0.025
-0.003
-0.222**
-0.037*
0.013
0.009
(0.054)
(0.022)
(0.096)
(0.020)
(0.050)
(0.028)
Sq. Cont. Distance to Core × CU (in '00 km)
0.003
0.003
0.018**
0.003*
0.007
0.002
(0.006)
(0.003)
(0.008)
(0.002)
(0.008)
(0.005)
0.555***
0.364***
0.024
0.083
0.402***
0.667***
Cont. Distance to Core × CM (in '00 km)
0.020
0.014
-0.184**
-0.061***
0.061**
0.054***
(0.051)
(0.012)
(0.082)
(0.015)
(0.025)
(0.020)
Sq. Cont. Distance to Core × CM (in '00 km)
0.001
-0.001
0.016**
0.006***
-0.005**
-0.007***
(0.005)
(0.001)
(0.007)
(0.001)
(0.002)
(0.002)
0.8846
0.5534
0.0236
0.0001
0.0395
0.0006
Individual Controls
YES
YES
NO
YES
YES
YES
Geographic Controls
YES
YES
YES
YES
YES
YES
Country-Year Fixed Effects
YES
YES
YES
YES
YES
YES
Observations
183,250
71,738
7,692
169,875
95,717
98,963
R-Squared
0.32
0.15
0.18
0.23
0.24
0.16
R-Squared -Within
0.31
0.10
0.17
0.23
0.24
0.16
Dependent Variable
Income
Afrobarometer
Notes: The results depicted in this table are estimated in the same way as those shown in Table 4.2 such that most
of the respective notes apply here. However, note the following important amendments: I use the continuous (road)
distances to both border crossings as well as core cities and also add their squared values as a separate regressor
interacted with the treatment time periods.
302
303
Chapter 5
Concluding Remarks
This dissertation provided three distinct micro-level analyses of pertinent development
challenges in sub-Saharan Africa, namely education, geography, and trade. Although these
factors have been studied separately throughout the chapters, they contribute jointly to better
explain underdevelopment, particularly in the context of sub-Saharan Africa: Whether it is
geography per se that influences growth via low levels of (agricultural) productivity (Gallup et
al. 1999), the hindrance to trade given its sizeable spatial extent and distance to (world)
markets (Frankel and Romer 1999), or the diminished incentive to invest in institutions or
human capital in face of a peculiar disease environment (Acemoglu et al. 2001; Sachs 2003),
these three factors are interlinked (Rodrik 2002). As such, the thesis speaks to the broader
discussion in the literature on factors related to development while taking a new perspective
by focusing on individual- and household welfare.
Acknowledging the complex interplay between these forces, the dissertation aimed to
contribute to the understanding of these developmental challenges by posing three specific
questions for the continent: (i) How does coastal access explain individual economic well-being
and what are the mechanisms behind? (ii) To what extent does education empower women,
specifically in their choice on fertility outcomes and their freedom to access the labor market?
And (iii) What are the spatial welfare consequences of market integration among three highly
polarized economies?
Chapter 2 concerns itself with the first question and showed that physical geography remains
a significant predictor of individual living standards within a comprehensive set of 28 African
countries. In particular, individuals more distant to major harbors are significantly deprived
across an array of wealth proxies such as consumer durables, the basic consumption items of
food, water, or medical care, and lack access to cash employment. Importantly, these links
remained when controlling for individual-level covariates, country-time specific influences via
fixed-effects, as well as an extensive set of other established geographical influences of
development. This underlines that there are most likely relevant development costs of
remoteness that need to be addressed. Even though the results presented in the chapter cannot
necessarily be interpreted as causal, they are informative and the focus on the individual level
provides new insights in comparison to the existing literature. When exploring potential
304
channels and mechanisms, we found two distinct factors which may provide starting points for
policymakers in trying to rectify these and similar disparities across space. For one, we see that
the presence of basic infrastructural endowments like paved roads, electricity grids, and
healthcare dampens the negative effect across distance. And second, educated individuals fare
significantly better even in these disadvantaged regions, which emphasizes the interlinked
nature of the studied topics once more.
Chapter 3 picks up on education as a key instrument of development policy. In particular,
the chapter studied the impact of a country-wide tuition abolishment policy on women’s
outcomes. What the chapter showed is that while educational attainment was positively
impacted for individuals at all strata, the behavioral (downstream) changes induced by added
schooling are almost exclusively identified for the poorer subgroup of the sample. Poor women
benefit from added schooling in the form of increased literacy, remunerated employment
opportunities as well as lowered desired and realized fertility. None of these effects of added
schooling are documented for women from wealthier households. This confirms the notion that
educating previously neglected groups provides significant marginal returns to schooling. These
findings also provide clear considerations for policymakers, donor agencies, and governments:
Interventions directly at budget-constrained individuals may be more cost-effective and
ultimately more equity-enhancing than one-size-fits-all policies. Given that the external validity
of our empirical strategy is limited, future research needs to investigate whether our findings
can be confirmed in other settings.
Chapter 4 picks up on the recent concerns that specific trade arrangements may be inequality
enhancing and investigated the distributional effect of regional market integration in the East
African Community. Given that the founding members were already highly polarized economies
before trade liberalization, a study of this case was deemed particularly promising. The theory
constructed in the chapter made specific provisions for these spatial facts in an attempt to give
the empirical investigation a prior on expected effects. Following the results of the theory, trade
liberalization in the East African Community was predicted to draw economic activity to the
border and thereby decrease previous spatial inequalities unless strong economies of scale
present in agglomerations outweigh these effects. The empirical evidence drawn from
individual-level data shows a relative welfare increase accompanied with accelerated
agglomeration in the interior economic hubs. Given that these findings are in contrast to the
theoretical predictions and also to what other recent empirical studies have shown, they are
noteworthy and provoke further investigation. While the chapter provides some of these
deliberations in a speculative way, the reduced-form empirics are ultimately not apt to inform
on specific parameter coefficients, which would permit isolating a key factor driving these
observed effects.
305
Additional Caveats. Next to the specific limitations and caveats mentioned within the scope
of the chapters and the preceding remarks, there are some further overarching considerations
that prompt future study on these topics. Despite its distinct benefits, the utilization of
household surveys imposes specific constraints on the analysis. For instance, challenges such as
the temporal limitations and the consistency of data, non-responses and the underreporting of
sensitive topics, as well as practical concerns like the cost of conducting surveys and tracking
long-term changes (in the same) households, prevail. However, given their usefulness, it is
evident that future research must not only rely on complementary data but also make use of
new approaches to data collection in developing contexts, such as high-frequency phone surveys
recently introduced in many countries. Independent of the circumstances, exploring and
combining a wide range of data sources will be crucial for advancing our understanding of the
present and future developmental challenges across the world.
