Migration, Human Capital Formation (Brain Drain)

World Development Vol. 40, No. 5, pp.
938955, 2012
2011 Elsevier Ltd. All rights reserved
0305-750X/$ - see front matter
www.elsevier.com/locate/worlddev
doi:10.1016/j.worlddev.2011.11.011
Migration, Human Capital Formation, and Growth:

An Empirical Investigation
CORRADO DI MARIA and EMILIYA A. LAZAROVA *
University of Birmingham, Edgbaston, UK
Summary. We study the eect of skilled emigration on human capital formation and growth in a sample of developing countries. We
nd that the migration rate exerts statistically signicant eects on both the level and the composition of human capital. We are able to
trace the impact of these changes on the growth rate of sending countries via regression analysis and simulations. Our results show that
while there are both winners and losers, almost 70% of the population in our sample suers lower growth as a consequence of skilled
migration. Moreover, the losses are concentrated in countries with low levels of technological sophistication.
2011 Elsevier Ltd. All rights reserved.
Key words education, brain drain, migration, human capital, economic growth, Asia, Africa, South America
1. INTRODUCTION
revised version of the previous paper, Beine et al. (2008) use

the recent data set by Docquier and Marfouk (2006) to test
for the existence of incentive eects in human capital accumulation, that is, the positive eect of migration probabilities
on human capital accumulation. The authors conclude that
these eects are indeed positive and go on to perform counterfactual simulations to compare the ex-post level of human capital in sending countries, when skilled workers do not benet
from a higher migration rate. In this instance their conclusions
are not clear-cut, as more than half the countries in their sample suer from brain drain, rather than benet from a brain
gain. In very recent additions to this literature, Beine, Defoort,
and Docquier (2011a) provide further evidence in favor of the
benecial brain drain hypothesis using a new panel data set
that allows them to explicitly address issues of endogeneity,
while Batista, Lacuesta, and Vicente (2012) test the brain
gain hypothesis for Cape Verde, using specially collected
data. Both papers provide additional evidence in favor of
the existence of substantial incentive eects.
The fact that migration possibilities exert a positive incentive eect conforms to economic intuition and represents an
important result. The literature mentioned above, however,
neglects another, more subtle aspect of the brain gain, in that
it abstracts from the possibility that migration might change
not only the level, but also the composition of human capital.
Indeed, elsewhere in the literature, a number of authors have
emphasized how the possibility of migration encourages
would-be migrants to concentrate on disciplines that are associated with higher probabilities of migration, especially in
health care and ICT (Clemens, 2007; Commander, Chanda,
Kangasniemi, & Winters, 2008; Connell, Zurn, Stilwell,
Awases, & Braichet, 2007; Gibson & McKenzie, 2011;
Kangasniemi, Winters, & Commander, 2007; Lorenzo,
Over the last decades, an increasing number of developed

countries have put in place dierent mechanisms to encourage
the immigration of only the most talented, skilled individuals
from developing countries. 1
As a consequence of such arrangements, the world has witnessed a dramatic modication in the composition of the pool
of migrants moving from developing to developed countries.
Over the last two decades, the share of highly skilled migrants
in the total number of migrants has increased dramatically.
Docquier and Marfouk (2006), for example, estimate that during 19902000 the number of foreign-born workers with tertiary schooling living in OECD member countries increased
by 63.7%, while for unskilled migrants the increase was only
14.4% over the same period. Such accelerating brain drain is
arguably one of the most striking features of globalization.
Whether the ow of skilled migrants from developing to
developed countries is a curse or a blessing for sending countries has been a contentious issue among economists for several decades. 2 One recent strand of literature, started by the
pioneering work of Oded Stark and co-authors (Stark, Helmenstein, & Prskatwetz, 1997, 1998), recognizes that the possibility of migration raises the returns to education and leads
to an increase in the level of human capital that may ultimately
prove benecial for sending countriesa Benecial Brain
Drain or Brain Gain (BG). This inuential literature has
shaped the direction of much of the recent debate on skilled
migration (Beine, Docquier, & Rapoport, 2001; Beine et al.,
2008; Dustmann, Fadlon, & Weiss, 2011; Mountford, 1997;
Stark & Wang, 2002; Vidal, 1998, to name but a few), and
has spawned an empirical literature aimed at testing its theoretical predictions. The earliest contribution in this respect is
the paper by Beine et al. (2001). The authors, however, use
gross migration rates to proxy for the migration rate of skilled
workers. As a consequence, their ndings in support of the BG
hypothesis need to be taken with caution. Beine, Docquier,
and Rapoport (2003) use the data on immigration rates toward the US by level of education published by Carrington
and Detragiache (1998), and also nd empirical support for
the existence of a BG in a cross-section of 50 developing countries. Their regressions, however, show that migration has a
negative growth eect in most developing countries. In the
* We are grateful to seminar participants at the Queens University of

Belfast, the University of Stirling, the National University of Ireland,
Maynooth, as well as conference participants at the European Economic
Association 2010 meeting in Glasgow. We greatly beneted from the comments of four anonymous referees and of the Journal Editor. We would
also like to thank Hillel Rapoport for making the data set for the paper
Beine et al. (2008) available to us for comparison purposes. Final revision
accepted: September 27, 2011.
938
MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH
Galvez-Tan, & Javier, 2007). 3 To date, few theoretical contributions have suggested a connection between the possibility of
migration and the type of skills the agents choose to acquire.
Mariani (2007) discusses the allocation of talent in a rent-seeking framework a` la Murphy, Shleifer, and Vishny (1991). He
concludes that if skills traditionally associated with rent-seeking, such as legal ones, are less conducive to emigration, an increase in the probability to migrate might encourage a shift
toward skills connected to entrepreneurship, like engineering,
ultimately beneting growth. While Mariani (2007) discusses
an alternative channel leading to a benecial brain drain, he
does not allow for the level eect discussed above. In a contribution more focussed on technological change, Di Maria and
Stryszowski (2009) argue that the possibility of migration,
while potentially leading to an increase in the level of human
capital, produces the wrong type of skill composition and
slows down the process of economic development in source
countries. Within a framework inspired by the literature on
appropriate institutions (Gerschenkron, 1962), they observe
that certain skills are relatively more valuable, and hence more
rewarded, in countries closer to the technological frontier.
This is due to the fact that in technologically advanced countries productivity advances are due to innovation, while in less
developed ones imitation plays a major role. By allowing for
the endogenous accumulation of skills on the part of workers,
who base their decision on the relative rewards such skills entail, they show that the possibility of migration distorts the
optimal formation of human capital, and hinders economic
growth. As this eect is stronger the less developed the sending
country, Di Maria and Stryszowski conclude that this provides a potential explanation for the co-existence of winners
and losers among sending countries emphasized by Beine
et al. (2008). From their theoretical analysis, then, it emerges
that neglecting composition eects might generate misleading
results.
In this paper we build on the theoretical insights of Di Maria and Stryszowski (2009) and assess empirically the eect of
migration on the formation of human capitalboth on its level and compositionand on economic growth in a sample of
developing countries. Using a data set covering 130 developing
countries for 1990 and 2000, we nd evidence that the possibility of migration does indeed aect both the level of human
capital, and the type of skills accumulated in sending countries. Furthermore, consistent with the prediction of Di Maria
and Stryszowski (2009), our results show that both eects depend on the level of technological sophistication of the sending
country. 4 We also provide evidence that the pace of technological development (and hence the growth rate of the economy) is aected by the composition rather than the level of
human capital. We conclude that migration, by aecting the
process of skills accumulation has potentially detrimental impacts on growth. To illustrate the substantive implications of
our ndings, we conclude our analysis by simulating the eects
of changes in the skilled migration rate on the composition of
human capital and on productivity growth. Our simulations
allow us to conclude that there are indeed winners and losers
among developing countries. In our data set roughly one third
of countries suer as a result of increased migration; these losers, however, account for over 70% of the total population in
the sample (over 2.7 billion people), and are characterized by a
lower level of technological development.
Besides being closely related to the literature on the brain
drain, our work is also related to the literature studying how
the composition of human capital aects growth. While their
main contribution focuses on the allocation of talent between
entrepreneurship and rent-seeking, the empirical evidence
939
found in Murphy et al. (1991) provides an early example of

linking growth to the skill composition of the work force. Iyigun and Owen (1999) build on these ndings and show that
dierent ratios of entrepreneurs to professionals are optimal
in dierent phases of economic development. The present paper is closer in spirit to the work of Vandenbussche, Aghion,
and Meghir (2006), who explicitly emphasize the interaction
between skill composition and distance from the technological
frontier in determining a countrys growth performance. Their
focus is, however, quite dierent from ours, given that they
discuss secondary vs. tertiary education, do not deal with
international migration, and concentrate their empirical eorts
on OECD countries.
In the next section we describe the conceptual framework
underlying our empirical analysis. The data used in the empirical investigation are described in Section 3, and the results of
the analysis are discussed in Section 4. The substantive eects
of these results for the developing countries in the sample are
the object of Section 5. Finally, Section 6 concludes the paper
with a summary of the results, and some implications for
migration policy.
2. MIGRATION, HUMAN CAPITAL ACCUMULATION
AND GROWTH
Before we proceed to our empirical investigation, we outline
the theoretical framework that underpins that analysis. In
what follows we sketch a model of endogenous growth, where
heterogenous labor inputs are used to produce a nal good
and to generate technological improvements. The model is
an extension of the framework in Di Maria and Stryszowski
(2009). Its structure is presented in Figure 1.
The structure of production is the familiar one of nonscale
Schumpeterian models, where the nal good (Y) is produced
using a continuum of mass M of intermediate products (xi)
and the labor input of lower-skilled workers L, according to 5:
Z M
a
Y t Lt =M 1a
A1a
1
it xit di:
0
where Ait represents the productivity of intermediate i at

time t.
If, following Young (1998), we assume that the number of
intermediates per worker converges to a constant, the production structure above can be shown to be consistent with sustained growth without scale eects.
Producers of intermediate goods maximize their prots in
each period by choosing optimally the levels of production,
and the pace of productivity change. Productivity improves
as the result of both bona-de innovation and imitation activities (Benhabib & Spiegel, 1994; Vandenbussche et al., 2006).
The former pushes the world technological frontier outwards,
whereas the latter consists in copying technologies from the
world technological frontier. Following Di Maria and
Stryszowski (2009), we assume that both activities require
two types of highly-skilled labor inputs, that we refer to as
technically skilled (T) and generalists (G), so that productivity
changes according to:
1r
r
Ait Ait1 Ait1 T /nit G1/
nit At1 Ait T mit Gmit ;
where the subscripts n and m identify skilled labor inputs employed in innovation and imitation activities, respectively, and
A represents the frontiers productivity level. In what follows,
we further assume that workers endowed with general skills
are relatively more productive when employed in creative
940
WORLD DEVELOPMENT
U N SKILLE D W O RKE RS
S KILLE D W O RKE RS
T E C H N ICA L ( T ) AND
G E N ERAL (G)
I M ITATIO N (M)
I N N OVA TIO N (N)
T E C H N O LOGY (A)
P RO D UC TIO N ( Y )
Figure 1. The structure of the model.
endeavors (innovation) rather than in imitative activities (e.g.,

reverse engineering), that is, we let r > /. 6
Given that countries further away from the technological
frontier have a larger advantage of backwardness (Gerschenkron, 1962), entrepreneurs in less developed countries
specialize in imitation; both imitation and innovation take
place at intermediate levels of the technological development;
and nally only innovation occurs once the technological frontier has been reached. On the one hand, this implies that the
growth rate of output declines over time as development progresses; on the other, since general skills become more productive closer to the frontier, the relative wages of generalists tend
to increase as countries develop.
Successive generations (each of mass 1) of heterogeneous
workers are born, accumulate skills, work, and die within each
period. Each worker is born with a dierent level of talent,
which determines the relative opportunity cost (in terms of
working time) of acquiring skills. Thus, comparing the relative
costs and returns to accumulating skills, workers divide themselves into lower-skilled, generalists, and technologists. Ordering workers in terms of talent, we can represent the
endogenous split of workers in the three groups as in Figure 2.
Any increase in the relative wages of skilled workers increases the share of skilled workers in the economy (reduces
L). Moreover, if the relative rewards accruing to G skills increase, so does the share of G-skilled workers.
Given the absence of spillovers, the maximizing behavior of
agents in the decentralized equilibrium described above maximizes GDP growth, and guarantees that each country converges to the frontier of technology over time. 7
Allowing for the possibility of migration changes the results
of the model by distorting the workers skills accumulation
decisions. Since this paper focuses on skilled migration, we
Unskilled workers
Skilled workers
Figure 2. The skill composition of the workforce.
simplify the model by assuming that only skilled workers migrate, while lower-skilled workers are internationally immobile.
When skilled workers are able to emigrate with some positive (exogenous) probability, their expected wages change.
Letting xj,t indicate the expected wage for a worker of type j
at time t, we have:
xj;t 1 rj xHj;t rj xFj;t ;
where the H and F superscripts identify domestic and foreign

wages, respectively, while rj is the probability with which a
worker of type j expects to secure a job abroad. 8 Clearly, emigration is only from less to more developed countries, since it
is driven by productivity dierences. Thus, Eq. (3) implies that
the expected returns to skills increase with the probability of
migration, which means that also the share of skilled workers
increases with it. This is what we call the level eect of migration. In terms of Figure 2, L shrinks as G + T increases. Furthermore, the wage gap shrinks with a source countrys
approach to the technological frontier, thus, the level eect
diminishes with technological development.
In addition to this conventional eect, there is a more subtle
mechanism at work in this framework. As migration is particularly appealingand empirically most relevant 9from less
developed countries to destination ones at (or close to) the
technological frontier, migrants move from countries where
imitation activities are still commonplace to countries where
productivity can be improved only via innovation. This suggests that the relative productivity, and hence the relative
wage, of G-skilled workers is higher in destination countries
than in source ones. Thus G-skills benet from a larger
migration premium relative to T-skills. When migration is
possible and workers consider foreign jobs when deciding on
skills accumulation, G-skills become relatively more attractive
and more workers elect to acquire such skills. We refer to this
as the relative productivity eect.
Eqn. (3), however, claries that the relative attractiveness of
dierent skill types does not only depend on labor remuneration in the destination country, but also on the dierent migration probabilities of workers of dierent type. Indeed, one of
the stylized facts on brain drain is that technically skilled
workers tend to be more successful in emigrating. Thus, if rT is

suciently larger than rG, the possibility of migration might
have the opposite eect from the one previously discussed,
and lead to an increase in the share of technically skilled workers. We call this eect the relative probability eect.
Since both the relative productivity and relative probability
eects have implications for the share of technically skilled
workers, we refer to them collectively as the composition eect.
Given that the two eects operate in opposite directions, the
sign of the composition eect is ambiguous. All else equal,
the relative migration premium of a G-skilled worker is higher
the further from the frontier is his/her country of origin. Thus,
one would conclude that based on the theoretical model discussed above, the relative productivity eect dominates further
away from the frontier (and the overall composition eect is
negative), while the relative probability eect dominates closer
to the frontier (implying a positive composition eect). 10
Finally, it is clear that migration-induced changes in the level and composition of the workforce induced by the probability of migration also aect the growth path of source
countries. Since we have argued above the growth rate is maximized in the market equilibrium without migration, we must
conclude that the probability of migration may reduce the rate
of growth in sending countries.
From this brief discussion we can take away the following
testable implications: (i) the probability of migration aects
positively the proportion of skilled workers in the labor force
(level eect), (ii) the level eect is smaller the closer the country
is to the technological frontier, (iii) the probability of migration aects the proportion of technically skilled individuals
in the labor force (composition eect), (iv) the sign of the composition eect is expected to be negative for countries lagging
far from the frontier, and positive as the frontier is approached. Finally, our theoretical discussion points to the fact
that the growth rate of a country should depend on both the
share of skilled workers (level of human capital) and the distribution of skill types (composition of human capital). Thus, to
gauge whether the distortionary eects of migration enhance
or hamper technological development (and growth), one
should investigate the impact of both the level and the composition of human capital on growth. In the rest of the paper, we
provide empirical tests of the above hypotheses.
3. DATA DESCRIPTION
The data set needed to pursue our empirical strategy is constructed from ve dierent sources. Docquier and Marfouks
(2006) data set provides data on the level of human capital
by educational attainment for 194 countries in 1990 and
2000. 11 Given that skilled migration introduces a wedge between the educational attainment of natives and the amount
of skilled workers available on the domestic labor market,
we need to dene the level of human capital before and after
migration. The ex ante measure, Ha, is dened as the ratio
of working-age nationals with tertiary education (i.e., working-age residents with tertiary education plus the workingage stock of emigrants with tertiary education) to total working-age nationals (that is the sum of working-age residents and
working-age emigrants). The corresponding ex-post variable,
Hp, is instead dened as the proportion of working-age residents with tertiary education divided by the total number of
working-age residents. For our estimations we also use another of the series provided by Docquier and Marfouk
(2006), namely the stock of working-age emigrants from a given source country i to OECD countries with secondary edu-
941
cation (MSsec,i) to instrument for the rate of skilled migration

of that country.
The Docquier and Marfouks (2006) data set also contains
emigration rates to OECD countries by educational level for
194 source countries. These series, however, are constructed
based on census data from destination countries and they
may overestimate the extent of skilled migration, as they fail
to control for where the skills have been acquired. Fortunately
an alternative data source is provided by Beine, Docquier, and
Rapoport (2007) who use additional survey data from a subset of OECD countries on the age of entry, in order to control
for where tertiary education was acquired. According to the
authors, the data on age of entry is available for 77% of skilled
immigrants to the OECD. Therefore, the data provide a reliable indicator of the age-of-entry structure of immigration
to the OECD. 12 Focussing attention only on OECD countries
is clearly a limitation of these data, however, since about 90%
of all high-skilled emigration is toward the OECD, the emigration rates provided by Beine et al. (2007) are a good proxy for
the overall high-skilled emigration rate. 13 In what follows, rh
is the emigration rate of working-age natives with tertiary education to OECD countries, for individuals who rst reached
their destination after the age of 22. The data set contains
information for 168 source countries in 1990, and 192 in 2000.
We exclude from our analysis 29 countries which are considered immigration-receiving. 14 We further remove from the
sample 26 former socialist countries, since human capital formation in these countries in the early 1990s was severely affected by the transition from a centrally-planned to a
market-driven economy. 15 Overall, our sample comprises
130 countries the majority of which are either low (34) or lower middle income (47) countries. 34 countries in our sample are
upper middle income countries, while the remaining 14 are
high income nonOECD countries. 16
The descriptive statistics of the main variables, Ha, Hp, and
rh are presented in Table 1. There is a sucient variation in all
of the variables, in particular the data set presents a wide
range of migration rates. At one end of the range, the country
with the lowest emigration rate of tertiary educated workers in
the sample is Swaziland. Other countries with emigration rates
below 0.005 are the United Arab Emirates, Oman, and Mongolia. At the other extreme, the countries with the largest
skilled emigration rates are Samoa (93% in 1990), Guyana, Palau, and Tonga, all above 80%. Among these, all but Guyana
are island states and they all have relatively small populations.
The case of Guyana is particularly striking since it is one of the
Table 1. Descriptive statistics

Variable
Obs.
Mean
Std. Dev.
Min.
Max.
rh
Ha
Hp
S&T
PROXIM
gTFP,5yr
DENS
EDU
REMIT
PHONE
ROAD
GDPpc
MSsec
POP
262
262
262
129
164
154
387
256
302
299
216
216
262
387
0.19
0.08
0.06
0.37
0.60
0.0044
231.02
4.36
4.04
197.18
36.31
2534.341
48915.92
3.51E+07
0.22
0.06
0.05
0.16
0.11
0.0209
1342.58
2.37
7.17
272.60
28.02
4969.159
167590.30
1.40E+08
0.00
0.00
0.00
0.00
0.37
0.0773
1.34
0.42
0.00
1.27
0.80
129.8224
14.00
15122
0.94
0.28
0.22
0.76
0.84
0.0478
16776.24
15.35
64.87
1849.56
100.00
46605.66
2408250.00
1.30E+09
942
WORLD DEVELOPMENT
countries with the highest ex-ante rate of human capital accumulation (0.21 in 2000), but the post emigration share of
skilled workers in the total is just 4% in 2000. Mongolia, on
the other hand, thanks to the low migration rate emerges as
one of the countries with higher ex-post proportion of tertiary
educated labor force. Several African states like Malawi,
Mozambique, Niger, Rwanda, and Uganda are among the
countries with the lowest level of human capital.
Data on the composition of human capital are taken from
the UNESCO Education Statistics. To proxy for the share
of technically skilled workers we focus on the proportion of
students enrolled in science and technology out of total enrollment in tertiary education, our S&T variable. Given the
restructuring of the International Standard Classication of
Education (ISCED) after 1997, data are only available for
1970, 1980, 1985, 1990, 1995, 1996, and 1997. Due to the great
number of missing values in the series, however, we use values
in 1985 and 1980 to represent the values in 1990 when these
are missing. To match the composition of human capital data
with the series on the level of human capital and emigration
rates, we take the most recent values (1997) to represent
S&T in 2000. Where the 1997 data are missing, we use the
1996 or 1995 values as available. In this way we are able to
construct data on S&T for 83 countries in 1990, and 46 countries in 2000. The average proportion of tertiary students enrolled in science and technology specialities is 37% in our
sample and there is a sucient variation among countries, as
shown in Table 1. Countries with negligible shares of students
enrolled in any science and technology specialities in higher
education are Brunei, Djibuti, Lao, and Seychelles. At the
other end of the spectrum are those countries for which
S&T is above 70% such as Algeria, Angola, Dominica (attains
the maximum in the sample in 1990), El Salvador, Jamaica,
and Trinidad and Tobago.
To control for the degree of technological sophistication in
sending countries, we construct an indicator of proximity to
the world technological frontier, PROXIM, as the ratio
of the total factor productivity (TFP) of country i to that of
the US. Thus, a proximity index close to 1 indicates that a
country is close to the technological frontier, whereas technological laggards are characterized by an index close to 0. As
standard in the literature, we calculate TFP as the log of output per worker, minus the log of capital per worker times the
capitals share. Finding accurate estimates of the capital share
for developing countries is not easy (Gollin, 2001). Following
the recent practice in the literature (e.g., Caselli, 2005; Caselli
& Coleman, 2006) we take the labor shares estimates provided
by Bernanke and Gurkaynak (2002) in Table 10, page 42. 17
Given the limited coverage of developing countries, we can
only collect labor share data for 29 countries in our sample.
For the countries for which data is not available we proceed
as Bernanke and Gurkaynak (2002) and take the labor share
to be 0.65 (i.e., the capital share is 0.35). 18 To construct the
capital stock series we follow Vandenbussche et al. (2006),
and use a perpetual inventory method with a 6% depreciation
rate. As capital investment, we take gross capital formation in
constant 2000 US dollar, Ii,t, from the World Development
Indicators (World Bank, 2009). Thus, the initial level of capital for country i, Ki,0 is given by:
K i;0
I i;1
;
gi 0:06
where Ii,1 is the earliest available data on gross capital formation for country i; and gi is the growth rate of GDP of country
i in the period from the earliest till the latest date of available
data on gross capital formation. 19
As shown in Table 1, the average proximity index in the

sample is approximately 0.6 which we would expect as we have
a sample of developing countries. The variation in the sample,
however, is not very large. As a consequence, in some of our
estimations, we encounter issues of imperfect multicollinearity.
Two major issues that may arise in relation to imperfect mutlicollinearity are (a) the parameter estimates may be sensitive to
changes in sample composition and (b) the standard errors on
the individual coecients may be very high. Thus, we report
the results of robustness checks with respect to the number
of observations and we provide the results from joint hypothesis tests throughout the empirical investigation. The countries
in the sample closest to the technological frontier with an index above 0.8 are the Bahamas, Brunei (it attains the highest
value in the sample in 1990), Costa Rica, Kuwait, Saudi Arabia, United Arab Emirates. Not surprisingly these are all high
income countries, with the exception of Costa Rica (upper
middle income). Among the countries lagging furthest from
the frontier, we have Ecuador (the lowest value in the sample
in 2000) and Malawi both of which have a proximity index below 0.4.
Based on the TFP series, we also construct the dependent
variable used in the growth analysis. For each country, we
compute 5-year averages of TFP growth to smooth out any
cyclical movement in the data. The summary statistics for this
variable can also be found in Table 1. The country with the
lowest average TFP growth in the sample is Jordan and the
country with the highest average TFP growth is Gabon.
The remaining control variables in the regression analysis
come from the World Development Indicators (World Bank,
2009). We use population density (people per squared kilometer), DENS, total public spending on education as a percentage of GDP, EDU, mobile and xed-line telephone
subscribers per employee, PHONE, percentage of paved roads
in the country ROAD, remittances as a percentage of GDP,
REMIT, and initial value for GDP per capita GDPpc. In addition, as an instrument for the skilled migration rate, we use total population, POP. The descriptive statistics of these
variables are also shown in Table 1. The mean values and standard deviations testify once more of the diversity of countries
present in the sample.
4. EMPIRICAL RESULTS
In this section we put the theoretical implications described
in Section 2 to the test. Due to data availability, most of the
analysis will be performed using cross-sectional samples with
a relatively small number of observations. As the data provide
us with only a snapshot of events, we cannot make conclusive
inferences regarding causality. The small sample sizes, furthermore, may imply that some of our results are sensitive to sample composition. The analysis involves estimating three
equations. The rst two refer to the impact of migration on
the formation of human capital, and account for both its level
and composition. The third equation links human capital to
the growth performance of developing countries, specically
allowing for the role played by each countrys level of technological sophistication. As these eects are contemporaneous,
one could argue that there are common shocks aecting all
three regressions, calling for a system estimation approach.
The diculty of conducting our analysis within a system estimation approach arises from data availability. In fact, there
are only 16 countries for which our three general equations
can be estimated simultaneously. 20 Instead, we choose to focus on each component separately and investigate possible
endogeneity issues in depth. We start o with an investigation

of the level eect.
(a) Migration and the level of human capital
In studying the incentive eects of migration on human capital accumulation, the relevant denition of human capital includes not only the residents in the sending country, but also
skilled natives working abroad. For this reason, we use the
ex-ante, measure of human capital, Ha, dened in Section 3.
Our empirical specication extends that of Beine et al.
(2008), to control for the level of technological development
of the source country:
D logH a;0090 a0 a1 logrh;90 a2 logPROXIM 90
a3 logPROXIM 90 logrh;90 a4
logH a;90 a5 DENS 90 a6
logEDU 90 a7 REMIT 90 a8 SSA
a9 LAT e:
The dependent variable, Dlog(Ha,0090) = log(Ha,00)

log(Ha,90), is the growth rate of the ex-ante stock of human
capital during 19902000. We use the stock of human capital
at the beginning of the period, Ha,90, to control for the possibility of convergence across countries in the proportion of tertiary
educated workers. The migration rate of tertiary educated individuals in 1990, rh,90, captures the incentive eects discussed in
Section 2. To control for potential nonlinear impacts of migration at dierent levels of economic development of the source
country, we include log(PROXIM90) and the interaction term
log(PROXIM90) log(Rh,90) in some of our regressions, where
PROXIM90 is the proximity of the country to the world technological frontier. Population density in 1990, DENS90, is used
as a proxy for the cost of acquiring education, since one would
expect that the higher the population density, the smaller the
average distance from schools, the lower the cost of education.
Additionally, we introduce public spending on education in
1990, EDU90, to better proxy for the supply (both in terms of
quantity and quality) of education. 21 REMIT is workers
remittances in 1990, a control for return migration, and alleviated credit constraints on human capital investment. Finally,
SSA and LAT are regional dummies identifying Sub-Saharan
African and Latin American countries, respectively, as dened
by the World Bank. Note that due to the lag structure of the
specication, and data availability, (4) species a cross-sectional regression.
Since the accumulation of human capital and the migration
rate may be simultaneously determineda higher level of human capital may induce higher migration rate due to a reduction in the skill premium on the local labor market compared
to foreign ones, for exampleone possible source of concern
when estimating (4) is the possible endogeneity of the migration rate. Notice, however, that we calculate both the migration rate and the change in the skilled labor force using
immigrants stock, rather than ows, and that the rst dierence in skilled labor stock is calculated over a period of
10 years. As a consequence, endogeneity might not be such a
relevant issue in this model. Nevertheless, we believe it is
important to address potential endogeneity issues using an
instrumental variable approach. Of the various instruments
suggested in the literature (e.g., Durlauf, Johnson, & Temple,
2005) such as the countrys population size, the initial stock of
immigrants, life expectancy at birth, various indices of social
unrest and racial tensions, and the per capita GDP, Beine
et al. (2008) advocate the use of only the rst two, either be-
943
cause the others are very highly correlated with the initial level
of human capital or because of the insucient number of
observations. In our search for valid instruments, we also used
the CEPII data set that contains measures of geographical and
cultural distances between pairs of countries, including the
physical distance of a source country to each one of the six
major destination countries: Australia, Canada, France, Germany, United Kingdom, United States; information on the
colonial history of a source country; and whether a source
country has English, French, or German as one of its ocial
languages. Arguably, all of these variables could be considered
as valid instruments: they are relevant as they aect the cost of
migration like travel costs, cultural proximity, and language
barriers; and they should be exogenous as they should not affect the individual decision to acquire tertiary education. Our
analysis, however, shows that none of these variables passes
the (Stock & Yogo, 2005) test for weak instruments. Moreover, when they are included together with our instruments,
they weaken the signicance of the test statistics. This is why
we rely as instruments for the migration rate of tertiary educated individuals on the stock of immigrants of the same
nationality with secondary education (MSsec) and population
size of the source country (POP). We believe the stock of
immigrants to be a valid instrument for the migration rate because (i) a higher stock of immigrants, that is, a larger diaspora, reduces the cost of emigration, 22 but (ii) the stock of
immigrants with secondary education should not directly affect an individuals decision to acquire tertiary education. Furthermore, we use the stock of immigrants with secondary
education without correction for age of entry. These individuals have not acquired tertiary education in the destination
country post-migration, and thus arguably have even less related to the choice to acquire tertiary education at home. Population size should not be per se a factor in the human capital
equation either, if one accepts that the absence of scale eects
is a realistic feature of the underlying human capital accumulation model. On the other hand, a larger population might reduce the chances to emigrate, since restrictions on immigration
by destination countries do not fully reect the size of the pool
of would-be migrants. In what follows, to test whether the
instruments are relevant, we use the critical values tabulated
by Stock and Yogo (2005), while instrument exogeneity is
tested using a J-test. Finally, we test whether the migration
rate is in fact an exogenous regressor in the model employing
Wooldridges (1995) robust score test.
We use as a benchmark the model studied by Beine et al.
(2008), Model I in Table 2. The rst two columns present
the results from the OLS and IV estimations, respectively. 23
The estimate of the eect of the migration rate on skill formation that we obtain are comparable in magnitude to the results
in Beine et al. (2008): a 1% increase in the migration rate of
high-skilled workers increases the growth rate of the share of
high-skilled workers by about 0.05% points in both specications. We also nd evidence of convergence in human capital
levels among countries in the sample, given that the coecient
for the initial level of human capital has a statistically signicant negative value. According to our estimates, neither public
spending on education (EDU), nor population density
(DENS), nor workers remittances (REMIT) have a statistically signicant eect in either regression. 24
The same holds for the Latin America dummy (LAT). In
fact, a test on the joint signicance of these variables indicates
that they can be excluded from the model as a group (p-value
of 0.6167 in the OLS, and 0.5687 in the IV regression).
A comparison of the OLS and IV results reveals that the two
estimates are quite similar: all regressors have comparable size
944
WORLD DEVELOPMENT
Table 2. Level of human capitaldependent variable Dlog Ha,0090
Model I
log(rh,90)
Model II
Model III
OLS
IV
OLS
IV
OLS
0.0482
(0.0187)**
0.0500
(0.0206)**
0.0481
(0.0202)**
0.2824
(0.1429)*
0.0802
(0.0267)***
0.2870
(0.1425)**
0.2456
(0.0596)***
0.0002
(0.0002)
0.0587
(0.0415)
0.0007
(0.0019)
0.3434
(0.1269)***
0.0122
(0.0460)
0.1883
(0.1489)
0.2454
(0.0563)***
0.0002
(0.0002)
0.0584
(0.0394)
0.0008
(0.0018)
0.3432
(0.1191)***
0.0119
(0.0439)
0.1825
(0.1381)
0.1896
(0.0312)***
0.0006
(0.0002)**
0.0261
(0.0629)
0.0272
(0.0161)*
0.2164
(0.0716)***
0.0210
(0.0669)
0.1674
(0.1753)
0.1574
(0.0350)**
0.0009
(0.0002)***
0.0748
(0.0681)
0.0257
(0.0151)
0.2127
(0.0785)***
0.1178
(0.0692)*
0.5050
(0.2318)**
0.0246
(0.0580)
0.1663
(0.3510)
0.1588
(0.1169)
0.1738
(0.0314)***
0.0006
(0.0002)***
0.0230
(0.0658)
0.0274
(0.0162)*
0.2270
(0.0709)***
0.0407
(0.0632)
0.0241
(0.2194)
82
0.4946
82
55
0.5843
45
log(PROXIM90)
log(rh,90) log(PROXIM90)
log(Ha,90)
DENS90
log(EDU90)
REMIT90
SSA
LAT
Const.
No. of Obs.
R2
Weak
J-test
Endogeneityc
H0: a1 = a3 = 0
346.615a
(0.2896)
(0.7661)
55
0.5947
94.382b
(0.2154)
(0.1555)
(0.0255)**
Notes: Heteroskedasticity robust standard errors are reported in parenthesis.

*
Indicate signicance at 10% levels, respectively.
**
***
a
Robust F-statistic reported.
b
KP Wald statistic.
c
p-Value of the robust score stats reported.
of coecients and the same level of signicance under both

OLS and IV. The instruments pass the test of relevance with
a test statistics of 346, which is much above the critical value
of 10. The test on instrument exogeneity supports our view
that the instruments are exogenous as we do not reject the null
hypothesis at 56%. Moreover, our analysis suggests that
migration rate is not endogenous as we cannot reject the null
that it is exogenous.
Our next specication, Model II in Table 2, extends the
benchmark model to include a measure of the degree of technological development of the sending country. Based on our
theoretical discussion in Section 2, we would expect an increase in the portion of tertiary educated labor force as a
country approaches the technological frontier. This hypothesis
is supported by our regression results as demonstrated by both
the OLS and IV estimations of the model. In the OLS regression, the eect of the migration rate on the proportion of tertiary educated natives is very similar in magnitude to the one
found in the previous model. The IV estimate of the same effect is almost twice as high, 0.08%, which may indicate a
downward bias in the OLS. Also in this case, as instruments
for the migration rate we use the stock of immigrants with secondary education and the countrys population size. In addition, we also instrument for the proximity to the frontier in
order to account for possible reverse causality. As an instrument we use the 10-year lag of that variable, that is, PROXIM80. As an indication of the instruments relevance we use
the Kleibergen and Paap (2006) rank Wald statistics and we
can reject the null hypothesis of weak instruments based on

the Stock and Yogos (2005) critical values at the 10% maximal IV size. The test for instruments exogeneity also indicates
that they are valid instruments. Finally, the test on the endogeneity of the regressors indicates that both the migration rate
and the proximity to the frontier may be treated as exogenous.
The inclusion of the proximity to the frontier as a regressor,
and the change in the samples composition altered the statistical signicance of some of the other explanatory variables.
The coecient of population density (DENS), which is insignicant in the previous regressions, becomes signicant at
the 5% and 1% in the OLS and IV estimations, respectively.
The sign of the coecient, however, is negative. This is contrary to what one would expect assuming that this variable
captures accessibility to higher education. The negative sign,
instead, is consistent with a Malthusian interpretation of population density as a proxy for social conict and environmental pressure (see Malthus, 1798, and more recently Urdal,
2005). In addition, the eect of remittances becomes positive
which complies with the intuition that higher remittances play
a role in relieving credit constraints, allowing more individuals
to pursue higher education. This variable is only statistically
signicant in the OLS regression, and only at 10%.
To see whether the probability of migration aects the level
of human capital dierently at dierent stages of development,
we include an interaction term between these two variables in
our last specication, Model III in Table 2. 25 None of the
three main variables of interestmigration rate, proximity
to the technological frontier, or their interactionis shown to

have a statistically signicant eect. This may be explained by
the high correlation between the interaction term and the log
of migration rate, 0.7463, and that between the interaction
term and the log of proximity to the technological frontier,
0.6004. 26 Indeed, a test for the joint signicance of all three
coecients, indicates that they are signicant at 1% (p-value of
0.0059). A joint hypothesis test on the signicance of the (nonlinear) eect of migration (i.e., H0: a1 = a3 = 0) indicates that
this eect is signicant at 5%.
Based on the signs of the estimated coecients for the
migration rate and the interaction term, we nd some support
for the theory described in Section 2: the positive eect of
migration on human capital accumulation is stronger the further the countries are from the technological frontier. 27 This
eect fades as the frontier is approached, and eventually vanishes once the frontier is reached. 28
It is worth noting that, compared to the OLS estimation of
Model II, the coecient estimates for the other variables are
robust to the inclusion of the interaction term.
(b) Migration and the composition of human capital
The possibility that certain skills may be more demanded
and more rewardedin destination countries relatively to the
home market, adds another layer of complexity to the analysis
of human capital formation under migration. Our aim of this
section is to put to the test the theoretical prediction that the
possibility for migration distorts the composition of human
capital in the source country, and that this distortionary eect
depends on its distance from the technological frontier.
Ideally, we would like to use data on migration rates of tertiary educated natives by eld of study. Such data are unfortunately unavailable, and we can only use migration rates by
educational level. We thus have to assume that all tertiary educated workers face the same migration rate. 29 This implies
that the ex-ante and ex-post skill composition of human capital is the same, and simplies our notation as we do not need
to introduce subscripts to distinguish between gross and net
variables. Furthermore, due to lack of data we use the proportion of enrollment in tertiary education with scientic and technical major, S&T, as a proxy for the proportion of science and
technology graduates in the stock of skilled workers. While
this can be seen as a limitation for our growth regressions below, enrollment is indeed the best point of analysis to identify
the incentive eects of migration on the composition of human
capital, since enrollment reacts to work prospects much faster
than the stock of university graduates.
We model the empirical relation linking the composition of
human capital (before migration) to the migration rate of
skilled workers as
S&T t b0 b1 rh;t b2 PROXIM t b3 PROXIM t rh;t
b4 Qt b5 EDU t b6 D2000 b7 SSA b8 LAT
tt ;
where the variables S&Tt, rt, PROXIMt, EDUt, SSA, and

LAT have all been dened above and are measured at time
t 2 {1990, 2000}. The public expenditure on education is included in this model to control for potential distortionary effect of government involvement in education, the
presumption being that higher public expenditure in education
is associated with more centrally planned education sectors,
and higher distortions. Qt here indicates one of dierent proxies for the domestic demand for technically skilled individuals.
We use two alternative variables to check the robustness of
945
our results: phones per employee and the percentage of paved

roads. 30 The additional variable D2000 is a dummy variable
for year 2000.
We start o by estimating a baseline specication in which
we abstract from possible nonlinear eects of migration, that
is, we assume b3 = 0 in both Models I and II of Table 3. Models I and II dier for the proxy used to control for domestic
demand for technical skills. As in Section 4(a), we present both
OLS and IV results for each specication. In the IV estimations we correct for the possible endogeneity of the migration
rate and the proximity to the technological frontier for reasons
similar to the one discussed in the context of human capital
accumulation above. We use the same set of instruments as
in the IV estimates in Section 4(a): the stock of immigrants
with secondary education, the population size, and the 10-year
lag of the proximity to the technological frontier. When we use
phones per employee as our proxy, the performance of the
instruments is the most satisfactory, both in terms of their relevance and exogeneity. When the percentage of paved roads is
used as a proxy, the instruments pass the test for relevance at
15% Maximal IV Size 31 and pass the test for exogeneity at the
10% level but not at the 5%. 32 Further investigation of the
endogeneity of migration rate and proximity to the frontier
indicates that they can be treated as exogenous. The eect of
skilled migration on the proportion of students enrolled in science and technology specialties appears robust in the two specications: a percentage point increase in the skilled migration
rate, leads to an increase in the proportion of higher education
students enrolled in science and technology degrees of about
0.19% points for all specications in the rst four columns
of Table 3. All other coecients, except for the one of proximity, are rather stable across OLS and IV estimates. For the
coecient on proximity, the OLS and IV regressions produce
dierent signs when phones per employee is used to proxy for
the demand of technicians. 33 Going back to our theory, one
would expect to see a negative sign on this variable as generalists become relatively more productive, the higher the degree
of technological sophistication. The demand proxies have the
theoretically predicted positive sign, while the negative sign of
the coecient for public expenditure in education is rather
puzzling. This may indicate that a tertiary education sector
characterized by a high involvement of the government faces
a cap on the enrollment in technical degrees.
Next, we investigate the evidence on relative probability and
relative productivity eects identied in Section 2. In the presence of such eects, we would expect to nd a statistically signicant estimate for b3. Indeed, we nd evidence that the
interaction term is statistically dierent from zero in both
Models III and IV in Table 3. Furthermore, the results indicate that for countries further away from the technological
frontier the eect of skilled migration on the proportion of students enrolled in technical degrees is negative, while for those
closer to the frontier the eect is positive. The estimated
threshold values for the reversal in sign are 0.5022 and
0.5336 in Models III and IV, respectively. 34 In both cases
the threshold is statistically smaller than 1 (p-values for H0:
b1 = b3 equal 0.0000 and 0.0012 in Models III and IV,
respectively). 35 Thus, in countries with relatively low levels
of technological sophistication the possibility of migration reduces the enrollment in science and technology specialties,
compared to a situation in which no emigration is allowed.
The opposite occurs in relatively more developed countries.
The result complies with our intuition: the relative productivity eect, which predicts a negative eect of migration on the
proportion of science in technologies, dominates the relative
probability eect, which predicts a positive eect of migration
946
WORLD DEVELOPMENT
Table 3. Composition of human capitaldependent variable: S&Tt
Model I
rh,t
PROXIMt
rh,t PROXIMt
log(ROADt)
Model III
Model IV
OLS
IV
OLS
Model II
IV
OLS
OLS
0.2060
(0.0933)**
0.4159
(0.2202)*
0.1924
(0.1113)*
0.2802
(0.2167)
0.1874
(0.0980)*
0.0032
(0.2033)
0.2016
(0.1130)*
0.0440
(0.1995)
1.1402
(0.3838)***
0.3017
(0.2070)
2.1367 (0.5793)***
0.0418
(0.0275)
0.0271
(0.0081)
0.8262
(0.4501)*
0.6191
(0.2226)***
1.6452 (0.6721)**
0.0262
(0.0289)
0.0339
(0.0204)
0.0256
(0.0082)***
0.0644
(0.0445)
0.0592
(0.0379)
0.0014
(0.0409)
0.3890
(0.1123)***
0.0282
(0.0257)
0.0292
(0.0075)***
0.0400
(0.0492)
0.0536
(0.0397)
0.0027
(0.0386)
0.3889
(0.1192)
82
0.2952
73
log(PHONEt)
EDUt
D2000
SSA
LAT
Const.
No. of Obs.
R2
Weaka
J-testb
Endogeneityb
0.0215
(0.0084)**
0.0051
(0.0341)
0.0232
(0.0467)
0.1192
(0.0464)**
0.5244
(0.1252)***
0.0271
(0.0081)***
0.0045
(0.0328)
0.0357
(0.0435)
0.0796
(0.0464)*
0.5408
(0.1219)
68
0.3293
60
13.293
(0.0851)*
(0.9463)
0.0229
(0.0084)***
0.0003
(0.0329)
0.0424
(0.0492)
0.0877
(0.0546)
0.7256
(0.1496)***
0.0381
(0.0202)*
0.0274
(0.0074)***
0.0749
(0.0419)*
0.0596
(0.0368)
0.0120
(0.0398)
0.5718
(0.1205)***
68
0.3734
82
0.3740
15.506
(0.4791)
(0.8882)

*
**
***
a
Sheas partial R2 reported for each endogeneous variable.
b
p-Value of the robust score stats reported.
on S&T, for countries lagging further behind in terms of technological development. It is easy to imagine that science and
technology skills, that are more prone to obsolescence than
other types of skills, may be deemed less marketable if acquired in countries that lag far away from the frontier. As a
consequence, students in the least developed countries could
be more inclined to acquire other skills, studying arts and
humanities, for example, in view of migration. 36 At the same
time, emigration is generally believed to be easier for science
and technology graduates (provided of course that their skills
are current), explaining our results for the relatively more
developed countries in our sample. 37
Proximity to the technological frontier is shown to also affect the composition of enrollment in tertiary education directly in both Model III and IV. Indeed, the p-value of the
test for the joint signicance of b2 and b3 is 0.0019 in Model
IV. The direction of the overall eect depends on the migration rate: for countries with a migration rate higher than a certain threshold of migration, approaching the frontier has a
positive eect on the share of enrollment in science and technology degrees; for countries with a migration rate below
the threshold, proximity to the frontier has the opposite eect.
The value of the estimated threshold is not very robust across
specications: it is estimated at 35% according to Model III
and 14% according to Model IV, this latter value, moreover,
is found not to be statistically dierent from 0 at 10%. Despite
the empirical uncertainty as refers to the value of the threshold, these results are consistent with our theoretical framework. In particular, they conform with an interpretation
focussed on two mechanisms. First, as countries move closer

to the frontier, workers incentives to acquire technical skills
may decrease if complementarities across skills become more
pronounced, and a broader variety of skills are demanded. It
has been suggested that this is indeed what happens when
countries complete the transition from being mere imitators
of foreign technologies, a phase in which a high share of technically inclined graduates is preferable, to being properly innovative, when a wider range of skills become necessary
(Feinstein, 2006; Yusuf, 2007). This view implies that the share
of technically skilled labor force should decline as proximity
increases. Second, as migration becomes easier, a larger proportion of workers actively seek jobs abroad. Since most destination countries favor technically skilled immigrants,
investing in S&T skills equips potential migrants with competencies that are in high demand overseas (provided that the
country is advanced enough). This latter mechanism is more
prominent in high emigration countries. The interplay between
these two mechanisms explains our results above: at low levels
of migration the former eect dominates, while for higher levels of migration it is the latter eect that has the stronger impact.
(c) Human capital and growth
The last step in our empirical endeavor is to gauge the signicance that changes in human capital formation due to
migration may have on economic growth. To this end, we
study the eect of both the level and composition of human
capital on the growth rate of the countrys TFP. Here we build

upon the empirical growth model of Vandenbussche et al.
(2006) who study the eect of human capital on growth as a
function of the countrys technological development. While
the aim of Vandenbussche et al. (2006) is rather dierent from
ours, their model is well suited to study the hypothesis that different skill compositions of human capital are better for
growth at dierent stages of development.
The empirical relation between the level and composition of
human capital and TFP growth is given by the following equation:
gTFP ;5yr;t c0 c1 logH p;t c2 PROXIM t logH p;t c3
log S&T t10 c4 PROXIM t logS&T t10
c5 PROXIM t c6 GDP pc;t10 c7 SSA c8 LAT
6
mt :
The dependent variable, gTFP,5yr,t is the average annual growth

rate of TFP over 5 years. 38 Hp,t is the level of human capital
after migration. S&T, the proportion of enrollment in tertiary
education with technical and science specialty, is used here to
proxy for the proportion of the stock of workers in the labor
market with the same characteristics. Changes in enrollment
only gradually manifest themselves as corresponding changes
in the stock of human capital, for this reason in this equation
we use a 10-year-lag for this variable. To capture possible nonlinear eects of the level of human capital and its skill composition, we include interaction terms of both variables with the
index of technological development, PROXIM. The initial value of GDP per capita, GDPpc,t10, is included to gauge the
existence of catching-up eects, and therefore we expect it to
have a negative coecient. Finally, SSA and LAT are the
country group dummies described above. Given the lag structure of the model, almost all specications are based on crosssectional data. The only exception is the specication in which
the eect of S&T is assumed to be zero.
947
The set of relevant regression results are presented in Table 4.

From our point of view the main insight is that the eect of the
composition of human capital is statistically signicant in all
specications. Models IIIV allow for possible nonlinear eects
of the proportion of workers with technical education.
Although the individual coecients of S&T and the interaction
term are not statistically signicant in any of these specications, joint signicance tests indicate that the null hypothesis
can be rejected at 5% in Model III and at 10% in Models IV
and V. This is easily explained by the high correlation (0.8501)
between S&T and its interaction term in the sample. All these
regressions indicate that a higher proportion of workers with
technical skills are benecial for TFP growth, but that this eect
is decreasing with the proximity to the technological frontier. 39
These results conrm our theoretical priors derived by the work
of Di Maria and Stryszowski (2009).
While we nd statistically signicant eect of the type of
composition of the labor force on TFP growth, we fail to identify any statistically signicant eect of the proportion of tertiary educated workers in any of the regression specications.
When we include an interaction term between the level of human capital and the proximity to the frontierin Models II
and IIIour estimates show that the eect of the proportion
of tertiary educated people on TFP growth is weaker, the
higher the degree of technological sophistication. This eect
goes counter the evidence provided by Vandenbussche et al.
(2006) for OECD countries.
For completeness, the second column of Model I reports the
results of the IV estimation of the basic regression model without nonlinear eects in which we control for the endogeneity
of the proximity to the technological frontier. We follow the
estimation strategy of Vandenbussche et al. (2006) and we
use a lagged value of proximity to instrument for PROXIM00.
In particular, we use PROXIM80. We cannot test for the exogeneity of this instrument as the system is exactly identied.
We argue, however, that this is an acceptable assumption, given the length of the lag. The instrument is highly relevant, the
Table 4. Total factor productivity growthdependent variable: gTFP,5yr,t

Model I
log(Hp,00)
log(Hp,00) PROXIM00
log(S&T90)
log(S&T90) PROXIM00
PROXIM00
GDPpc,t10
SSA
LAT
Const.
No. of Obs.
R2
H0: c1 = c2 = 0
H0: c3 = c4 = 0
Model II
Model III
OLS
OLS
OLS
0.0050
(0.0538)
0.0339
(0.0262)
0.0365 (0.0443)
0.0201
(0.0443)
0.0074
(0.0022)***
0.0066
(0.0066)
0.0010
(0.0044)
0.0493
(0.0354)
0.0560
(0.1814)
0.1002 (0.3047)
0.0347
(0.0262)
0.0355 (0.0433)
0.0306
(0.0757)
0.0075
(0.0024)***
0.0070
(0.0079)
0.0012
(0.0052)
0.0434
(0.0459)
OLS
IV
OLS
0.0027
(0.0516)
0.0015
(0.249)
0.0916
(0.1650)
0.1823 (0.2654)
0.0138
(0.058)**
0.0161
(0.0062)**
0.0568
(0.0223)**
0.0074
(0.0024)***
0.0066
(0.0072)
0.0005
(0.0050)
0.0275
(0.0152)*
0.0424
(0.0249)*
0.0058
(0.0038)
0.0058
(0.0072)
0.0016
(0.0059)
0.0288
(0.0226)
0.0715
(0.0564)
0.0064
(0.0022)***
0.0002
(0.0061)
0.0005
(0.0047)
0.0018
(0.0265)
56
0.2768
50
71
0.1276
(0.7263)

*
**
***
56
0.2850
(0.0453)**
Model IV
56
0.2864
(0.9473)
(0.0699)*
Model V
0.0343
(0.0258)
0.1047 (0.0435)
0.0188
(0.0552)
0.0073
(0.0024)***
0.0063
(0.0051)
0.0012
(0.0051)
0.0496
(0.0352)
56
0.2851
(0.0524)*
948
WORLD DEVELOPMENT
F-statistic is 216.91, which is much higher than 10, the critical

value for the Stock and Yogos (2005) weak instrument test.
The robust Wald test for the endogeneity of proximity indicates that we can reject the null hypothesis that PROXIM00
is exogenous in this model at 5% but we cannot reject it at
10% (p-value 0.0634). Given the remarkable stability of the
coecients on the level and composition of human capital,
and the indicator of proximity to the technological frontier between the OLS and IV estimates, however, we believe that the
OLS estimates are suciently reliable.
Finally, we want to remark that the coecient of the initial
value of GDP per capita has the statistically signicant negative sign, as expected. The regional dummy variables, on the
other hand, are not signicant in any of the specications.
5. MIGRATION, HUMAN CAPITAL AND GROWTH:
A DISCUSSION
So far we have provided empirical evidence that the possibility to emigrate aects the process of human capital formation
by changing the incentives to accumulate human capital and
to acquire certain types of skills. We have found that both
of these distortionary eects are statistically signicant and
comply with the underlying theoretical predictions. To gauge
the economic signicance of these eects, we have also investigated the role of human capital in the productivity growth
of a country. In this sample of developing countries, we nd
robust and statistically signicant eects of the composition
of tertiary education labor force on growth, but no such eect
for the proportion of tertiary educated labor force. This is why
in our subsequent discussion we focus solely on the composition of human capital. Our aim is to derive a better understanding of the substantive eects of the possibility of
emigration on growth operating through this channel. To this
end, we adopt a simulation-based approach, as discussed by
King, Tomz, and Wittenberg (2000) and employ the statistical
software CLARIFY developed by these authors.
Our rst step is to study the eect of the migration rate on
the proportion of individuals enrolling in a science and technology degree for the set of 82 observations for which we
could estimate Model IV of Table 3. We simulate an increase
in the rate of skilled migration in each country in the sample
by 40% (e.g., a country with an initial migration rate of 0.19
would face an increase to 0.19 1.4 = 0.27.) while keeping
the level of technological development, the percentage of public expenditure on education, and the number of phones per
employee at their actual levels. 40 Our statistical exercise uses
stochastic simulations techniques to simulate the dierence
in the share of workers with S&T major, by drawing 1000 sets
of simulated parameters for each country, from the sampling
distribution of the parameters estimates. The condence interval for each observation is obtained by ordering the simulated
values and considering the 5th and 95th percentile. The results
of these simulations are presented in Table 5. 41
Clearly, an increase in the emigration rate by 40% represents
a small change in percentage point terms for countries with
low emigration ratesit is less than 1% point for Brazil, for
examplewhile for countries with high emigration rates, this
is a large change in absolute termsit is equivalent to 19%
points for Barbados, for example. Table 5 illustrates that on
average, a country like Argentina could have seen its share
Table 5. The eect of an increase of the migration rate by 40% on S&T

Country
Year
Drh
DS&T
90% conf. Interval
Trinidad and Tobago

Trinidad and Tobago
Barbados
Sri Lanka
Panama
Mauritius
Mauritius
Malaysia
Domenican Republic
Iran
Honduras
Costa Rica
Turkey
Egypt
El Salvador
Tunisia
El Salvador
Iran
Colombia
Colombia
Tunisia
Jordan
Nicaragua
Uruguay
Nicaragua
Argentina
Uruguay
Chile
Chile
Kenya
2000
1990
2000
1990
1990
1990
2000
1990
2000
1990
1990
1990
1990
1990
2000
1990
1990
2000
1990
2000
2000
1990
2000
2000
1990
1990
1990
1990
2000
1990
0.2699
0.2646
0.1900
0.0924
0.0498
0.2172
0.1805
0.0642
0.0512
0.0763
0.0560
0.0189
0.0278
0.0182
0.0730
0.0518
0.0836
0.0416
0.0253
0.0296
0.0339
0.0210
0.0776
0.0214
0.0800
0.0108
0.0184
0.0191
0.0170
0.1520
0.1434
0.1419
0.0872
0.0297
0.0284
0.0251
0.0243
0.0235
0.0154
0.0148
0.0115
0.0112
0.0111
0.0096
0.0089
0.0087
0.0087
0.0081
0.0076
0.0072
0.0065
0.0057
0.0057
0.0054
0.0052
0.0049
0.0047
0.0047
0.0046
0.0041
0.0828
0.0818
0.0505
0.0155
0.0164
0.0058
0.0010
0.0129
0.0077
0.0046
0.0040
0.0064
0.0063
0.0055
0.0015
0.0016
0.0033
0.0025
0.0038
0.0031
0.0019
0.0027
0.0059
0.0024
0.0070
0.0028
0.0021
0.0020
0.0021
0.0207
0.2052
0.2029
0.1238
0.0427
0.0407
0.0573
0.0502
0.0336
0.0225
0.0255
0.0192
0.0160
0.0159
0.0137
0.0195
0.0161
0.0215
0.0139
0.0110
0.0113
0.0112
0.0086
0.0181
0.0084
0.0181
0.0069
0.0073
0.0073
0.0069
0.0308
949
Table 5 (continued)
Country
Year
Drh
DS&T
Mexico
Cameroon
Namibia
Algeria
Yemen, Republic
Morocco
Djibouti
Brazil
United Arab Emirates
Zimbabwe
Saudi Arabia
Sudan
Cote dIvoire
Venezuela
Zimbabwe
Thailand
Indonesia
Philippines
Senegal
Benin
Swaziland
India
Swaziland
Philippines
Peru
Benin
India
Bangladesh
Burkina Faso
China
Botswana
Madagascar
Lesotho
Mali
Paraguay
Madagascar
Guinea
Lesotho
Togo
Vietnam
Guinea
Ecuador
Guyanaa
Ethiopia
Togo
Ethiopia
Uganda
Mozambique
Erithrea
Malawi
Ghana
Uganda
1990
1990
2000
1990
2000
1990
2000
1990
2000
1990
2000
1990
1990
1990
2000
2000
1990
1990
1990
1990
2000
2000
1990
2000
1990
2000
1990
1990
1990
1990
1990
1990
2000
1990
1990
2000
2000
1990
1990
1990
1990
1990
1990
1990
2000
2000
2000
1990
2000
1990
1990
1990
0.0217
0.0357
0.0109
0.0188
0.0195
0.0683
0.0298
0.0051
0.0028
0.0205
0.0023
0.0168
0.0083
0.0105
0.0357
0.0068
0.0112
0.0386
0.0380
0.0176
0.0016
0.0137
0.0006
0.0413
0.0165
0.0306
0.0090
0.0070
0.0039
0.0099
0.0064
0.0138
0.0158
0.0220
0.0096
0.0176
0.0393
0.0394
0.0302
0.0589
0.0406
0.0128
0.1428
0.0240
0.0601
0.0290
0.1228
0.0978
0.1116
0.0644
0.1354
0.1574
0.0037
0.0032
0.0030
0.0028
0.0028
0.0025
0.0025
0.0020
0.0017
0.0015
0.0013
0.0013
0.0012
0.0012
0.0011
0.0011
0.0007
0.0006
0.0004
0.0003
0.0003
0.0001
0.0001
1.32E05
1.91E05
3.26E05
0.0002
0.0003
0.0006
0.0008
0.0009
0.0009
0.0012
0.0015
0.0015
0.0018
0.0031
0.0031
0.0032
0.0033
0.0038
0.0041
0.0046
0.0051
0.0070
0.0070
0.0103
0.0145
0.0157
0.0188
0.0195
0.0204
90% conf. Interval

0.0007
0.0020
0.0015
0.0001
0.0000
0.0085
0.0018
0.0011
0.0010
0.0015
0.0008
0.0011
0.0001
0.0003
0.0046
0.0001
0.0011
0.0058
0.0060
0.0026
0.0000
0.0022
0.0000
0.0071
0.0029
0.0033
0.0018
0.0016
0.0016
0.0028
0.0024
0.0036
0.0044
0.0057
0.0038
0.0055
0.0109
0.0109
0.0096
0.0145
0.0121
0.0082
0.0307
0.0114
0.0200
0.0151
0.0350
0.0370
0.0410
0.0384
0.0505
0.0553
0.0068
0.0088
0.0045
0.0055
0.0056
0.0143
0.0073
0.0028
0.0025
0.0048
0.0019
0.0040
0.0024
0.0028
0.0074
0.0021
0.0025
0.0075
0.0073
0.0035
0.0005
0.0026
0.0002
0.0076
0.0030
0.0056
0.0015
0.0011
0.0004
0.0014
0.0007
0.0021
0.0022
0.0033
0.0009
0.0022
0.0056
0.0056
0.0038
0.0090
0.0054
0.0000
0.0230
0.0015
0.0072
0.0012
0.0170
0.0099
0.0118
0.0010
0.0140
0.0177
The skilled migration rate used in the simulations was capped at 100%, as a 40% increase would imply a migration rate in excess of 100%.
of students enrolled in science and technology in 1990 increase

by 0.49% points, had it had a skilled emigration rate of 3.78%
rather than the actual 2.7%. Out of our 82 observation, 26
indicate an average decrease in S&T, however, only one for
one of them, that of Ecuador, this eect is signicantly negative at the 95% condence level. This is not surprising as Ecuador is one of the countries lagging furthest from the frontier
(PROXIM of 0.38 in 1990), and our theory predicts that in this
case the relative productivity eect should dominate. Overall,
we can identify 36 observations for which the impact of the
change of the emigration rate is statistically signicant at the

90% condence level. The largest impact in magnitude, more
than 14% point, is observed in Trinidad and Tobago. Trinidad
and Tobago, unlike Ecuador, is a country close to the technological frontier (PROXIM of 0.78 in both 1990 and 2000). In
fact, there are two countries for which the eect of the increase
in the migration rate is much higher than for the rest, Trinidad
and Tobago and Barbados. Both of these countries have high
proximity and relatively high skilled migration, though they
are not the countries with this higher migration rate. Guyana,
950
WORLD DEVELOPMENT
Figure 3. Changes in S&T following a 40% increase in the rate of skilled migration vs. proximity to the frontier.
the country with the highest skilled migration rate in our sample, would face on average a negative eect on its S&T ratio as
a result of a further increase in migration: the reason for that is
that it has a relatively low degree of proximity (0.52). Figure 3
plots the simulations results against each countrys index of
proximity to the frontier and shows the reversal of the composition eect that occurs as the relative probability eect comes
to dominate the relative productivity eect.
Having obtained central predictions for the counterfactual
level of S&T, we are able to take our thought experiment
one step further and ask what would be the impact of such
migration-induced changes on the TFP growth for each country. Our second set of simulations is based on the estimation of
the parsimonious version of Eq. (6) presented in Table 4.3 as
Model III. Using a similar methodology as before, we simulate
the impact of changes in S&T on TFP growth for each of the
56 countries for which we are able to obtain a counterfactual
value of S&T for 1990, and compute the relative condence
intervals. Keeping the value of all other variables in the

growth equation at their actual values, we change the value
of ln(S&T) to ln(S&T + DS&T) where DS&T is the reported
change in S&T presented in Table 5. In this way we simulate
changes in growth rates for 50 countries, as presented in Table 6.
Just like Beine et al. (2007), we nd that sending countries
may end up winning or losing from an increase in skilled
migration. Out of the 50 countries in Table 6, 32 experience
a positive growth impact, for only 12 of them, however, the
change is signicantly positive at the 95% condence level.
Mauritius, which has the highest average positive eect fails
the signicance test at the 90% condence level. Among the
countries in our data set, Kenya seems to be the one that
has the most to gain from an increase in the skilled migration
rate. The remaining 18 countries experience a negative eect,
with a statistically signicant impact for 16 of them. Figure 4
plots changes in TFP growth rates against our index of
Table 6. The eect of an increase of the migration rate by 40% on TFP growth
Country
Mauritius
Malaysia
Trinidad and Tobago
El Salvador
Sri Lanka
Honduras
Kenya
Iran
Tunisia
Colombia
Jordan
Nicaragua
Uruguay
Zimbabwe
Cameroon
Mexico
DS&T
DTFP g rowth
0.02510
0.02350
0.14190
0.00870
0.02970
0.01150
0.00410
0.90870
0.00870
0.00760
0.005670
0.00520
0.00470
0.00150
0.00320
0.00370
0.00121
0.00035
0.00338
0.00027
0.00023
0.00023
0.00022
0.00022
0.00018
0.00017
0.00013
0.00013
0.00010
0.00010
0.00008
0.00008
90% conf. Interval

0.00016
0.00117
0.00546
0.00002
0.00080
0.00005
0.00005
0.00011
0.00009
0.00016
0.00013
0.00001
0.00010
0.00002
0.00001
0.00002
0.00261
0.00173
0.00544
0.00056
0.00118
0.00052
0.00041
0.00054
0.00044
0.00048
0.00038
0.00026
0.00027
0.00017
0.00016
0.00017
951
Table 6 (continued)
Country
Morocco
Turkey
Chile
Sudan
Panama
Algeria
Cote dIvoire
Venezuela
Egypt
Argentina
Brazil
Philippines
Indonesia
Benin
Senegal
Swaziland
Peru
Costa Rica
Bangladesh
India
China
Burkina Faso
Madagascar
Botswana
Mali
Paraguay
Guinea
Togo
Ecuador
Lesotho
Guyana
Mozambique
Ghana
Malawi
DS&T
DTFP g rowth
0.00250
0.01110
0.00470
0.00130
0.02840
0.00280
0.00120
0.00120
0.00960
0.00490
0.00200
0.00060
0.00070
0.00030
0.00040
0.00010
1.02E10
0.01120
0.00030
0.00020
0.00080
0.00060
0.00090
0.00090
0.00150
0.00150
1.30221
0.00320
0.00410
0.00310
0.00970
0.01450
0.01950
0.01880
0.00008
0.00007
0.00007
0.00007
0.00005
0.00005
0.00004
0.00004
0.00003
0.00003
0.00003
0.00003
0.00002
0.00002
0.00001
2.73E06
4.07E12
6.07E06
8.83E06
9.97E06
0.00001
0.00002
0.00003
0.00007
0.00010
0.00010
0.00011
0.00012
0.00022
0.00022
0.00026
0.00045
0.00076
0.00291
technological sophistication (PROXIM). The diagram illustrates that the winners are found among the countries closer
90% conf. Interval

0.00002
0.00029
0.00008
0.000003
0.00279
2.01E06
0.00001
2.71E06
0.00114
0.00026
0.00007
0.00001
3.76E06
4.22E06
2.15E06
7.05E07
0.00000
0.00121
0.00002
0.00002
0.00003
0.00004
0.00005
0.00012
0.00018
0.00019
0.00020
0.00021
0.00043
0.00040
0.00050
0.00080
0.00136
0.00534
0.00014
0.00040
0.00021
0.00014
0.00258
0.00010
0.00008
0.00007
0.00108
0.00029
0.00013
0.00005
0.00004
0.00004
0.00002
6.15E06
0.00000
0.00107
1.84E06
2.03E06
2.31E06
4.19E06
6.26E06
0.00001
0.00002
0.00001
0.00002
0.00003
3.35E6
0.00005
0.00003
0.00009
0.00017
0.00032
to the technological frontier, while the losers among countries

lagging further from it.
Figure 4. Changes in TFP growth following a 40% increase in the rate of skilled migration vs. proximity to the frontier.
952
WORLD DEVELOPMENT
According to our simulations, the losers group contains

among others China, India, and Bangladesh. Overall, the losers account for over 69% of the total population in our sample, and 2.6 billion people in total (in 2000). In terms of
welfare, our results have even stronger implications. Not only
is almost 70% of the population aected by losses, but also, given the decreasing marginal utility of income, losses among
the poorest countries should be weighted more than gains to
more developed ones.
6. CONCLUSION
By bridging two strands of literature, that on the economic
consequences of brain drain, and the growth one focussing on
the role of human capital formation, we provide several new
insights on the relation between migration, human capital formation and growth. First, we nd support for the existence of
an incentive eect on the level of (ex-ante) human capital accumulation. Second, we present evidence that the possibility of
migration also aects the types of skills that agents choose
to acquire. This underscores that the level eect exists alongside a composition eect. Third, in line with our theoretical
priors, we show that both these eects depend on the level
of technological development of the sending country (its proximity to the frontier). Dierences in wages and the degree of
marketability of migrants skills depend on the level of technological development, thus the eect of migration needs to be
discussed taking explicitly into account the technological gap
of each sending country. Fourth, our simulations show that
more than one third of the countries in our data set, representing almost 70% of the total population, suer as a result of an
increase in skilled migration. The losers are found among the
relative less developed countries, implying that, overall, the
welfare costs of skilled migration may be large.
As is the case for any empirical endeavor, our results are
highly inuenced by the quality of the data used. In this respect, data on educational attainment by eld of study leave
much to be desired, especially when focussing on nonOECD
countries as we do. The need to invest resources in the generation of better data to help empirical research cannot be overstated in this eld. Despite this caveat, our analysis provides
clear empirical support to the claim made by developing countries that recent immigration policies of OECD countries may
have dire consequences for the migrants countries of origin.
While selecting the most talented individuals from developing
countries has a clear economic rationale for destination countries, our work focuses on its implications for sending countries: by changing both the level and the composition of
human capital, an increase in the possibility of migration for
(certain types of) skilled workers reduces the growth rate of
TFP in many source countries. We draw two conclusions from
our work. First, we stress the need for better quality data to
support additional research eorts in this area, to test the
robustness of our ndings, and to better inform the policy process. Second, based on our ndings, the need for a more concerted approach to migration policy among developed and
developing countries emerges very starkly from our analysis.
NOTES
1. A recent review of the debate, including a survey of existing and
proposed policies and of their consequences, is oered by ILO (2006).
7. Di Maria and Stryszowski (2009) provide formal proof of this result in

a closely related framework.
2. Commander, Kangasniemi, and Winters (2004) present an excellent

review of this literature. See also the discussions in Beine, Docquier, and
Rapoport (2008) and Di Maria and Stryszowski (2009).
8. This modeling choice reects the idea of the small door proposed by
Stark et al. (1998): while the incentives to accumulate human capital
improve for all workers due to the possibility of migration, only a small
fraction of them eventually leaves, resulting in a more skilled, albeit
somewhat smaller, workforce. This might generate welfare gains rather
than losses, a phenomenon to which these authors refer as brain gain.
3. Anectodal evidence in this respect is abundant. Lorenzo et al. (2007),

for example, report that in the Philippines the number of nursing colleges
rose from 170 in 1999 to 460 in 2005, and that most of these new colleges
have curricula specically tailored to foreign health systems. Kangasniemi
et al. (2007) survey Indian doctors working in the UK and nd that 30% of
them acknowledge that migration prospects inuenced their education
plans and eort. Finally, Commander et al. (2008) nd evidence of
migration-induced skills accumulation in their survey of the Indian ICT
sector. The International Organization for Migration summarizes this
evidence stating that prospects of working abroad have increased the
expected return to additional years of education, and led many people to
invest in more schooling, especially in occupations in high demand overseas
(IOM, 2003).
4. Beine et al. (2011a) also nd that level eects are stronger for poorer
countries, but do not consider composition eects, nor the overall impact
of migration on growth.
5. For our purposes, higher skilled workers are those with postsecondary education, thus lower-skilled workers are all the rest.
6. Note that this does not imply that generalists are more productive in
absolute terms than technologists in either imitation or innovation, but
simply that complex tasks like the management and smooth running of
advanced research ventures require more than just technical skills.
9. Migrants moving from less developed to OECD countries account for

85% of the total according to the calculations of Docquier and Marfouk
(2006).
10. Clearly, this assumes a sign reversal in the composition eect,
however, such an eect might not occur for specic parameters combinations.
11. Here we use a measure of educational attainmentthe share of
working age natives who have completed tertiary educationas a proxy
for human capital. This is only one of many possible variables (imperfectly) measuring the quantity of human capital. Azariadis and Drazen
(1990) use adult literacy rates, Mankiw, Romer, and Weil (1992) resort to
school enrollment ratios, while Barro and Lee (1993) develop internationally comparable data on average years of schooling for a large sample
of countries and years. An interesting parallel literature, associated with
the seminal work of Eric Hanushek and co-authors (see Hanushek &
Woessmann, 2008, for a review), has emerged in recent years. They suggest
to complement such measures of attainment with measures of achievement, or educational quality, such as the results of the OECD PISA
international tests. Their results show that after controlling for cognitive
skills, educational attainment either becomes statistically insignicant, or

is shown to have very little additional eect on economic growth in a panel
of developed countries. Since our focus here is on developing economies,
however, and given the dearth of comparable data on achievement for
such countries, we are forced to abstract from considering cognitive skills,
and to concentrate on the available attainment data.
12. We refer the interested reader to the original source for a comprehensive description of the data and for discussions of the data collection
and interpolation techniques. The authors also provide a fair discussion of
the strengths and pitfalls of their data.
13. See both Docquier and Marfouk (2006) and Beine et al. (2008) for
more detailed discussions on this issue.
14. These traditional immigration countries are: Australia, Austria,
Belgium, Canada, Cyprus, Denmark, Finland, France, Germany, Greece,
Hong Kong, Iceland, Ireland, Israel, Italy, Japan, the Republic of Korea,
Luxembourg, Malta, the Netherlands, New Zealand, Norway, Portugal,
Singapore, Spain, Sweden, Switzerland, the United Kingdom, and the
United States. Our choice mirrors a similar strategy pursued by Ivlevs and
de Melo (2008).
15. The 26 countries to be dropped from the sample are: Albania,
Armenia, Azerbaijan, Belarus, Bosnia and Herzegovina, Bulgaria, Croatia, Czech Republic, Estonia, Georgia, Hungary, Kazakhstan, Kyrgyzstan, Latvia, Lithuania, Macedonia, Moldova, Poland, Romania, Russia,
Slovakia, Slovenia, Tajikistan, Turkmenistan, Ukraine, and Uzbekistan.
16. This classication is based on the current World Bank taxonomy, not
the historical situation. The World Bank divides economies into groups
according to their 2009 GNI per capita, calculated using the World Bank
Atlas method. The groups are: low income, US$995 or less; lower middle
income, US$996US$3,945; upper middle income, US$3,946US$12,195;
and high income, US$12,196 or more.
17. Following Caselli (2005), we take the values from the Actual
OSPUE column whenever available, failing that, we choose column
Imputed OSPUE, and nally the LF column when neither of the
previous two is available. The interested reader is referred to Gollin (2001)
and Bernanke and Gurkaynak (2002) for extensive discussions on the
underlying issues, the methodology of calculating labor shares and data
availability for developed and developing countries.
18. Notice that this assumption is common in the literature given the lack
of appropriate data on labor shares. Topel (1999), Caselli (2005) and
Vandenbussche et al. (2006) all make similar assumptions.
19. Our series on gross capital formation starts in 1960. Due to missing
observations, however, the earliest available data for some countries can
be as late as the 1980s.
20. Notice, however, that when estimating the three equations as a
system of seemingly unrelated equation, we cannot reject the null
hypothesis that three models are indeed unrelated. Using the Breusch
Pagan test for independent equations we obtain a p-value of 0.9813.
Moreover, 74% of the countries used in the overall analysis are employed
in at least two of the models and 43% are employed in the analysis of the
level, composition, and growth models. We have also computed the
summary statistics of the main variables in our analysis using the samples
employed in the level, composition, and growth models, respectively.
These statistics are remarkably similar across samples. The tables with
these descriptive statistics are available from the authors upon request.
21. Beine et al. (2008) do not include this variable because of its high
correlation with the initial level of human capital in their sample. In our
sample, however, the pairwise correlation between the log of initial human
953
capital and the log of the public expenditure on education as a percentage

of GDP is fairly small (0.0693), and we include both variables.
zden (2011b) nd that a larger diaspora
22. Beine, Docquier, and O
increases the migration rates for both skilled and unskilled migrants.
23. Throughout the paper, in the interest of brevity, we do not present
the rst stage regressions of the IV procedures. The results are available
from the authors upon request.
24. The coecient estimates on EDU, DENS, and REMIT are not robust
to changes in specication as evident from the estimates of Models II and
III in Table 2. These dierences are also driven by dierences in sample
size across models. When we re-estimate Model I with the same samples
used in Model II, we nd that (a) DENS is signicant at the 10% and 1%
level in the OLS and IV estimates, respectively; (b) the coecient on
REMIT is positive and signicant at 10% in both the OLS and IV
estimates; and (c) the coecient on EDU is negative and not statistically
signicant in both the OLS and IV estimates. There are no signicant
changes in the size and sign of the other coecients. These additional
estimations are available from the authors upon request.
25. As neither the migration rate nor proximity to the technological
frontier are found to be endogenous in the previous specications, we do
not conduct an IV analysis here.
26. The correlation between the log of the migration rate and the log of
proximity in this sample is 0.0089.
27. Notice that since PROXIM is a variable ranging between 0 and 1, its
natural logarithm, ln(PROXIM), is negative.
28. We cannot reject the null hypothesis that a1 = a3 at 10%: the pvalue equals 0.2934. These results conrm the ndings of Beine et al.
(2011a) who control for potential nonlinear eects by introducing a
dummy for the rich/poor countries in their data set. Beine, Docquier, and
Rapoport (2010) follow a similar strategy but nd no evidence of
nonlinear eects.
29. Note that the main conclusions of Di Maria and Stryszowski (2009),
also sketched in the extended framework in Section 2that the distortional eect of migration possibilities slows down growth in sending
countriesdoes not hinge upon the migration rates being dierent across
skill types. It is, however, true that the distortions are strongest when the
migration probabilities dier for dierent types of skills. As a consequence, our analysis might underestimate the distortionary eect of
migration.
30. As a robustness check, we re-estimate Models IIV presented in
Table 3 using a sample of countries for which both proxies for the
domestic demand for technically skilled workers are available. In these
regressions we use 65 observations for the OLS and 58 observations for
the IV estimations. The results are by and large robust with respect to
sample size with two notable exceptions discussed below. These regressions are available from the authors upon request.
31. The critical value for the 10% maximal IV size is 13.43 which is only
slightly higher than the obtained KleibergenPaap Wald F-statistics of
13.293.
32. It should be noted that the J-test produced a p-value of 0.2246 when
we used only the 58 observations for which data on both log(ROADt) and
log(PHONEt) are available, that is, 2 observations less.
33. Such sign change is not observed in the IV when we restrict ours to
countries for which both log(ROADt) and log(PHONEt) are available.
954
WORLD DEVELOPMENT
34. To calculate the thresholds, compute the marginal eect of a change

in rh,t on S&T from (5), obtaining: oS&Tt/orh,t = b1 + b3 PROXIMt.
The sign of this eect is positive when PROXIMt exceeds b1/b3, that is
using the coecient estimates from Model III, for examplewhen
PROXIMt > 0.5022.
35. In our sample there are twenty countries with a proximity indicator
below 0.5022 in 1990 or 2000. Among these countries we nd Burkina
Faso, the Central African Republic, Ethiopia, Guinea, Madagascar,
Malawi, and Togo.
36. Alternatively, as suggested by one of the referees, it might simply be
the case that for very poor countries, it is too dicult or expensive to oer
S&T degrees. This is, however, not the case for countries in our data set
given the positive value of S&T for all of them.
37. Abella (2006) discusses policies implemented by several OECD
countries involved in the on-going international competition for global
talent. Point systems, temporary admissions under skill-based categories
(e.g., H-1B visas), and facilitation of family migration for specic
categories of workers have all been used to favor workers in the eld of
science and technology, markedly health professionals, geneticists and IT
specialists.
38. As customary in the literature, we use 5-year-period averages for the

growth rate, reducing the risk of capturing business cycle eects.
39. We could not nd any statistical evidence that the eect of the
proportion of workers with technical skills on TFP growth becomes
negative close enough to the frontier. The threshold for the change in sign
(^c3 =^c4 0:9820 in Model III, 0.9497 in Model IV, and 0.9820 for
Model V. None of these estimates are statistically dierent from 1 with pvalues 0.9748, 0.9226, and 0.9535 for the tests in Models III, IV, and V,
respectively.
40. We chose to increase the emigration rate by 40% as the average rate
of increase in the emigration rate during 19902000 for the countries in
our sample was 38%.
41. For a detailed discussion of this method, together with illustrative
examples, please refer to King et al. (2000). The Clarify documentation,
available on-line at www.stanford.edu/~tomz/software/clarify.pdf, provides additional details on the method and its applications.
REFERENCES
Abella, M. (2006). Global competition for skilled workers and consequences. In C. Kuptsch, & E. F. Pang (Eds.), Competing for global
talent. Geneva: International Labour Oce.
Azariadis, C., & Drazen, A. (1990). Threshold externalities in economic development. The Quarterly Journal of Economics, 105(2),
501526.
Barro, R. J., & Lee, J. (1993). International comparisons of educational
attainment. Journal of Monetary Economics, 32(3), 363394.
Batista, C., Lacuesta, A., & Vicente, P. C. (2012). Testing the brain gain
hypothesis: Micro evidence from cape verde. Journal of Development
Economics, 97(1), 3245.
Beine, M., Defoort, C., & Docquier, F. (2011a). A panel data analysis of
the brain gain. World Development, 39(4), 523532.
zden, C. (2011b). Diasporas. Journal of
Beine, M., Docquier, F., & O
Development Economics, 95(1), 3041.
Beine, M., Docquier, F., & Rapoport, H. (2001). Brain drain and
economic growth: Theory and evidence. Journal of Development
Economics, 64, 275289.
Beine, M., Docquier, F., & Rapoport, H. (2007). Measuring international
skilled migration: A new database controlling for age of entry. The
World Bank Economic Review, 21(2), 249254.
Beine, M., Docquier, F., & Rapoport, H. (2008). Brain drain and human
capital formation in developing countries: Winners and losers.
Economic Journal, 118, 631652.
Beine, M., Docquier, F., & Rapoport, H. (2003). Brain drain and LDCs
growth: Winners and losers. IZA working paper, 869.
Beine, M., Docquier, F., & Rapoport, H. (2010). On the robustness of
brain gain estimates. Annales dEconomie et de Statistiques, 97/98,
143165.
Benhabib, J., & Spiegel, M. M. (1994). The role of human capital in
economic development evidence from aggregate cross-country data.
Journal of Monetary Economics, 34(2), 143173.
Bernanke, B., & Gurkaynak, R. (2002). Is growth exogenous? Taking
Mankiw, Romer, and Weil seriously 2001. NBER Macroeconomics
Annual, 16, 1172.
Carrington, W. J., & Detragiache, E. (1998). How big is the brain drain?
IMF working papers, 98/102.
Caselli, F. (2005). Accounting for cross-country income dierences. In P.
Aghion, & S. Durlauf (Eds.). Handbook of economic growth (Vol. 1,
pp. 679741). Amsterdam: North-Holland (chap. 10).
Caselli, F., & Coleman, J. (2006). The world technology frontier. The
American Economic Review, 96(3), 499522.
Clemens, M. (2007). Do visas kill? Health eects of African health
professional emigration. Center for global development working paper,
114.
Commander, S., Chanda, R., Kangasniemi, M., & Winters, L. A. (2008).

Must skilled migration be a brain drain? Evidence from the Indian
software industry. The World Economy, 31(2), 187211.
Commander, S., Kangasniemi, M., & Winters, L. A. (2004). The brain
drain: Curse or boon? A survey of the literature. In R. E. Baldwin, &
L. A. Winters (Eds.), Challenges to globalization: analyzing the
economics. Chicago, IL: Chicago University Press.
Connell, J., Zurn, P., Stilwell, B., Awases, M., & Braichet, J.-M. (2007).
Sub-Saharan Africa: Beyond the health worker migration crisis?.
Social Science & Medicine, 64(9), 18761891.
Di Maria, C., & Stryszowski, P. (2009). Migration, human capital
accumulation and economic development. Journal of Development
Economics, 90(2), 306313.
Docquier, F., & Marfouk, A. (2006). International migration by educational attainment, 19902000. In C. Ozden, & M. Schi (Eds.),
International Migration, Remittances and the Brain Drain
(pp. 151199). New York: Palgrave Macmillan.
Durlauf, S. N., Johnson, P. A., & Temple, J. R. (2005). Growth
econometrics. In P. Aghion, & S. N. Durlauf (Eds.). Handbook of
economic growth (Vol. 1, pp. 555677). Amsterdam: Elsevier (chap. 8).
Dustmann, C., Fadlon, I., & Weiss, Y. (2011). Return migration, human
capital accumulation and the brain drain. Journal of Development
Economics, 95(1), 5867.
Feinstein, J. S. (2006). The nature of creative development. Stanford, CA:
Stanford University Press.
Gerschenkron, A. (1962). Economic backwardness in historical perspective:
A book of essays. Cambridge, MA: Belknap Press of Harvard
University Press.
Gibson, J., & McKenzie, D. (2011). The microeconomic determinants of
emigration and return migration of the best and brightest: Evidence
from the Pacic. Journal of Development Economics, 95(1), 1829.
Gollin, D. (2001). Getting income shares right. Journal of Political
Economy, 110(2), 458474.
Hanushek, E. A., & Woessmann, L. (2008). The role of cognitive skills in
economic development. Journal of Economic Literature, 46(3),
607668.
ILO (2006). In C. Kuptsch, & E. F. Pang (Eds.), Competing for global
talent. Geneva: International Labour Oce.
IOM (2003). World migration 2003: Managing migrationChallenges and
responses for people on the move. Geneva: International Organization
for Migration.
Ivlevs, A., & de Melo, J. (2008). FDI, the brain drain and trade: Channels
and evidence. CEPR Discussion Paper, No. DP7002.
Iyigun, M. F., & Owen, A. L. (1999). Entrepreneurs, professionals, and
growth. Journal of Economic Growth, 4(2), 212232.

Kangasniemi, M., Winters, L. A., & Commander, S. (2007). Is the medical
brain drain benecial? Evidence from overseas doctors in the UK.
Social Science & Medicine, 65(5), 915923.
King, G., Tomz, M., & Wittenberg, J. (2000). Making the most of
statistical analyses: Improving interpretation and presentation. American Journal of Political Science, 44(2), 347361.
Kleibergen, F., & Paap, R. (2006). Generalized reduced rank tests using
the singular value decomposition. Journal of Econometrics, 133,
97126.
Lorenzo, M., Galvez-Tan, J., & Javier, L. (2007). Nurse migration from a
source country perspective: Philippine country case study. Health
Services Research, 42, 14061418.
Malthus, T. A. (1798). An essay on the principle of population (1986th ed.).
London: W. Pickering.
Mankiw, N. G., Romer, D., & Weil, D. N. (1992). A contribution to the
empirics of economic growth. The Quarterly Journal of Economics,
107(2), 407437.
Mariani, F. (2007). Migration as an antidote to rent-seeking?. Journal of
Development Economics, 84(2), 609630.
Mountford, A. (1997). Can a brain drain be good for growth in the source
economy?. Journal of Development Economics, 53(2), 287303.
Murphy, K. M., Shleifer, A., & Vishny, R. W. (1991). The allocation of
talent: Implications for growth. Quarterly Journal of Economics,
106(2), 503530.
Stark, O., Helmenstein, C., & Prskatwetz, A. (1997). A brain gain with a
brain drain. Economic Letters, 55, 227234.
Stark, O., Helmenstein, C., & Prskatwetz, A. (1998). Human capital
formation, human capital depletion, and migration: a blessing or a
curse?. Economic Letters, 60, 363367.
955
Stark, O., & Wang, Y. (2002). Inducing human capital formation:

Migration as a substitute for subsidies. Journal of Public Economics,
86, 2946.
Stock, J., & Yogo, M. (2005). Weak instruments in linear IV regression. In
D. Andrews, & J. Stock (Eds.), Identication and inference in
econometric models: Essays in honor of Thonas J. Rothenberg (chap.
5). New York, NY: Cambridge University Press.
Topel, R. (1999). Labor markets and economic growth. In O. Ashenfelter,
& D. Card
(Eds.). Handbook of labor economics (Vol. 3C,
pp. 29432984). Amsterdam: Elsevier (chap. 44).
Urdal, H. (2005). People vs. Malthus: Population pressure, environmental
degradation, and armed conict revised. Journal of Peace Research,
42(4), 417434.
Vandenbussche, J., Aghion, P., & Meghir, C. (2006). Growth, distance to
frontier and composition of human capital. Journal of Economic
Growth, 11(2), 97127.
Vidal, J.-P. (1998). The eect of emigration on human capital formation.
Journal of Population Economics, 11(4), 589600.
Wooldridge, J. M. (1995). Score diagnostics for linear models estimated by
two stage least squares. In G. S. Maddala, P. C. B. Phillips, & T. N.
Srinivasan (Eds.), Advances in econometrics and quantitative economics:
Essays in honor of Professor C. R. Rao (pp. 6687). Oxford: Blackwell.
World Bank (2009). World development indicators. Washington, DC: The
World Bank.
Young, A. (1998). Growth without scale eects. Journal of Political
Economy, 106(1), 4163.
Yusuf, S. (2007). From creativity to innovation. Policy research working
paper, 4262. World Bank.

Migration, Human Capital Formation (Brain Drain)

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Migration, Human Capital Formation (Brain Drain)

Uploaded by

Copyright:

Available Formats

World Development Vol. 40, No. 5, pp.

Migration, Human Capital Formation, and Growth:

revised version of the previous paper, Beine et al. (2008) use

Over the last decades, an increasing number of developed

* We are grateful to seminar participants at the Queens University of

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

found in Murphy et al. (1991) provides an early example of

where Ait represents the productivity of intermediate i at

I N N OVA TIO N (N)

endeavors (innovation) rather than in imitative activities (e.g.,

Figure 2. The skill composition of the workforce.

where the H and F superscripts identify domestic and foreign

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

workers tend to be more successful in emigrating. Thus, if rT is

cation (MSsec,i) to instrument for the rate of skilled migration

Table 1. Descriptive statistics

As shown in Table 1, the average proximity index in the

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

endogeneity issues in depth. We start o with an investigation

The dependent variable, Dlog(Ha,0090) = log(Ha,00)

Notes: Heteroskedasticity robust standard errors are reported in parenthesis.

of coecients and the same level of signicance under both

can reject the null hypothesis of weak instruments based on

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

to the technological frontier, or their interactionis shown to

where the variables S&Tt, rt, PROXIMt, EDUt, SSA, and

our results: phones per employee and the percentage of paved

Notes: Heteroskedasticity robust standard errors are reported in parenthesis.

focussed on two mechanisms. First, as countries move closer

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

capital on the growth rate of the countrys TFP. Here we build

The dependent variable, gTFP,5yr,t is the average annual growth

The set of relevant regression results are presented in Table 4.

Table 4. Total factor productivity growthdependent variable: gTFP,5yr,t

Notes: Heteroskedasticity robust standard errors are reported in parenthesis.

F-statistic is 216.91, which is much higher than 10, the critical

Table 5. The eect of an increase of the migration rate by 40% on S&T

90% conf. Interval

Trinidad and Tobago

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

90% conf. Interval

of students enrolled in science and technology in 1990 increase

change of the emigration rate is statistically signicant at the

intervals. Keeping the value of all other variables in the

90% conf. Interval

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

90% conf. Interval

to the technological frontier, while the losers among countries

According to our simulations, the losers group contains

7. Di Maria and Stryszowski (2009) provide formal proof of this result in

2. Commander, Kangasniemi, and Winters (2004) present an excellent

3. Anectodal evidence in this respect is abundant. Lorenzo et al. (2007),

9. Migrants moving from less developed to OECD countries account for

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

capital and the log of the public expenditure on education as a percentage

34. To calculate the thresholds, compute the marginal eect of a change

38. As customary in the literature, we use 5-year-period averages for the

Commander, S., Chanda, R., Kangasniemi, M., & Winters, L. A. (2008).

MIGRATION, HUMAN CAPITAL FORMATION, AND GROWTH

Stark, O., & Wang, Y. (2002). Inducing human capital formation:

You might also like