Policy Research Working Paper 9622

When Distance Drives Destination,
Towns Can Stimulate Development
Joachim De Weerdt
Luc Christiaensen
Ravi Kanbur

Social Protection and Jobs Global Practice
April 2021
Policy Research Working Paper 9622

Abstract
While city migrants see their welfare increase much more trumps the attraction from their promise of greater wealth,
than those moving to towns, many more rural-urban making towns more appealing destinations. Education
migrants end up in towns. This phenomenon, docu- mitigates these effects (lesser deterrence from distance,
mented in detail in Kagera, Tanzania, begs the question greater attraction from wealth), while poverty reduces the
why migrants move to seemingly suboptimal destinations. attraction of wealth, consistent with the notion of urban
Using an 18-year panel of individuals from this region and sorting. With about two-thirds of the rural population in
information on the possible destinations from the census, low-income countries living within two hours from a town,
this study documents, through dyadic regressions and con- these findings underscore the importance of vibrant towns
trolling for individual heterogeneity, how the deterrence for inclusive development.
of further distance to cities (compared to towns) largely

This paper is a product of the Social Protection and Jobs Global Practice. It is part of a larger effort by the World Bank to
provide open access to its research and make a contribution to development policy discussions around the world. Policy
Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted
at lchristiaensen@worldbank.org.

The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.

Produced by the Research Support Team
When Distance Drives Destination,
*
Towns Can Stimulate Development

§
Joachim De Weerdt Luc Christiaensen Ravi Kanbur

JEL Codes: J61, O15, O55.
Key Words: Africa, internal migration, urbanization, secondary towns.

* We thank Kalle Hirvonen for useful comments and Economic Development Initiatives (EDI), the Tanza-
nian National Bureau of Statistics (NBS) and the Organisation for Economic Co-operation and Development
(OECD) for making the underlying data publically available. We acknowledge funding from Excellence of
Science (EOS) Research Project 30784531 at the Research Foundation – Flanders (FWO).
IOB, University of Antwerp & LICOS, KU Leuven. E-mail joachim.deweerdt@uantwerp.be
The World Bank. E-mail lchristiaensen@worldbank.org
§ Cornell University. Email sk145@cornell.edu
1 Introduction

In most of the developing world, wealth is unequally spread geographically and increases
with city size (Ferre et al. 2012; Young, 2013; Gollin et al., 2021). This suggests migration
as an important avenue for growth and poverty reduction, aﬀecting both those who move
(McKenzie et al. 2010, Beegle et al. 2011, Bryan et al. 2014) and those who remain behind
(Giles and Yoo 2007, Kinnan et al. 2018, De Weerdt and Hirvonen 2016, Morten 2019).
Moves to big cities are, on average, much more lucrative than moves to small towns. But
survey data show that few people make it to the cities and many more migrants end up with
more modest income gains realized through moves to smaller towns (Christiaensen et al.,
2019). A key question is therefore how migrants choose their destinations: why do migrants
move to seemingly suboptimal locations?

One consideration is that there are perhaps a number of nonmonetary amenities, such as
health, public goods, crime or pollution that serve to compensate for the lower consumption
levels (Dustmann and Okatenko, 2014). Gollin et al. (2021) argue against this possibility by
showing that in 20 developing countries and across a range of non-monetary amenities the
urban gradient remains clearly visible: practically every amenity they can measure seems
to improve with population density. The authors interpret this as evidence in favor of a
friction model in which people’s mobility is somehow restricted, which allows the spatial
disequilibrium to persist. Support for this view from our study setting in Tanzania comes
from Ingelaere et al. (2018) who show that there is a widespread desire to move out of
the rural areas and into larger urban agglomerations, especially by younger people. But
despite their acute awareness of the spatial diﬀerences in living standards, for the average
young person living in a remote village, moving is also considered a daunting, sometimes
impossible, endeavor.

Much of the migration literature over the past couple of decades has further emphasized the
importance of selection, with the younger, better educated and richer found more likely to
move. They are often better equipped to overcome migration costs and could stand to beneﬁt
more from the employment opportunities for the skilled that urban areas oﬀer (Young, 2013;
Lucas 2015; Hamory, 2020). As those categories are also smaller in number, this could help

1
explain the smaller number eventually making it to the city, compared to the towns.

We use household survey and census data from Tanzania to explore our key question, why
many more migrants move to seemingly suboptimal towns. As in other settings documented
in the literature, the data conﬁrm the existence of rising living standards across the urban
spectrum, from rural areas over towns to cities. That urban gradient is visible both in terms
of assets and housing wealth as well as amenities. Overlaying the wealth distributions, we
ﬁnd that living standards of the 90th percentile in rural areas still lie below those of the 10th
percentile in urban areas.

This is problematic for the basic Todaro (1969) model, which assumes that the rural wage
needs to equal the weighted average of the high and low urban wage, with the weights being
the likelihood of getting each. Kanbur et al. (2019) consider an extension of the basic
model that introduces migration costs.1 These can bring the Todaro model back in line
with the empirical observation in our setting that the rural wage is lower than both the
lower and higher urban wage. Their second extension to the model is to consider multiple
destinations (town or city), each characterized by diﬀerent formal and informal destination
wages (Kanbur et al., 2019).2

Guided by this model, we use a novel data set on migrants for whom we have pre-migration
and post-migration survey data to explore how migration costs and expected income at
destination aﬀect destination choice. At baseline in the early 1990s, the respondents all lived
in Kagera, a large, remote and primarily rural region in the northwestern part of Tanzania.
By 2010 about half of those who were still alive had moved out to other villages, towns
or cities. We use the pre-migration survey data to create a dyadic data set of all possible
migrant-destination pairs, among which the migrant will have chosen one and not chosen all
others. We then use census data to construct measures of the basic Todaro building blocks
(expected wages and migration costs), allowing us to test the relative importance of each.

Particularly, we proxy migration costs with distance. The use of distance as a proxy for
1
Migration costs are broadly deﬁned here, including transport costs but also the risk associated with the
move as well as setup and job search costs at the destination, and so forth. They typically increase with
distance.
2
Similarly, Lagakos et al. (2020) emphasize the importance of heterogeneous migration costs in under-
standing observed migration patterns and returns across destinations (and individual attributes).

2
migration costs (broadly deﬁned) and migration deterrent has a long history in the migra-
tion literature (Ravenstein, 1885; Stewart, 1941; Schwarz, 1973; Lucas, 2001; Lucas 2015).3
Migration networks (such as the presence of a family member at destination) could help
overcome migration costs, even over large distances (Munshi, 2003; Bao et al. 2007; Stuart
and Taylor, 2019), though the presence of a migration network is often itself inversely related
with distance.

The results conﬁrm that destinations that are closer and have higher local living standards
are more likely to be chosen, but distance is, by quite some margin, the most important
determinant. Expected wages would need to go up by 5.7 standard deviations to oﬀset a
1 standard deviation increase in distance. Education increases the attractiveness of higher
living standards at destination. Nonetheless, simulations show that towns still trump cities
in our sample over most of the education gradient. Only after completing upper secondary
education do cities become more attractive. The poor, on the other hand, are even more
deterred by the eﬀects of distance, while being less inclined to respond to the beneﬁts from a
higher (expected) living standard at destination. With the larger share of rural populations
being poorer, less educated and living closer to towns than to large urban centers, it does
not surprise that a smaller than expected share of rural migrants eventually ends up in the
larger, more distant cities.

The paper proceeds as follows. We ﬁrst present the data and characterize the study area
of Kagera and its migrants in more detail (section 2). Section 3 motivates the regression
speciﬁcation and identiﬁcation strategy analyzing the eﬀects of distance and destination
attractiveness on destination choice, as well as how they change given key migrant charac-
teristics. Section 4 presents the empirical results and interprets them with respect to the
observed migration patterns of Kagera’s population (more to towns than to cities). Section
5 concludes.
3
Ravenstein’s ﬁrst law of migration, formulated in 1885 and derived from observing place of birth in the
British 1871 and 1881 censuses, states that most migrants move only a short distance (Lucas, 2001). The
importance of distance for migration has further been generalized into a gravity model of migration, in which
the number of people migrating is inversely proportional to the square of the distance between origin and
destination and proportional to the product of their populations (Stewart, 1941).

3
2 Context, data and the urban gradient

This section introduces our two data sources: the Kagera Health and Development Survey
(KHDS) and the 2002 Tanzania Population and Housing Census. Within each we demon-
strate rising living standards from villages across towns to cities, which we refer to as the
urban gradient. In KHDS the gradient is demonstrated for changes in migrant welfare in a
panel data set, while in the census it is shown to hold cross-sectionally for all residents of
these areas.

2.1 Kagera

The migrants we study all originate from the Kagera region of Tanzania. Figure 1 provides a
map showing where the region is situated within Tanzania. Kagera, being primarily depen-
dent on agriculture and far away from the ocean, is badly connected to the global economy.
It is typical of remote parts of Africa where policy prescriptions for economic growth are
elusive (De Weerdt, 2010). Given small plot sizes (typically 3 or 4 acres per household)
and relatively low-productivity agriculture, internal migration will be an important aspect
of growth and poverty reduction for this region.

In the context of this paper it is useful to highlight some of the geographical speciﬁcities of
the Kagera region, which may be relevant for the study of destination choice and in particular
whether urban migrants end up in large cities or smaller towns. The top panel of Figure 1
shows the baseline villages from the KHDS survey where all our respondents originate (the
survey is described in the next subsection) and the bottom panel shows the urban areas in
the country, split into small towns (10-100k), big towns (100-500k) and cities (>500k).

We see that Kagera lies at the periphery of the country and has, within its borders, only
one large town (Bukoba) and no cities. The closest city is Mwanza on the southern shores
of Lake Victoria, while Tanzania’s prime city, Dar es Salaam, is located at the other end of
the country. In short, Kagera is a relatively remote region and that should be kept in mind
when analyzing urban destination choices of migrants originating from Kagera.

4
2.2 KHDS

The Kagera Health and Development Survey (KHDS) is a study into the long-run wealth
dynamics of households and individuals within Northwest Tanzania. This study entails the
resurvey of a panel of households, originally interviewed for 4 rounds from 1991 to 1994.
Resurveys were then organized in 2004 and 2010. A multi-topic household questionnaire is
administered to all split-oﬀ households originating from the baseline households, including
those that have moved out of the baseline location.

This constitutes one of the longest-running African panel data sets of this nature and oﬀers an
unprecedented set of research opportunities for examining long-run (nearly 20 years) trends
in and mechanisms of poverty persistence and economic growth in rural households. As the
children of the original respondents have now formed their own households, intergenerational
and migration issues can be addressed by the survey data. The data is of particularly high
quality and the 2010 round of the survey was conducted using electronic survey questionnaires
administered on handheld computers, with automated skips and validation checks run during
the interview when errors could be corrected at source. The resulting improvements in data
quality have been formally assessed by Caeyers et al. (2012).

KHDS has maintained a highly successful tracking rate. Table 1 shows that in 2010 88%
of the original 6,353 respondents had either been located and interviewed, or, if deceased,
suﬃcient information regarding the circumstances of their death collected.

In 2010, households were found in three cities: Dar es Salaam, Mwanza and Kampala. This
is deﬁning cities as agglomerations with more than 500,000 inhabitants. The 2012 census
puts the population of Dar es Salaam at 4.36m and Mwanza at 0.7m.4 There are a further 21
respondents who moved to areas that, while administratively recognized within Tanzania as
cities, have a population below 500,000 (Arusha, Tanga and Mbeya). Results do not change
4
For Mwanza this is the sum of Nyamagana Municipal Council (363,452) and Ilemela Municipal Council
(343,001). More detailed analysis using 2002 census data (Wenban-Smith and Ambroz, 2014) showed that
while the municipal districts of Mwanza counted a population of 596,885, only 65% of these lived in strictly
urban wards (others lived in rural or mixed wards). The current census does not yet disclose the new
categorizations, but applying the same rural-urban ratio of 65% to the current census ﬁgure gives us an
urban population of 456,631. This estimate lies under 500,000, but we decided to treat Mwanza as a city
based on the district-level ﬁgures.

5
if we change the deﬁnition from population-based to administrative. All KHDS locations
were matched with their census ward-level classiﬁcation, which distinguish between urban,
mixed and rural areas. All urban and mixed areas are classiﬁed as towns (or cities if in
Mwanza, Dar or Kampala); all rural areas as rural. The KHDS CAPI application had a
series of conditional drop-down menus where the interviewer chose the region, district, ward
and village in which the household was located. These were pre-populated with the exact
census locations and codes, so that the matching exercise is perfect. The census classiﬁcation
is based on 2002 data and the average migrant respondent in our survey moved in 2003.

The consumption data originate from extensive food and non-food consumption modules
in the survey, carefully designed to maintain comparability across survey rounds and con-
trolling for seasonality. The consumption aggregate includes home produced and purchased
food and non-food expenditure. The non-food component includes a range of non-food pur-
chases, as well as utilities, expenditure on clothing/personal items, transfers out, and health
expenditures. Funeral expenses and health expenses prior to the death of an ill person were
excluded. Conservatively, rent is also excluded from the aggregate to avoid large diﬀerences
in prices for similar quality housing being the driver of any measured urban-rural disparities.
The aggregates are temporally and spatially deﬂated using data from the price question-
naires included in each survey round. As household size may diﬀer between urban and rural
households, consumption is expressed in per adult equivalent units rather than per capita.
The poverty line is set at 326,474.2 Tanzanian shillings (TSh),5 calibrated to yield for our
sample of respondents who remained in Kagera the same poverty rate as the 2007 National
Household Budget Survey estimate for rural areas (37.6 percent).

Unless otherwise indicated the analysis below is conducted at the individual level.

Decomposition analysis by Christiaensen et al. (2019), illustrated again in Table 2, shows
that moves to cities are more lucrative for those who make them, but that many more
migrants end up in secondary towns. Furthermore, while there was similarly clear stochastic
dominance between future town migrants and future city migrants by 2010, there was no
stochastic dominance during the early 1990s at baseline (Figure 2), suggesting that town
and city migrants were observationally equivalent (at least by consumption per capita). For
5
At the time of the survey, one US dollar was worth around TSh 1,450.

6
migrants to rural areas the stochastic dominance was already present at baseline and gets
exacerbated by the end line (except for consumption per capita values to the far right of the
distribution, where one would unlikely want to place a poverty line). There is a clear urban
gradient to poverty, just as in Ferre et al. (2012) and Lanjouw and Marra (2018).

2.3 Census

The KHDS survey informs us on the plight of the average Kagera migrant at the diﬀerent
destinations. This section considers the general population characteristics across the whole
of Tanzania. In 2002, about halfway between the baseline and endline KHDS rounds, the
Tanzania Population and Housing Census was conducted. The Integrated Public Use Mi-
crodata Series (IPUMS) provides access to the long-form census data on over 3.7 million
individuals in the country, which can be disaggregated by district, and within each district
by rural and urban areas, giving a total of 254 geographical areas to potentially consider
(Minnesota Population Center, 2014).

The long-form sample contains basic census information, such as the age and sex of all
members, from which we can derive household size and a dependency ratio. Information on
housing characteristics includes the number of bedrooms in the dwelling and the materials
out of which the roof, walls and ﬂoors are made. Further details on utilities allow us to
calculate who has electricity, piped water and a ﬂush toilet. We further know about asset
ownership. For each individual there is information on their employment status and the
industry or sector they work in. The location means for each variable are presented in
Table 3.

Many of the variables collected reﬂect the wealth of the household. Lacking weights, such
as prices, to meaningfully add up the various wealth components available in our data, we
reduce the dimensionality of our wealth data by extracting the ﬁrst principal component of
a number of housing, utility and asset outcomes of sampled households.

We use the following variables for the principal component analysis: number of bedrooms
per person, whether the dwelling has an iron sheet roof (or better), walls from baked bricks

7
(or better), a tile or cement ﬂoor, electric lighting, piped water, a ﬂush toilet and whether
it owns a radio, owns a phone and owns an iron.

Our asset index combines housing and other assets with amenities. Yet, amenities (such
as contentment with local public services and area security) often explain more of people’s
intention to migrate than satisfaction with their personal assets and living standards (Dust-
mann and Okatenko, 2014). To help distinguish between the eﬀect of these two types of
wealth (private/public) on attracting migrants, we also create a separate asset index based
on variables that relate purely to housing and other asset wealth (number of bedrooms per
person, type of roof, wall, or ﬂoor and ownership of radio, iron and phone), as well as an
amenities index based purely on the subset of variables relating to amenities (availability of
electric lighting, piped water and ﬂush toilet).

Housing, utilities and assets outcomes improve as one moves from villages over towns to
cities, in line with what Gollin et al. (2021) document for much of the developing world
(Table 3). This is reﬂected in the constructed wealth index which is much higher in towns
than in rural areas and higher still in cities compared to towns (see also Figure 2). We
also see that, with respect to the constructed wealth index, the 10th percentile in cities
already outperforms the 90th percentile in rural areas. More broadly, the town distribution
of wealth mostly dominates this of rural areas, while the city wealth distribution dominates
both (Figure 3).

Furthermore, urban households are smaller, younger and their members have more years of
formal education. City households are also slightly more educated than town households (0.6
more years of formal education on average – maximum years of education in the household is
8.64 years on average in the city versus 8.01 in towns). The census asks whether the household
experienced any death among its members in the 12 months preceding the interview, followed
by a question on the sex and age of the deceased. Somewhat fewer urban households were
bereaved and in cities the age of the deceased is higher, although the size eﬀects are not
large. Very few rural households have a non-seasonally employed member, but respectively
24% and 38% of town and city households do. Finally, while only 9% of households have
a member working outside of the agricultural sector in rural areas, 54% and 74% do so in
towns and cities.

8
3 Determinants of destination choice

3.1 The extended Todaro model

The previous analysis has highlighted at least two stylized facts, which are not in line with
the basic Todaro (1969) model. First, urban welfare levels clearly dominate rural welfare
levels, across the whole distribution. This is true when we look at data on those who
migrate to these locations in the KHDS, or on all those residing there in the census data.
In the census, for example, Table 3 shows the constructed wealth index in the rural areas
going from -0.99 at the 10th percentile to 0.29 at the 90th percentile, while in cities these
same percentiles span 0.33 to 2.73: even the lowest city welfare level, measured this way,
outperforms that of the highest rural ones. Similarly, KHDS shows that moves to towns or
cities clearly outperform not moving or moves to rural areas.

These facts are problematic for the original Todaro model, as the model requires rural wages
to equate the expected city wages in equilibrium. Kanbur et al. (2019) extend the model
by introducing a cost to moving for each migrant (ﬁnancial, social, psychological,...). The
equilibrium condition now requires that rural wages plus migration costs need to equal the
expected urban wage, so that there is a range of costs that can make the Todaro model ﬁt
the stylized facts.

Second, both KHDS and the census show that the economy is not dualistic, but displays a
clear gradient from villages, over towns to cities. Thus, the second extension the authors
make to the Todaro model is the introduction of multiple destinations.

The model highlights the role of a number of key variables when it comes to destination
choice: the cost of moving and the destination wages. The next section introduces these
variables into a regression framework to assess their relative importance and their interaction
with poverty, education and age.

We use distance to destination as a proxy for the cost of moving. While the use of distance
has been somewhat ignored in the more recent development literature, it has a long tradition

9
in migration (Lucas, 2001). The focus on distance as proxy for migration deterrent in this
study, is further motivated by the work of Ingelaere et al. (2018), who conduct detailed life
history interviews with 75 respondents sampled from the Kagera Health and Development
Survey (KHDS), the same survey underlying the analysis in this paper. They show how
respondents themselves, when narrating their life histories, talked about distance as an
impediment to moving: it raises the cost of moving, it complicates returning home to visit
or when things go wrong at destination, it makes it diﬃcult to maintain ties, can erode
informal property rights and typically comes with higher cultural barriers to moving.

3.2 Econometric speciﬁcation

If we are to understand destination choice among migrants then it is as important to know
where the migrant moved to as it is to know which potential destinations the migrant did
not migrate to. Within the census we can identify 78 locations that were destinations for
KHDS migrants, representing either urban or rural locations in 57 districts. Our analysis
makes the assumption that these 78 locations are the potential destinations for our sample
of migrants. They revealed themselves as practically feasible destinations for at least one of
our migrants.

Each of the migrants in our sample has chosen to move to one potential destination; and
has therefore also chosen not to move to the 77 other potential destinations. In order to
understand that choice better, we create a dyadic data set that contains 78 observations for
each migrant i; one observation for each potential destination d. Our dependent variable
Yid is a dummy equal to one if i was found in location d during the last survey round and
zero otherwise. Our analysis will consist of studying the correlates of Yid . This kind of
dyadic analysis has been used before to understand destination choice by Fafchamps and
Shilpi (2013). We follow these authors in analyzing only those who have actually migrated
in order to not confound the destination choice with the migration choice. We also analyze
separately those who have moved to rural areas and those who have moved to urban areas.

Destination choice Yid will depend on the observed and unobserved characteristics of the
individual (including those related to household and community circumstances), the desti-

10
nation (such as the standard of living the migrant would expect to achieve at destination)
and the relation between the individual and the location (such as distance to destination or
the interaction between the individual’s wealth level and the destination’s wealth level).

Our data setup has 78 observations per individual, which allows us to address observed and
unobserved individual characteristics of the individual by including an individual ﬁxed eﬀect
αi in the regression. Another econometric concern is whether the structure of the error
terms is inﬂuenced by the baseline survey’s two-stage sampling design. We therefore report
standard errors that account for clustering within each of the 51 KHDS villages of origin,
which were the primary sampling units in the baseline survey (Abadie et al., 2017).

The key independent variables of interest include the characteristics of the destination district
in vector Dd and relational variables in vector Rid that are speciﬁc to the i − d pair, such as
distance of d to i’s baseline village and how the eﬀect of destination characteristics changes
with individual characteristics such as education interacted with destination characteristics.
We do not include any individual, household or community level eﬀects. These are controlled
for by the inclusion of the individual ﬁxed eﬀect αi , which helps protect our estimated
coeﬃcients on the destination and relational variables of interest from bias due to unobserved
individual heterogeneity (selection).

We then estimate the following equation:

Yid = Dd β1 + Rid β2 + αi + id , (1)

where id is an error term.

In our most basic regression setup we populate Dd average wealth at destination and Rid
with the natural logarithm of the km distance of the destination district to the baseline
community.

11
4 Results

4.1 Distance is not dead

In the ﬁrst column of Table 4 we estimate Equation 1 on a dyadic data set that links all
migrants to all destinations. Consistent with the insights from the literature, both wealth and
distance are signiﬁcantly correlated to destination choice, with opposite signs: the further
away is the potential destination, the less likely it will be chosen; the wealthier the location
is, the more likely it will be chosen. The (absolute) magnitude of the distance coeﬃcient is,
however, 4 times that of the wealth coeﬃcient. Multiplying those coeﬃcients by the standard
deviations of the corresponding variables in the dyadic data set (0.97 for distance and 0.71
for wealth), we see that average wealth at destination would need to go up by 5.7 standard
deviations to oﬀset the negative eﬀect of a one standard deviation increase in distance.

These results hold controlling for individual characteristics. However, this does not exclude
that the deterring and appealing eﬀects of distance and wealth, respectively, diﬀer by indi-
vidual characteristics. We explore this by interacting distance and wealth at destination with
baseline poverty status (1=poor) (Table 4, column 2), years of formal education (column 3),
and baseline age (column 4). Coming from a poor household increases the negative eﬀect of
distance and decreases the attraction of high wealth at destination. Similarly, education at-
tenuates the friction related to distance and enlarges the pull of high wealth areas, consistent
with high wealth areas oﬀering more opportunities for higher educated people. Interactions
with age show that older people display a slightly lower pull-eﬀect of wealth and are equally
aﬀected by distance as young people.

Our wealth index includes both housing and other asset wealth, as well as amenities. We
investigate whether either of these components dominates destination choice and enter sep-
arate indices for each of these components of wealth as well as their interaction (column 5).
Only the interaction of the two indices is statistically signiﬁcant, suggesting that neither the
assets nor the amenities component dominate the eﬀects of the wealth index, but both are
complementary to each other.

12
Finally, we explore whether distance and wealth have diﬀerent eﬀects depending on whether
the potential destination is urban or rural (column 6). To do so, we include a dummy variable
indicating the urban status of the destination as a regressor, as well as its interactions with
distance and wealth. Urban destinations are substantially less likely to be chosen, but
controlling for whether the destination is urban or rural, we do not see any diﬀerential eﬀect
in the pull eﬀect of the destination’s wealth. The deterring eﬀect of distance also remains,
but is twice as strong for rural than for urban destinations.

We explore these diﬀerences between urban and rural migrants further by altering how we
form the dyads. In Table 5 we link all urban migrants to all 47 urban locations. For urban
migrants selecting urban locations distance and wealth are both signiﬁcantly associated with
destination choice, with coeﬃcients roughly of equal absolute magnitude. At a per standard
deviation change the absolute magnitude for distance (sd=0.91 in this dataset pairing urban
migrants with urban destinations) is 1.6 times greater than that for wealth (sd=0.50). In
Table 6 we link all rural migrants to all 31 rural locations. Again, both distance and wealth
matter, but expressed in standard deviation change (sd=1.03 for distance and a much lower
sd=0.20 for wealth in the rural-rural dyadic dataset) distance outweighs wealth by a factor
of 7.5.

For urban migrants, interactions with poverty and educational status at the outset further
show that the pulling power of urban destination wealth is reduced for poorer households.
Better educated individuals, on the other hand, tend to ﬁnd wealthier urban destinations
more attractive than lesser educated individuals and also tend to be less deterred by distance.
Nonetheless, the deterring eﬀect of distance remains, even for highly educated individuals
(Table 5). It is only by around three years of postsecondary education that distance loses its
eﬀect completely. Urban migrants with such high levels of education would not be deterred
by distance and, furthermore, the pull eﬀect of destination wealth would be 50% higher
for them, compared to those who only completed primary education (the modal level of
education in our sample).

Poor rural migrants, on the other hand, are more deterred by distance than poor urban
migrants, and as their urban counterparts, they are also less inclined to respond to larger
wealth at destination. As with urban migrants, more education reduces the deterring eﬀect

13
of distance for rural migrants (albeit at a lesser gradient) and increases the appeal of wealth
at destination (this time at a steeper gradient). There is also a small negative interaction
eﬀect between age and wealth.

A Shapley decomposition for the regressions in columns 1 of Table 4 shows that distance
accounts for 98% of explained variation of destination choice in the full sample. A similar
exercise for column 1 of Table 5 shows that distance accounts for 75% of the explained
variation once we restrict the analysis to dyads formed by urban migrants choosing between
urban locations. We should bear in mind that the R-squares are quite low in all regressions
(between 4 and 17 percent). This is in large part due to the fact that we are considering a
large number of possible destinations for each person (47 urban and 31 rural destinations).
But it also a reminder that distance does not tell the whole story. Despite the low R-square,
it is obvious from the F-test for joint signiﬁcance of all regressors reported in the table that,
taken together, wealth and distance are important determinants of destination choice.

4.2 Distance, towns and cities

For all three dyadic samples of Table 4, Table 5 and Table 6 the regressions highlight the
dominance of distance over wealth as a correlate of destination choice, both as a share of
explained variation and when comparing the size of standardized regression coeﬃcients. In
this subsection we ask to what extent distance and wealth diﬀerences between cities and
towns explain the observed preference for towns as a destination choice in our study setting,
and how that varies across types of households.6

Our baseline villages lie, on average, 3.06 logged kilometers from the closest town and 5.64
from the closest city, so the diﬀerence is 2.58. Plugging that diﬀerence into the regression
equation estimated in column 1 of Table 5 (urban migrants only) shows, ceteris paribus,
an 8.5 (=0.033*2.58) percentage point higher likelihood of choosing towns because they
are closer. Only 2.6 percentage points of this advantage of towns over cities gets undone
6
As pointed out in subsection 2.2, Kagera is a relatively remote region located far from the nearest cities
(Figure 1). The choice between city or town would play out diﬀerently, obviously, for a migrant living in
close proximity to a big city.

14
through higher destination wealth in cities, which we know, from Table 3, is 0.69 in favor
of cities (0.038*0.69=2.6). The ﬁrst panel (‘ALL’) in Figure 4 visualizes this horse race
between distance and wealth for urban migrants. We can also conduct this exercise using
all migrant-destination pairs in Table 4. Here towns have a 6.5 percentage point higher
likelihood of being chosen (0.025*2.58) and the wealth eﬀect does little in favor of cities
reducing that number by only 0.4 percentage points (0.006*0.69). Clearly, the pull eﬀect of
higher wealth in cities is much lower than the discouragement eﬀect of the corresponding
increase in distance. We conjecture that that is the reason many people prefer towns over
cities.

Who does end up going to the city? Our heterogeneity analysis in Table 5 shows how
wealth at destination has a larger pull eﬀect for the non-poor and the educated and how the
negative eﬀect of distance on destination choice is attenuated by schooling, but exacerbated
by poverty. Our simulations depicted in Figure 4 show that the distance eﬀect becomes very
attenuated for high levels of education. The deterrent eﬀect of distance is over ﬁve times
stronger for those without education compared to those with higher secondary education.
In fact only for this latter group is the distance eﬀect is so attenuated that the wealth eﬀect
(which increases with education, although much more modestly) becomes stronger and cities
become more attractive than towns. As the regression results do not simultaneously interact
with other individual characteristics, we do not conclude from this analysis that raising
education levels would increase migration to cities. Rather the results indicate that there
exists a type of respondent for whom the trade-oﬀ between distance and wealth no longer
favors towns as a destination choice. Such diﬀerential frictions would lead to the kind of
sorting documented by Hamory et al. (2020) and Young (2013). This may result from
education itself as well as other characteristics of the household where the more educated
person grew up, such as wealth levels, access to networks, as well as mindset which may all
help overcome migration costs.

15
5 Concluding remarks

This paper starts from the observation that in our study area of Kagera, Tanzania, much
larger shares of the rural population tend to migrate to towns than to cities, despite greater
welfare gains from moving to the latter than from moving to the former. With many of the
rural migrants originally also poor, town migration contributed more to poverty reduction
than city migration. To shed light on this seemingly suboptimal destination choice of rural-
urban migrants, we draw on an extension of the Todaro model which explicitly considers
diﬀerent destinations (towns and cities) as well as spatial friction/migration costs to motivate
an empirical speciﬁcation to examine the drivers of this equilibrium.

In particular, using a dyadic regression structure applied to our two decade long panel of
individuals from Kagera, augmented with the wealth to be expected at the diﬀerent pos-
sible destinations from the census, we estimate the diﬀerential eﬀect on destination choice
of migration costs and expected wealth gain at destination controlling for individual char-
acteristics (i.e. selection) through individual ﬁxed eﬀects. Migration costs were proxied by
distance to the destination, following a longstanding (though recently somewhat ignored)
tradition in the migration literature (Lucas, 2001). Acknowledging the well-documented
importance of selection in migration, individual heterogeneity in the anticipated deterrent
eﬀect of distance and the expected pull eﬀects of expected wealth at destination were further
explored through interaction with the migrants’ poverty status, education level, and age be-
fore migrating. Simulations were ﬁnally conducted to explore the empirical strengths of the
diﬀerent factors in explaining the migration ﬂows.

The ﬁndings underscore the continuing relevance of distance for destination choice (rural
and urban); in our sample, expected wealth at destination would have to increase by 5.7
standard deviations on average to oﬀset a 1 standard deviation increase in (log) distance. In
this, it is consistent with Ravenstein’s ﬁrst law of migration, put forward in 1885, stating that
most migrants move only a short distance. When conﬁned to urban destinations only, the
deterrent eﬀect of distance remains, but the attraction of wealth increases, with the strength
of these eﬀects in turn aﬀected by the migrant’s poverty (reduced attraction of wealth) and
educational status (reduced deterrence of distance, increased attraction of wealth). With the

16
closest cities on average much further away for most of the rural population than the closest
towns and rural populations often poorer and less educated (Beegle and Christiaensen 2019),
the ﬁnding that more rural-urban migrants end up in towns than in cities, contributing in
the aggregate more to poverty reduction than city migration, does not surprise in light of the
regression ﬁndings. The importance of distance in understanding the decomposition results
is also conﬁrmed by the simulations.

Broadly interpreted, these ﬁndings could be seen as supporting the New Urban Agenda
(United Nations, 2017), which calls for balanced territorial development policies and plans
that strengthen the role of small and intermediate cities and towns, especially since a large
share of the global rural population gravitates around small and intermediate urban centers.
Globally, 28% of the rural population lives within 1 hour travel time from a city of 1 million
or more; 27% within one hour travel time from an intermediate city (250,000-1 million)
and 27% within one hour traveling from a town (250,000 or less) (Cattaneo, Nelson, and
McMenomy, 2021). In low-income countries, the share of the rural population living within
one hour of a town rises to 43%, with another 20% living within 2 hours; 13% live within one
hour from an intermediate city and 7% within one hour from a city (>1 million). Among
the extreme poor, 80 percent are rural (82% in Sub-Saharan Africa) (Beegle et al. 2019).
African cities thus face a double burden to act as engines of poverty reduction; they need
to become economically dense—not merely crowded, to generate the virtuous agglomeration
economies associated with cities in the developed world (Lall, Henderson, and Venables,
2017), and they are also far from where the poor live.

At a minimum, the ﬁndings indicate that the role of distance in migration decisions, high-
lighted in the earlier work on migration, continues to be important empirically, especially in
more remote areas, and that it may have been neglected prematurely more generally. The
paper did not try to uncover why distance plays such an important role. The aforementioned
in-depth life histories with 75 respondents from the KHDS survey (Ingelaere et al. 2018)
point among other things to transport costs and the liquidity constraints to overcome this,
the socio-cultural aﬃnity with the destination, and the ability to maintain ties with the
home community and return home when things go wrong (safety net). Networks can reduce
the deterrence of distance (especially over long distances and including deterrence related
to socio-cultural aﬃnity) (Lucas, 2001; 2015). It explains their importance in determining

17
migration and destination. But their emergence often also depends on distance in the ﬁrst
place. Some distance-related aspects, such as transport costs and information, may be more
amenable to policy, and can reduce friction even over short distance, as illustrated by the ef-
fect of cash transfers on seasonal migration in Bangladesh (Bryan, Chowdhury and Mobarak,
2014). Others may take more time (such as the development of social safety nets). Together
it would suggest that the eﬀect of distance on destination choice is likely to continue for
some time.

Finally, two potential spill-overs following the interplay between distance, towns and mi-
gration may further accelerate the structural transformation and deserve further attention.
First, many new towns are arising in Tanzania, as they are across much of the African con-
tinent. OECD/SWAC (2020) combines demographic data with satellite and aerial imagery
to document a total of 7,617 urban agglomerations of at least 10,000 people in Africa in
2015. In 1950 that number stood at just 624. Growth in the number of urban agglomera-
tions has been especially strong in recent years. From 2000 to 2015, 2,475 new urban areas
came into existence through in-situ urbanization. With every new small town (a subset
of) rural dwellers will see the distance between their home village and the nearest urban
agglomeration shrink. Given the primacy of distance as a determinant of destination choice
documented here, we would expect frictions to rural-urban migration to come down through
the rise of small towns. Such rural-urban migration is an important component of structural
transformation, so the rise of small towns is not only a consequence of this process, but could
actually speed it up. Relatedly, there are good reasons to believe other frictions, for example
related to the ﬂow of goods, services and information between urban and rural areas will
also come down with the emergence of small towns. While important studies like Dorosh
and Thurlow (2013, 2014) and Gibson et al. (2017) have studied how urban growth in large
versus small towns aﬀects rural poverty reduction, we still know very little about the exact
mechanisms through which this happens.

Second, given their proximity, towns can be particularly important in facilitating the ﬁrst
move, which is often the hardest, thereby potentially instigating a virtuous cycle of physical
and economic mobility. The results from the Kagera life histories highlight the diﬃcultly
and importance of the ﬁrst move, an oft ignored phenomenon. It is diﬃcult because it is
often a step in the dark that prospective migrants need to undertake with little preparation

18
and before all the right elements (ﬁnances, skills, networks and the like) are in place to
ensure that the move will be successful. It often comes with many risks and hardships. It is
important because it often sets in motion a virtuous circle of physical and economic mobility.
In this sense we should not just judge the value of towns as points of ﬁnal destinations, but
also as enablers of a ﬁrst move that builds networks, ﬁnancial resources and human capital
potentially (but not automatically) leading to further migration - and in particular, further
migration that would not have been possible without the intermediary step. With towns
being more easily accessible, rural dwellers will become more physically mobile, which will
open opportunities to make life-improving choices. The importance of transit migration has
been highlighted by Artuc and Ozden (2018) in the context of international migration, but
it remains under-documented in low-income contexts as in Africa.

19
References
Abadie, Alberto, Susan Athey, Guido Imbens and Jeﬀrey Wooldridge. 2017. When Should
You Adjust Standard Errors for Clustering? NBER Working Paper No. 24003.
Artuc, Erhan and Caglar Ozden. 2018. Transit Migration: All Roads Lead to America.
Economic Journal 128(July):F306–F334.
Bao, Shuming, Orn B. Bovarsson, Jack, W., Hou, and Yaohui Zhao. 2007. Interprovincial
Migration in China: The Eﬀects of Investment and Migrant Networks. IZA Discussion
Paper 2924.
Beegle, K., J., De Weerdt, and S., Dercon, 2011, Migration and Economic Mobility in
Tanzania: Evidence from a Tracking Survey, Review of Economics and Statistics, 93-3:
1010-1033.
Beegle, Kathleen, Luc Christiaensen. 2019. Accelerating Poverty Reduction in Africa. Wash-
ington, DC: World Bank. https://openknowledge.worldbank.org/handle/10986/32354
Bryan, G., Chowdhury S. and Mobarak A. 2014. Underinvestment in a Proﬁtable Technol-
ogy: The Case of Seasonal Migration in Bangladesh, Econometrica 82(5): 1671–1748.
Caeyers, Bet, Joachim De Weerdt and Neil Chalmers. 2012. Improving Consumption Mea-
surement and Other Survey Data Through CAPI: Evidence from a Randomized Experi-
ment. Journal of Development Economics 98:19–33.
Cattaneo, Andrea, Andrew, Nelson, and Theresa, McMenomy. 2021. Global Mapping of
Urban-Rural Catchment Areas Reveals Unequal Access to Services. PNAS 118(2): 1-8.
Christiaensen, L., J. De Weerdt, and Y. Todo, 2013. Urbanization and Poverty Reduction:
the Role of Rural Diversiﬁcation and Secondary Towns,” Agricultural Economics, 44(4-
5):435-447.
Christiaensen, Luc, Joachim De Weerdt and Ravi Kanbur. 2019. Decomposing the con-
tribution of migration to poverty reduction: methodology and application to Tanzania.
Applied Economics Letters 26(12):978-982. DOI: 10.1080/13504851.2018.1527436
De Weerdt, Joachim. 2010. Moving out of Poverty in Tanzania: Evidence from Kagera.
Journal of Development Studies 46(2): 331-349.
De Weerdt, Joachim and Kalle Hirvonen. 2016. Risk Sharing and Internal Migration.
Economic Development and Cultural Change 65(1):63-86.
Dorosh Paul and James Thurlow. 2013. Agriculture and small towns in Africa. Agricultural
Economics 44(2013):449–459.
Dorosh, Paul, and James, Thurlow, 2014. Can Cities or Towns Drive African Development?
Economy-wide Analysis for Ethiopia and Uganda. World Development 63:11.
Dustmann, C., and A. Okatenko. 2014. “Out-migration, Wealth Constraints, and the
Quality of Local Amenities.” Journal of Development Economics 110: 52–63.
Fafchamps, Marcel and Forhad Shilpi. 2013. Determinants of the Choice of Migration

20
Destination. Oxford Bulletin of Economics and Statistics, 75(3): 388-409.
e, C´
Ferr´ eline, Francisco, H.G., Ferreira, and Peter Lanjouw. 2012. Is There a Metropolitan
Bias? The Relationship between Poverty and City Size in a Selection of Developing
Countries. World Bank Economic Review 26-3: 351-382.
Gibson, John, Gaurav Datt, Rinku Murgai and Martin Ravallion. 2017. For India’s Rural
Poor, Growing Towns Matter More Than Growing Cities. World Development 98(2017):413-
429.
Giles, John and Kyeongwon Yoo. 2007. Precautionary behavior, migrant networks, and
household consumption decisions: An empirical analysis using household panel data from
rural China, Review of Economics and Statistics 89(3):534-551.
Gollin, Douglas, Martina Kirchberger and David Lagakos. 2021. In search of a spatial
equilibrium in the developing world. Journal of Urban Economics 121, 103301.
Hamory, Joan, Marieke Kleemans, Nicholas Li and Edward Miguel. 2020. Reevaluating
agricultural productivity gaps with longitudinal microdata. Journal of the European
Economic Association, jvaa043.
Ingelaere Bert, Luc Christiaensen, Joachim De Weerdt and Ravi Kanbur. 2018. Why Sec-
ondary Towns Can Be Important for Poverty Reduction – A Migrant Perspective. World
Development 105:273-282.
Kanbur, Ravi, Luc Christiaensen and Joachim De Weerdt. 2019. Where to Create Jobs to
Reduce Poverty: Cities or Towns? Journal of Economic Inequality 17(4):543-564.
Kinnan, Cynthia, Shing-Yi Wang and Yongxiang Wang. 2018. Access to Migration for Rural
Households. American Economic Journal: Applied Economics 10(4):79-119.
Lagakos, David, Samuel Marshall, Ahmed Mushﬁq Mobarak, Corey Vernot, and Michael E.
Waugh. 2020. Migration costs and observational returns to migration in the developing
world. Journal of Monetary Economics 113: 138-154.
Lanjouw, P. and Marra M. 2018. Urban poverty across the spectrum of Vietnam’s towns
and cities. World Development 110(2018):295-306.
Lucas, Robert. 2001. The Eﬀects of Proximity and Transportation on Developing Country
Population Migrations. Journal of Economic Geography 1(3): 323-339.
Lucas, Robert. 2015. Internal Migration in Developing Economies: An Overview. KNO-
MAD Working Paper 6: Washington D.C.
McKenzie, David, John Gibson and Steven Stillman. 2010. How Important Is Selection? Ex-
perimental vs. Non-Experimental Measures of the Income Gains from Migration. Journal
of the European Economic Association 8(4):913–945.
Minnesota Population Center. 2014. Integrated Public Use Microdata Series, International:
Version 6.3 [Machine-readable database]. Minneapolis: University of Minnesota.
Morten, Melanie. 2019. Temporary Migration and Endogenous Risk Sharing in Village

21
India. Journal of Political Economy. Forthcoming doi10.1086/700763.
Munshi, Kaivan. 2003. Networks in the Modern Economy: Mexican Migrants in the U. S.
Labor Market. The Quarterly Journal of Economics, 118(2), 549-599.
OECD/SWAC (2020), Africa’s Urbanisation Dynamics 2020: Africapolis, Mapping a New
Urban Geography, West African Studies, OECD Publishing, Paris.
Ravenstein, E.G.. 1885. The Laws of Migration. Journal of the Royal Statistical Society 48:
167-235.
Stewart, John. 1941. An inverse distance variation for certain social inﬂuences. Science 93:
89-90.
Stuart, Bryan and Evan Taylor. 2019. Migration Networks and Location Decisions: Evidence
from U.S. Mass Migration. IZA Discussion Paper No. 12709.
Todaro, Michael. 1969. A Model of Labor Migration and Urban Unemployment in Less
Developed Countries. American Economic Review 59(1):138-148.
United Nations. 2017. New Urban Agenda. Habitat III. https://uploads.habitat3.org/hb3/NUA-
English-With-Index-1.pdf.
Wenban-Smith, H. and Ambroz, A. 2014. Urbanisation in Tanzania. IGC Working Paper
E-40104-TZA-1/2/3, International Growth Centre.
Young, Alwyn. 2013. Inequality, the Urban-Rural Gap and Migration. Quarterly Journal
of Economics 128(4):1727-1785.

22
Tables and Figures

23
Table 1: KHDS response rates

2004 2010
Interviewed 4430 (70%) 4339 (68%)
Deceased 961 (15%) 1275 (20%)
Untraced 962 (15%) 739 (12%)
TOTAL 6353 (100%) 6353 (100%)

24
Figure 1: Tanzania and Kagera

This map shows the location of baseline KHDS enumeration areas, as well as the spatial
distribution of urban agglomerations in Tanzania. Agglomeration and population data are
from Africapolis (OECD/SWAC, 2020).

25
Table 2: Decomposing growth and poverty reduction by 2010 location

Growth (consumption per capita)
Share in
1991-94 Change in total
2010 Location N average 2010 average average growth
Rural 1,086 347,433 573,281 225,848 0.29
Town 702 390,934 906,228 515,293 0.43
City 285 400,836 1,229,495 828,659 0.28
TOTAL 2,073 369,617 776,247 406,630 1.00

Poverty headcount
Share in
total net
Change in poverty
2010 Location N 1991-94 2010 headcount reduction
Rural 1,086 0.56 0.35 -0.21 0.40
Town 702 0.45 0.14 -0.31 0.38
City 285 0.45 0.02 -0.42 0.21
TOTAL 2,073 0.50 0.23 -0.27 1.00

26
Figure 2: Comparison of baseline and endline consumption per capita, by 2010 location
(KHDS)
CDF of baseline KHDS cons pc, by future location CDF of endline cons pc, by current location
1 1

.8 .8
Cumulative Probability

Cumulative Probability
.6 .6

.4 .4

.2 .2

0 0
0 200000 400000 600000 0 500000 1000000 1500000
annualized HH total consumption per capita in 2010 TZS annualized HH total consumption per capita in 2010 TZS

c.d.f. of village c.d.f. of town c.d.f. of village c.d.f. of town
c.d.f. of city c.d.f. of city
The 10% highest cons pc values were dropped for a better rendition of the graph The 10% highest cons pc values were dropped for a better rendition of the graph

Baseline CDF Endline CDF

Figure 3: CDF of census-based wealth index, by location (population-weighted)
1

.8
Cumulative Probability

Rural
.6

Town
.4

City
.2

0
-2 0 2 4
Wealth index

27
Table 3: Census location means

Rural Town City
At least corrugated iron sheet roof 0.33 0.84 0.99
At least baked brick walls 0.18 0.52 0.88
At least tile or cement ﬂoor 0.11 0.62 0.88
Electric lighting 0.01 0.29 0.45
Piped water 0.21 0.68 0.77
Flush Toilet 0.00 0.11 0.16
No. bedrooms 2.29 2.20 2.08
No. bedrooms per person 0.71 0.71 0.72
GoodElec 0.01 0.27 0.42
Owns Radio 0.45 0.67 0.75
Owns Phone 0.01 0.08 0.20
Owns Iron 0.03 0.07 0.09
No. members age 19 or younger 2.68 2.14 1.83
No. members age 20-60 1.82 1.86 2.11
No. members age 61 or older 0.26 0.15 0.10
Household size 4.76 4.15 4.04
Dep. ratio (as share of 20-60 year olds) 1.71 1.27 0.94
Maximum years of formal education in HH 6.70 8.01 8.64
HH experienced death past 12 months 0.06 0.05 0.05
Age died (if deceased past 12 months) 23.20 27.81 27.13
At least one member non-seasonal employee 0.05 0.24 0.38
At least one member works in non-ag sector 0.09 0.54 0.74
Wealth index -0.46 0.79 1.48
Wealth index at 10th percentile -0.99 -0.65 0.33
Wealth index at 20th percentile -0.99 -0.23 0.75
Wealth index at 80th percentile -0.09 1.80 2.34
Wealth index at 90th percentile 0.29 2.65 2.73
Amenities index -0.37 0.73 1.18
Asset index -0.44 0.69 1.38
Notes: Averages are constructed using the census long form survey weights. Only Dar es Salaam and Mwanza are considered
cities. The constructed wealth index is the ﬁrst principal component using the following variables: number of bedrooms per
person, whether the dwelling has an iron sheet roof (or better), walls from baked bricks (or better), a tile or cement ﬂoor,
electric lighting, piped water, a ﬂush toilet and whether it owns a radio, owns a phone and owns an iron.

28
Table 4: Destination choice of migrants - dyadic regressions

(1) (2) (3) (4) (5) (6)
Distance to destination (ln km) -0.025*** -0.023*** -0.039*** -0.023*** -0.027*** -0.035***
(-21.176) (-18.077) (-31.178) (-14.512) (-17.531) (-13.328)
(-”-) * (poor HH) -0.003*
(-1.825)
(-”-) * (yrs schooling) 0.002***
(12.043)
(-”-) * (age in years) -0.0001
(-1.589)
(-”-) * (urban destination) 0.019***
(5.527)
Wealth index 0.006*** 0.009*** -0.006*** 0.009*** 0.022***
(6.062) (8.244) (-5.880) (6.031) (4.390)
(-”-) * (poor HH) -0.006***
(-4.642)
(-”-) * (yrs schooling) 0.002***
(12.496)
(-”-) * (age in years) -0.0001***
(-2.821)
(-”-) * (urban destination) -0.003
(-0.535)
Asset index -0.003
(-0.856)
Amenities index -0.002
(-0.486)
(Asset index)*(Amenities index) 0.020***
(7.231)
Urban destination -0.141***
(-6.455)
R-square .0443 .0449 .0497 .0444 .049 .0561
F 827 427 568 414 590 433
p-value F 0.000 0.000 0.000 0.000 0.000 0.000
N 156,936 155,688 156,936 156,936 156,936 156,936
Notes: LPM estimates of Equation 1 with standard errors clustered by the 51 origin enumeration areas. Regression coeﬃcients
with t-statistics in parentheses. The dyads are formed by linking all 2,012 KHDS migrants to all possible 78 destinations.
Destination wealth is the average household wealth index at destination calculated from the census. The wealth index is the
ﬁrst principal component using the following variables: number of bedrooms per person, whether the dwelling has an iron sheet
roof (or better), walls from baked bricks (or better), a tile or cement ﬂoor, electric lighting, piped water, a ﬂush toilet and
whether it owns a radio, owns a phone and owns an iron. The amenities index is constructed in a similar fashion using a subset
of these variables (electric lighting, piped water and ﬂush toilet) and the asset index is constructed using all the other variables.
Distance is the natural logarithm of distance in kilometers between the baseline location and the potential destination, as the
crow ﬂies.

29
Table 5: Destination choice of urban migrants - interactions

(1) (2) (3) (4)
Distance to destination (ln km) -0.033*** -0.035*** -0.063*** -0.031***
(-12.249) (-10.987) (-15.508) (-5.657)
(-”-) * (poor HH) 0.003
(0.786)
(-”-) * (yrs schooling) 0.004***
(8.516)
(-”-) * (age in years) -0.000
(-0.327)
Wealth index 0.038*** 0.042*** 0.024*** 0.040***
(11.866) (12.680) (3.124) (10.626)
(-”-) * (poor HH) -0.008**
(-2.274)
(-”-) * (yrs schooling) 0.002***
(2.787)
(-”-) * (age in years) -0.000
(-0.383)
R-square .0485 .049 .0547 .0485
F 240 130 294 133
p-value F 0.000 0.000 0.000 0.000
N 45,449 45,120 45,449 45,449
Notes: LPM estimates of Equation 1 with standard errors clustered by the 51 origin enumer-
ation areas. The dyads in the ﬁrst and fourth column are all migrants linked to all possible
destinations. Regression coeﬃcients with t-statistics in parentheses. The dyads in the second
column are all urban migrants linked to all potential urban destinations. The dyads in the third
column are all rural migrants linked to all potential rural destinations. Destination wealth is
the average household wealth index at destination calculated from the census. The wealth index
is the ﬁrst principal component using the following variables: number of bedrooms per person,
whether the dwelling has an iron sheet roof (or better), walls from baked bricks (or better), a
tile or cement ﬂoor, electric lighting, piped water, a ﬂush toilet and whether it owns a radio,
owns a phone and owns an iron. The amenities index is constructed in a similar fashion using
a subset of these variables (electric lighting, piped water and ﬂush toilet) and the asset index
is constructed using all the other variables. Distance is the natural logarithm of distance in
kilometers between the baseline location and the potential destination. Poverty information
comes from the baseline household in which i was residing in 1991-94. Education is measured
as years of formal education at baseline.

30
Table 6: Destination choice of rural migrants - interactions

(1) (2) (3) (4)
Distance to destination (ln km) -0.071*** -0.066*** -0.083*** -0.071***
(-29.180) (-26.434) (-22.623) (-23.494)
(-”-) * (poor HH) -0.007**
(-2.181)
(-”-) * (yrs schooling) 0.002***
(5.180)
(-”-) * (age in years) 0.000
(0.146)
Wealth index 0.049*** 0.065*** 0.017* 0.063***
(6.186) (6.401) (1.961) (6.282)
(-”-) * (poor HH) -0.026**
(-2.214)
(-”-) * (yrs schooling) 0.006***
(4.612)
(-”-) * (age in years) -0.000**
(-2.203)
R-square .172 .172 .174 .172
F 740 380 390 376
p-value F 0.000 0.000 0.000 0.000
N 32,395 32,116 32,395 32,395
Notes: LPM estimates of Equation 1 with standard errors clustered by the 51 origin enumer-
ation areas. Regression coeﬃcients with t-statistics in parentheses. The dyads are all rural
migrants linked to all potential rural destinations. Destination wealth is the average household
wealth index at destination calculated from the census. The wealth index is the ﬁrst principal
component using the following variables: number of bedrooms per person, whether the dwelling
has an iron sheet roof (or better), walls from baked bricks (or better), a tile or cement ﬂoor,
electric lighting, piped water, a ﬂush toilet and whether it owns a radio, owns a phone and owns
an iron. The amenities index is constructed in a similar fashion using a subset of these variables
(electric lighting, piped water and ﬂush toilet) and the asset index is constructed using all the
other variables. Distance is the natural logarithm of distance in kilometers between the baseline
location and the potential destination. Poverty information comes from the baseline household
in which i was residing in 1991-94. Education is measured as years of formal education at
baseline.

31
Figure 4: Likelihood of choosing towns over cities

This ﬁgure shows how much more likely it is for KHDS migrants to choose a town over a
city as an urban destination because of diﬀerences in distance to and wealth at the urban
destination. The y-axis measures this as a predicted likelihood by plugging into the regres-
sions estimates of Table 5 (i) the diﬀerence between average distance to the nearest town
and average distance to the nearest city, and (ii) the diﬀerence between average town wealth
and average city wealth. The solid bars add these two eﬀects together. The ﬁgure shows
that towns are attractive because they are closer and cities because they are wealthier. The
distance eﬀect dominates the wealth, eﬀect explaining the preference for towns, for all except
the most highly educated.