WPS5153
Policy Research Working Paper 5153
International Growth Spillovers,
Geography and Infrastructure
Mark Roberts
Uwe Deichmann
The World Bank
Development Research Group
Environment and Energy Team
December 2009
Policy Research Working Paper 5153
Abstract
There is significant academic evidence that growth in The analysis finds significant evidence for heterogeneity
one country tends to have a positive impact on growth in growth spillovers, which are strong between OECD
in neighboring countries. This paper contributes to countries and essentially absent in SubSaharan Africa.
this literature by assessing whether growth spillovers The analysis further finds strong interaction between
tend to vary significantly across world regions infrastructure and being a landlocked country. This
and by investigating the contribution of transport suggests that growth spillovers from regional "success
and communication infrastructure in promoting stories" in SubSaharan Africa and other lagging world
neighborhood effects. The study is global, but the regions will depend on first strengthening the channels
main interest is on SubSaharan Africa. The authors through which such spillovers can spreadmost
define neighborhoods both in geographic terms and importantly infrastructure endowments.
by membership in the same regional trade association.
This papera product of the Environment and Energy Team, Development Research Groupis part of a larger effort in
the department to understand the role of geography in economic development. Policy Research Working Papers are also
posted on the Web at http://econ.worldbank.org. The authors may be contacted at mr10013@cam.ac.uk and udeichmann@
worldbank.org.
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Produced by the Research Support Team
International Growth Spillovers, Geography and Infrastructure*
Mark Roberts
Department of Land Economy
University of Cambridge
19 Silver Street
Cambridge
CB3 9EP, UK
Email: mr10013@cam.ac.uk
Uwe Deichmann
Development Research Group
The World Bank
1818 H Street, NW
Washington, DC
20433, USA
Email: udeichmann@worldbank.org
*
The authors would like to thank participants in the conference "Spatial Economics and Trade" which was
jointly organized by the University of Strathclyde and the Scottish Institute for Research in Economics
(SIRE) and held on 26th July, 2008 for their very useful comments. They would similarly like to thank
participants of the International Symposium "Development Prospects for the 21st Century", organised by
the International Celso Furtado Center for Development Policies and held in Rio de Janeiro on 67th
November, participants of a seminar at the University of Cambridge, UK, Souleymane Coulibaly and
Nancy Lozano. All errors and omissions, however, remain the sole responsibility of the authors. A
previous version of this paper was prepared as background for the World Development Report 2009
"Reshaping Economic Geography." The findings, interpretations, and conclusions expressed in this paper
are entirely those of the authors. They do not necessarily represent the views of the International Bank for
Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive
Directors of the World Bank or the governments they represent.
1. Introduction
Significant economic growth experiences for countries rarely occur in isolation. More
typically, countries do well if their neighbors do well. The best known example is the
industrial revolution. Originating in England, it quickly (by the standards of the 18th and
early 19th century) spread to continental Europe in an almost contagionlike process.
More recently, the "East Asian miracle" saw Japan's dynamic growth pull many of its
neighbors along to middle or even high income status. Conversely, in world regions
where no initial growth poles emerge and where mechanisms for transmission of
spillover benefits are absent, entire regions may remain stagnant for long periods.
In this paper we aim to contribute to the analysis of spillovers between countries in
different regions of the world, with a particular emphasis on growth in SubSaharan
Africa (SSA). We take as a starting point studies by Easterly and Levine (1998) and
Collier and O'Connell (2007) and extend them in two main ways. First, we look at
spillovers across contiguous countries in a spatial econometric framework, but our main
focus is on interaction processes that spread through regional trade agreements. Secondly,
we assess a mechanism for the transmission of benefits, namely through infrastructure
investments which facilitate interaction between countries through trade and
communication of ideas. We are particularly interested to see if spillovers spread in a
spatially homogenous process, as Easterly and Levine (1998) assume; or whether these
effects are likely to be spatially heterogeneous on account of differences in the degree of
spatial or institutional integration, which is closer to the assumptions in Collier and
O'Connell (2007).
1
Understanding the scale and geographic scope of spillovers is an important piece in the
overall growth puzzle. For example, one reason why crosscountry growth spillovers
might be localized is because spillovers of knowledge between countries are also
localized. This may be the case if knowledge is embodied in intermediate goods and trade
in such goods is more probable between countries which are geographically proximate1
or which share a history of strong trading relations (Coe and Helpman, 1995, pp 861
863). Growth theory suggests that these trading partners will form socalled convergence
clubs with economic growth correlated across neighboring countries (see, inter alia,
Grossman and Helpman, 1991, Ertur and Koch, 2005, 2007). This might help to explain
why successive waves of economic development have tended to be confined within
relatively welldefined geographic regions.
With localized crosscountry growth spillovers, not only will growth miracles tend to be
spatially correlated, but so too will growth disasters. This might form part of the
explanation for SSA's growth failure, both in absolute terms and relative to other
developing parts of the world, in recent decades. It might also account for the large and
statistically significant Africa dummy that has been a recurring finding in the empirical
growth literature (Easterly and Levine, 1998). In that case, a coordinated effort between
SSA nations to stimulate domestic economic growth will have a multiplier effect
stemming from the neighborhood interactions implied by the existence of localized
spillovers (ibid., pp 136137).
1
Empirical evidence from the estimation of gravity models of trade provides strong support for the
hypothesis that the strength of trade between two countries is inversely related to distance (see, for
example, Brakman et al., 2009, chapter 1).
2
The above assumes that crosscountry spillovers are equally strong between countries in
different geographic regions or at different levels of development. To the extent that
such spillovers are localized because they are mediated by trade, this seems unlikely.
Whereas advanced industrialized countries and the emerging economies of East Asia
have high levels of intraregional trade2, trade between SSA countries is negligible: about
onehalf of observed bilateral manufacturing trade flows between SSA countries are
equal to zero (Bosker and Garretsen, 2008, p 12). Formal barriers to trade in the form of
tariffs and quotas are a major reason, as are inefficient customs procedures and the
inadequate state of transportation infrastructure within the region (Buys et al.,
forthcoming). This implies that before they can even begin to contemplate the potential
leveraging of spillover effects, policymakers in the region need to cultivate such effects
through policies designed to encourage regional integration. This will be particularly
important for resource poor landlocked countries within the region whose only credible
development strategy is integration with coastal and resource rich countries.3 The hope is
that some of the countries with a more fortunate geographic location or with better
endowments will eventually takeoff, dragging the resource poor landlocked neighbors
along with them (Collier and O'Connell, 2007). Switzerland provides an example.
Despite being landlocked, it suffers no disadvantage because it is tightly integrated with
2
Intraregional trade in East Asia today approximates that within the European Union (EU) (World Bank,
2008, p 195).
3
SSA is distinguished from the rest of the developing world by the disproportionate percentage of its
population which lives in resource poor landlocked countries. According to Collier and O'Connell (2007, p
7), 35 % of the region's population lives in such countries compared to a mere 1 % in developing countries
outside of SSA. These figures are based on a comparison of 43 SSA countries with a sample of 56 other
developing countries and relate to the period 19902000.
3
the rest of Europe, importantly through highly developed levels of transportation and
telecommunications infrastructure.4
This paper addresses the above issues using empirical growth regressions incorporating
crosscountry spillover effects. Due to data limitations we employ two samples of
countries. The first, "long sample" consists of 131 countries for the period 19702000.
This is used to test for the global importance of localized spillovers of growth between
countries using panel data techniques which allow for control for both observable and
unobservable timeinvariant determinants of growth. We explore three different
neighborhood definitions. Two of these define neighbors in purely geographic terms;
with the third, on which most of our attention focuses, defining neighbors as sharing a
formal regional trade agreement (RTA). By splitting the sample into subsamples
defined according to geographic regions and income levels, the paper also investigates
the spatial heterogeneity of localized growth spillover effects.
The second sample consists of 142 countries and covers the period 19922000. This
"short sample" is used to analyze the interrelationships between landlockedness, a
country's level of transport and telecommunications infrastructure development, and the
strength of localized growth spillovers. Given the shorter sampleperiod, we need to rely
on crosssectional, rather than panel, techniques.
4
Also, in the absence of welldeveloped transportation infrastructure, the coastal neighbors of landlocked
countries in SSA are barriers to access to world markets. By contrast, Switzerland's coastal neighbors
themselves constitute the market (Collier and O'Connell, 2007, p 4).
4
We find some indication that growth spillovers are indeed heterogeneous and are more
pronounced for neighborhood specifications based on RTA membership than on simple
contiguity. Spillovers are most pronounced in OECD countries but appear basically
absent in SSA. In particular, landlocked countries experienced little benefit from their
coastal neighbors, and this can be attributed in large part to poor transportation and
communications infrastructure. We illustrate the implications of these findings in a
hypothetical simulation exercise designed to illustrate the potential welfare losses
associated with the lack of integration. If Switzerland, during the period 1970 to 2000,
had been exposed to the growth spillovers experienced by the equally landlocked Central
African Republic, it would have forgone more than $330 billion (2000 international
dollars) of GDP, all else being equal.
The layout of the remainder of the paper is as follows. Section 2 reviews related
empirical literature estimating the strength of crosscountry growth spillovers. Following
this, section 3 sets out the different methods of defining a country's neighbors used in this
paper, while section 4 presents the results of exploratory spatial analysis of our 1970
2000 panel data. This provides evidence on the extent of clustering of growth rates
across countries and is, therefore, suggestive of the possible presence of heterogeneous
spillover effects. Section 5 then presents results on the strength of localized crosscountry
spillovers using both the full sample of countries for the period 19702000 and various
geographically and income based subsamples. This is followed, in section 6, by the
presentation of results relating to the interactions between landlockedness, infrastructure
5
development and the strength of growth spillovers derived using the 19922000 sample.
Section 7 presents our hypothetical simulation exercise. Finally, section 8 concludes.
2. Literature on CrossCountry Growth Spillovers
A large literature models growth between subnational regions as being interdependent
due to various factors, including growth spillovers (see, inter alia,; Armstrong, 1995a,
1995b; Bernat, 1996; Fingleton and McCombie, 1998; Rey and Montuori, 1998; Roberts,
2004; Angeriz et al., 2008, 2009; Bosker, 2007). This literature typically estimates the
strength of spillovers as being given by the estimated partial correlation between the
growth of one such region and a weighted average of the growth rates of neighboring
regions conditional on controlling for various observed local determinants of growth.5
To deal with the simultaneity problem which this implies, such estimation is ordinarily
carried out within a single equation framework in a purely crosssectional setting relying
on maximum likelihood (ML) estimation techniques pioneered by Anselin (1988).6, 7
In contrast, the literature on international growth spillovers between countries which
employs similar techniques is comparatively thin. Ades and Chua (1997) do not
estimate the strength of growth spillovers per se, but rather the impact of neighboring
countries' political instability. They find that such instability has a significant negative
impact on domestic growth and reduces the steadystate level of income per capita in the
5
This generates a specification referred to as the spatial lag model. Another frequently estimated type of
spatial model is the socalled spatial error model. In this model, the disturbance term has a spatial
autoregressive structure (Anselin, 1988).
6
Although, applications which use spatial panel techniques are becoming more common (see, inter alia,
Angeriz et al, 2008, 2009; and Arbia and Piras, 2005).
7
As an alternative to ML, instrumental variable (IV) based estimation techniques are sometimes used (see,
for example, Roberts, 2004, for an application).
6
domestic economy (ibid., p. 287). This effect works, at least partly, through the
disruption which regional political instability causes to trade. Moreno and Trehan (1997)
directly examine growth spillovers, making a distinction between gross spillover effects
and net spillover effects. Gross (net) spillover effects are associated with the estimated
coefficient on neighboring country growth in a crosssectional growth regression when
all other explanatory variables are excluded (included) (ibid., p. 405). Using data for
19651989, they find evidence that both types of spillover are extremely strong.
Other studies which use crosssectional growth regressions include Abreu et al (2004),
and Ertur and Koch (2005, 2007). Abreu et al (2004) find strong and statistically
significant spillovers of total factor productivity (TFP) growth over the period 1960
2000. Ertur and Koch (2007) build technological interdependence directly into a modified
version of the Solow model, allowing them to derive a conditional convergence style
estimating equation for growth which includes spatial effects. When estimating this
equation, they find neighboring growth to have a significant positive impact on domestic
growth. Ertur and Koch (2005) obtain similar results, although, in this case, the authors
derive their estimating equation from a twosector model of growth. In both papers, Ertur
and Koch further estimate their conditional convergence equations by applying Bayesian
methods which allow for parameter heterogeneity, which they find to be important.
Easterly and Levine (1998) adopt a similar approach to the above studies. However,
rather than using purely crosssectional data, they use average growth rates for three
7
pooled 10year periods.8 In doing so, they place specific emphasis on SSA in both
motivating their analysis and interpreting their results. They note that including neighbor
growth in their regressions eliminates the significance of the dummy variable for the
region which has traditionally been found to be important in empirical growth regressions
(ibid., 1998, p. 133). They furthermore find large neighborhood multiplier effects,
contending that these effects might have locked SSA into a slow growth pattern because
they imply that slow growth in neighboring countries becomes mutually reinforcing.
However, as they note, such effects also have a potentially positive upside for SSA. They
imply that, if pursued in unison, a large policy change will not only have directly
beneficial effects on the growth of each country in the region, but also reinforcing
indirect effects. They calculate a neighborhood multiplier of 2.2, which is the factor by
which the direct effects on domestic growth of a policy change are amplified if that
policy change is pursued in unison (ibid., pp 136137).
All of the above studies are concerned with longrun growth spillovers. There also exist
several studies which instead focus on shortrun growth spillovers, using annual growth
data within a panel framework. Weinhold (2002) models a country's growth as being
dependent on both contemporaneous and lagged values of a neighboring growth variable.
She interprets the former as capturing spatial dependence that might arise through
common shocks and the latter as capturing genuine growth spillovers. In doing so, she
allows for parameter heterogeneity between industrialized and developing countries,
finding that significant growth spillovers exist only for the latter.
8
Contrary to our analysis in section 5, however, they do not take advantage of the panel structure of their
data set to control for timeinvariant unobservable determinants of growth that might be correlated with
their neighboring growth variable.
8
Meanwhile, Behar (2008) tests whether neighborhood spillover effects exist over and
above similar effects which operate at both the regional and global levels. Depending on
the exact definition of neighborhood adopted, he finds that a 1 % increase in the growth
rate for all countries in the neighborhood increases the domestic rate of growth by
between 0.068 % and 0.106 %. He also presents suggestive evidence that net
neighborhood spillover effects are stronger in North America and Asia than in SSA,
whilst, for Europe, there is no significant net neighborhood spillover effect. SSA is also
found to be subject to strong regional spillover effects and Behar concludes that this
result is largely driven by the commodity exporting countries, particularly the oil
exporting countries (ibid., 2008, p. 19). Unlike other studies which use ML or IV
techniques, Behar does not explicitly control for the endogeneity of neighboring growth.
He argues that, given that endogeneity only exists under the spillover hypothesis, tests
based on the null hypothesis of no spillovers will still be valid. He furthermore argues
that the feedback effects to own country growth implied by the existence of spillovers,
and, hence, the extent of bias, are likely to be minimal (ibid., p. 7).
Finally, Collier and O'Connell (2007) is the contribution which, along with Easterly and
Levine (1998), is of most direct relevance to the current study. Whereas Easterly and
Levine impose a spatially homogenous spillover parameter across countries in their
sample and conclude that the resulting evidence of strong crosscountry spillover effects
might help to explain SSA's poor growth performance over recent decades, Collier and
O'Connell offer a competing hypothesis. Namely, that crosscountry spillover effects are
likely to be much weaker between SSA countries than between countries globally on
9
account of the region's lack of integration. This is particularly so for resource poor
landlocked countries, whose performance, globally, can be expected to be more
dependent than coastal and resource rich countries on neighbor growth. Consistent with
this, Collier and O'Connell find that, globally, a 1 percentage point increase in neighbor
growth is associated with a, statistically significant, 0.392 % increase in the domestic
growth rate, rising to a 0.71% increase for resource poor landlocked countries. However,
for resource poor landlocked countries in SSA, there is no significant estimated influence
of neighbor growth. This implies that, contra Easterly and Levine, strong neighborhood
multiplier effects, which have the potential to enhance the growth returns from
coordinated policy actions, exist only outside of SSA.9, 10
In this paper we investigate in greater depth the validity of the hypotheses forwarded by
Collier and O'Connell. To do so, however, we follow the majority of the literature in
making use of longerrun growth data as opposed to annual data. The framework,
therefore, bears more resemblance to the standard empirical growth literature, with the
crucial difference that it allows for spatial effects. Through analyzing the
interrelationships between landlockedness, the level of transport and telecommunications
infrastructure, and the strength of crosscountry growth spillovers, this paper explores in
9
Collier and O'Connell (2007) include a full set of year dummies in their regressions. It is not clear,
however, whether or not they also include country FEs. Also unclear is precisely how they define a
country's neighbors. In using annual macroeconomic data, spurious correlation associated with non
stationary data may also pose a problem.
10
The results discussed in the main text are Collier and O'Connell's results based on OLS. They also report
qualitatively similar results using IV and least absolute deviation (LAD) estimation methods (see Collier
and O'Connell, 2007, p. 40, Table 20, and Table A3, p 55).
10
greater detail the possible reasons for heterogeneous growth spillover effects across
landlocked countries in different parts of the world.11
3. Defining a Country's Neighbors
In previous studies, purely geographical criteria have frequently been used to identify a
country's neighbors (Abreu et al, 2004; Ads and Chua, 1997; Behar, 2008; Ertur and
Koch, 2005, 2007; Moreno and Trehan, 1997; and Weinhold, 2002). Additionally, some
studies have used trade data to estimate the strength of neighbor interactions (Easterly
and Levine, 1998; Moreno and Trehan, 1997; and Weinhold, 2002). We use three
alternative definitions of neighborhood by specifying three alternative neighborhood, or
spatial, weights (W) matrices:12
W1: Contiguity definition
For wij wii, wij = 1 if countries i and j share a contiguity relationship; otherwise
wij = 0.13
W2: Distance definition
11
We have confined ourselves in this section to a discussion of literature which focuses more or less
directly on the relationship between the growth of an economy and the growth of its neighbors and which
uses parametric techniques. However, there also exist studies which have found evidence of strong cross
country growth spillovers using nonparametric techniques (see, for example, Conley and Ligon, 2002) and
which have identified spillovers from foreign R&D levels to domestic productivity (Coe and Helpman,
1995).
12
In all analysis we follow the standard convention of rowstandardizing all W matrices, so that
N
wj 1
ij 1.
13
We assume that wii = 0 to prevent a country's growth rate being included in the definition of the growth
of neighboring countries and to avoid using a country's growth rate to predict itself in the analysis of
sections 5 and 6.
11
For wij wii, wij = d ij 2 , where dij denotes the weighted average of the bilateral
distances between cities of population greater than 100,000 in countries i and j;
otherwise wij = 0. This implies every country is a neighbor of every other country,
2
but the strength of interaction diminishes with d ij .
W3: RTA definition
For wij wii, wij = 1 if countries i and j shared membership of a regional trade
agreement (RTA) in October 2003; otherwise wij = 0.14
In constructing W1, we adopt a liberal definition of contiguity such that wij = 1 if
countries i and j are separated by a land or river border or by less than 400 miles of open
water, uninterrupted by the territory of a third country.15 This allows us to maximize
sample size by reducing the number of countries which would otherwise have no
neighbors under more stringent definitions.16
Although W1 and W2 are clearly exogenous to the growth process, which is important for
the regression analysis of later sections, one might worry that the same is not true for W3.
In particular, for RTAs which came into force over the sample period, the concern exists
that the probability of two countries having entered into such an agreement may have
been related to their growth performances over the same period. However, we consider
this concern to be relatively minor. Even where a RTA came into force only at some
14
A full list of RTAs included in the definition of W3 is provided in table A1 of the data appendix.
15
400 miles is the maximum distance at which two 200 mile exclusive economic zones can interact
(Stinnett et al, 2002).
16
Our main results are robust to the use of more stringent definitions.
12
point during the sampleperiod, rather than before it, such an agreement usually embodies
longstanding trading relationships which predate both the formal RTA and the beginning
of the sampleperiod. In many cases, also, a new RTA builds on a previous RTA that
was in force at the beginning of the period.17 Furthermore, to assess the robustness of
some of our main results to this concern, we engaged in some experimentation which
involved redefining W3 to be based only on, respectively, pre1991 and pre1981 RTA
membership. Although this had the effect of reducing sample size, we found our main
results still to hold.18
Our three W matrices embody very different neighborhood structures. For example,
using a common sample of 110 countries, figure 1 shows that W1 is relatively sparse
compared to W3. While each country has, on average, just over four neighbors with W1,
the average number of neighbors with W3 is just over twenty, because many countries
belong to multiple, overlapping, RTAs. Furthermore, whereas 69.9 % of contiguous
country pairs, as defined by W1, share membership of a RTA, only 15.3 % of RTA
country pairs has a contiguity relationship.
17
To demonstrate, we can cite two examples. Firstly, in our underlying data source, membership of the
East African Community (EAC) is listed as dating back to 2000. This is consistent with the revival of the
EAC in 2000. However, the EAC was first founded in 1967 (before collapsing in 1977), which is prior to
the beginning of our 19702000 sampleperiod. Second, whilst membership of the Economic Cooperation
Organization (ECO) dates back to 1985, this agreement is the successor to the Regional Cooperation for
Development agreement, which was in force between 19621979.
18
A related endogeneity concern is the existence of countries which withdrew from a RTA during the
sampleperiod or of RTAs which collapsed. An example of the former is Lesotho's withdrawal from the
Common Market for Eastern and Southern Africa (COMESA) in 1997. We cannot ruleout the possibility
of the dissolution of RTAs and of decisions to withdraw from RTAs having been endogenous to the growth
process over the sampleperiod. Again, however, in many cases where RTAs have been dissolved over the
sampleperiod, they have been succeeded by other agreements (captured by W3) involving similar
configurations of countries. For example, the UK, Denmark, Sweden, Austria and Portugal all departed
from the European Free Trade Association (EFTA) during the sampleperiod to become members of the
European Community.
13
Figure 1: Comparison of the structure of the neighborhood weights matrices W1 and W3
(a) Structure of W1 weights matrix (b) Structure of W3 weights matrix
In the remainder of this paper, our main focus is on the results obtained using W3, only
highlighting results using W1 and W2 where these show important differences.19 This is
because our main interest is in identifying spillovers due to institutional linkages between
countries that are facilitated through shared infrastructure and W3 relates more closely to
this notion of a country's integration into its neighborhood. In specifying W3, no
distinction is made between the degrees of integration embodied in different RTAs.
Rather, in the exploratory analysis of section 4, we expect stronger clustering of growth
rates where RTAs have contributed to more effective integration. Likewise, it is
anticipated that the regression analysis of sections 5 and 6 will detect stronger cross
country spillover effects where integration has been more effective. Treating all RTAs
equally also mitigates endogeneity concerns.20
19
The full set of results using all three matrices is available upon request from the authors.
For the analysis in section 5 we also considered a variant of W3 which restricted attention to a subset of
20
RTAs which are more prominent. Again, although this had the effect of reducing sample size, it left the
main results unchanged.
14
4. Exploratory Analysis of the Clustering of CrossCountry Growth Rates
Before assessing the strength of spillovers in a regression framework, we apply
exploratory spatial data analysis (ESDA) techniques which assess the extent to which
growth rates across neighboring countries are spatially autocorrelated. We use the
average annual logarithmic growth rate of real GDP per capita as our measure of
economic growth. We make use of the same sampleperiod of 19702000 on which the
regression analysis of section 5 focuses, splitting this into 5yearly crosssections.
To provide a general indication of whether or not there is evidence of significant
clustering of growth rates for each 5year period, we use Moran's I statistic (Moran,
1948), which is defined as:
i j wij ( g i M )( g j M )
I [1]
i (gi M )2
where gi and gj denote the growth rates of countries i and j respectively, wij the
corresponding weight in the neighborhood weights matrix, and M the mean rate of
growth in the sample. Table 1 shows that, for W3 (RTA definition of neighborhoods),
Moran's I is positive for all subperiods, with the exception of 19952000. This indicates
the presence of positive global spatial autocorrelation.21 Furthermore, using a
permutation based approach to inference22, this spatial autocorrelation is significant at the
21
A possible explanation for the negative Moran's I statistic for 19952000 might rest with the impacts of
the 1997 Asian financial crisis.
22
For a discussion of different methods of inference see Anselin (1992, p 133135). In implementing the
permutation approach, we used 999 permutations.
15
1 % level for all subperiods between 1975 and 1995. This provides evidence of
clustering of similar growth rates across countries in the same RTA, consistent with the
general presence of localized growth spillover effects.
[table 1 about here]
Although Moran's I provides an indication of the general presence of clustering, it yields
no insight into the possible existence of spatial heterogeneity in this process across major
world regions. To explore the presence of geographically defined subgroups, we first
construct a scatter plot of W3y against y where y is an n 1 vector of observations on
country growth rates expressed in deviations from the sample mean. Such a scatter plot
divides countries into four categories. These categories correspond to fast growing
countries with fastgrowing neighbors (HH); slow growing countries with slowgrowing
neighbors (LL); fastgrowing countries with slowgrowing neighbors (HL); and slow
growing countries with fastgrowing neighbors (LH). Mapping these categories then
provides a visual impression of the possible spatial heterogeneity in growth clustering.23
Figure 2, which relates to 19901995, is typical of the results obtained. It shows several
clusters of countries which shared fast growth relative to the sample mean. Notably, these
clusters seem to be associated with regions of the world with higher levels of formal and
informal integration (in particular, the USACanada, Europe, South Asia, and East Asia
and Pacific). By contrast, SSA has more of a patchwork appearance with notable
incidences of fast growing countries sharing RTAs with slowgrowing countries. This
23
Such a map is referred to as a Moran scatter plot map (Anselin, 1996).
16
indicates a greater propensity of growth rates across neighboring SSA countries to be
independent of each other than in other major parts of the world or across the group of
advanced industrialized countries as a whole. Prima facie, this supports the hypothesis
that SSA is characterized by a relative absence of growth spillovers on account of the
region's lack of integration, both as a result of an absence of effective formal agreements
and of inadequate levels of development of transportation and telecommunications
infrastructure.24
Figure 2: Moran scatterplot map
(Economic growth: average annual logarithmic growth rate of real GDP per capita
growth rate, 19901995; neighborhood definition: belonging to the same RTA)
24
Figure 2 provides no indication as to the statistical significance of the various clusters. It does not allow
us, therefore, to distinguish between whether, for example, the spatial clustering of fast growth rates
observed in East Asia and the Pacific (EAP) reflects genuine spillover effects or could have occurred
simply as a result of some random spatial process. Local Moran statistics do, however, allow us to
comment on statistical significance. For 19901995, use of these allows us to reject the hypothesis of a
random spatial distribution of growth rates not only for the EAP region, but also for the USACanada, parts
of South America and parts of SSA (for a discussion of local Moran statistics and associated approaches to
inference see Anselin, 1995). Full results are available upon request from the authors.
17
Results using W1 (contiguity definition of neighborhood) and W2 (distance definition)
are, in general, very similar to those reported above. The main difference is for the period
19701975 for W1, for which Moran's I is statistically significant at conventional levels.
Therefore, on a contiguity definition of neighborhood, there is no evidence of clustering
for half of the subperiods.
5. CrossCountry Growth Spillovers and their Spatial Heterogeneity
5.1. Model specification
While the above results are suggestive, we cannot be certain whether the observed
patterns of clustering are attributable to genuine spillover effectsor their absence in the
case of SSAor the existence of crosscountry variations in policies, institutions and
physical geography, or even the existence of shared transitional dynamics. For instance,
even in a world of complete autarky, the Solow (1956) model predicts that neighboring
countries with similar policies and institutions will exhibit similar growth rates if they
startoff with similar initial levels of income per capita. In this section, we test whether
any crosscountry correlation of growth rates remains after controlling for observable
determinants of growth, as well as for unobservable timeinvariant determinants, using
data for 19702000.25 This allows us to examine whether there is evidence of spatial
heterogeneity in the strength of crosscountry spillover effects which might be related to
varying levels of integration.
25
The estimator we use is Elhorst's (2003) maximum likelihood (ML) estimator for a panel data model with
FEs and a spatially lagged dependent variable.
18
Expressed in matrix and stacked (in crosssections by time period) form, the basic
estimating equation, which we apply to both our global sample and our various sub
samples, is:
g = (T) + X + (ITW)g + [2]
with E[] = 0 and E['] = 2INT
where g is a NT 1 vector of country growth rates; is a N 1 vector of countryspecific
timeinvariant FEs; X is a NT k matrix of observations on k exogenous control
variables; and IT and INT are identity matrices of dimensions T T and N T
respectively. is a k 1 vector of parameter coefficients and is a NT 1 vector of
disturbance terms. The primary parameter of interest, however, is . This is because the
multiplication of the matrix (ITW) with the vector g yields a NT 1 vector of weighted
average growth rates, where the growth rates being averaged are those of a country's
neighbors. As such, is a scalar parameter which captures the strength of localized
crosscountry growth spillovers.
By controlling for country specific, timeinvariant, determinants of growth which might
otherwise be correlated with our observable independent variables, the above
specification follows what, since Islam (1995), has been standard practice in much of the
empirical growth literature. In this sense, our estimation approach represents an
improvement over many of the previous related studies, discussed in section 2, which
rely on purely crosssectional spatial models.
19
In estimating equation [2], we specify a relatively parsimonious set of control variables
which appear regularly in the standard empirical growth literature. Specifically, the
control variables we include are, firstly, the standard controls suggested by the Solow
model (Mankiw et al, 1992): namely, Aver(I/Y) (the share of investment in real GDP),
Pop. growth (the mean logarithmic growth rate of population) and log(GDP per
capita)initial (the log initial level of real GDP per capita). We also include Aver(G/Y) (the
share of government expenditure in real GDP), Openness (exports plus imports as a
proportion of real GDP) and Civil war (the number of years in each 5year period
characterized by civil war). By restricting ourselves to a relatively parsimonious set of
controls, we are able to maximize N and, in particular, to sample as many neighbors of
each country as possible in the specification of W1 and W3, which is desirable from the
viewpoint of correctly inferring the magnitude of crosscountry spillover effects. Overall,
our exact crosssectional sample size varies according to the W matrix used. However,
for W3, on which we mostly focus, N = 131. This is considerably more than any of the
studies employing longrun growth data surveyed in section 2, for which N is invariably
less than 100.26
It is worth noting that, in including control variables and country FEs in our regressions,
we are assuming that any spatial autocorrelation in the policy and institutions which they
capture are not themselves, in part, a manifestation of growth spillover effects. As a
26
We also experimented with the inclusion of a measure of human capital. This, however, dramatically
reduced N, making estimation for our various subsamples unreliable. For SSA, we also experimented with
additional controls relating to resource richness and the number of years in each 5year period a country
had been free of the various policy syndromes discussed in Collier and O'Connell (2007). Inclusion of these
covariates did not materially affect any of the results presented below. Finally, we experimented with a
sample period of 19602000. Again, because it dramatically reduced sample sizes, this made estimation for
our various subsamples unreliable and, hence, we do not report the results.
20
consequence, the estimates of spillover effects which we report are probably on the
conservative side.27
5.2. Results
For W3, table 2 starts by presenting results using three different estimatorsa pooled
OLS estimator which excludes all country FEs; a standard withingroup (WG) estimator
which eliminates country FEs through first differencing, but which fails to control for the
endogeneity of the neighbor growth variable (Wy); and our preferred ML estimator
which allows for both country FEs and explicitly takes account of the endogeneity of Wy.
Using pooled OLS leads to an estimated crosscountry growth spillover coefficient, ,
^
which is both large in absolute terms and highly significant. In particular, = 0.4569
^
indicates that an increase of 1 % in the weighted average growth rate of neighboring
countries generates a 0.46 % increase in the domestic growth rate. This is similar to
estimates from previous studies based on either the application of purely crosssectional
spatial estimators or the application of IV estimation to pooled data. For example,
Easterly and Levine (1998) obtain an equivalent estimate of of 0.55 based on a smaller
sample of countries using pooled data for 19601990. Including FEs and using the WG
estimator more than halves to 0.2083 without completely eliminating its statistical
^
significance, while also controlling for the endogeneity of Wy using an appropriate
estimator, removes all evidence of a significant crosscountry growth spillover effects in
global data.
27
In particular, the estimates we report correspond, in the terminology of Moreno and Trehan (1997), to
estimates of the strength of net growth spillover effects.
21
[table 2 about here]
Having found no evidence of significant spillover effects using global data, we now turn
to the question of the possible heterogeneity of such effects across various geographically
and incomedefined subsamples. In doing so, the main distinction that we draw is
between the OECD countries, which are fully globally integrated with each other,
countries belonging to SSA, between which levels of integration are low, and the
countries in the rest of the world (RoW). Results for these three subsamples are
presented in table 3. For completeness, we also report results for various other
geographically and incomedefined subsamples which together comprise the RoW sub
sample, although these are invariably insignificant on account of the small sample sizes.28
For SSA, 0 and we cannot reject the hypothesis of an absence of crosscountry
^
spillovers. Meanwhile, for the RoW, is somewhat larger, but still insignificant at
^
conventional levels. For the OECD, however, there is evidence of significant cross
country growth spillovers, at least at the 10 % level.29 In particular, indicates that a 1
^
% increase in the weighted average of neighbor growth rates increases an OECD
country's domestic growth rate by 0.20 %.
[table 3 about here]
28
The definition of regions corresponds to that used by the World Bank
(http://web.worldbank.org/WBSITE/EXTERNAL/COUNTRIES/0,,pagePK:180619~theSitePK:136917,00.
html). These regions are EAP (East Asia & Pacific), ECA (Europe & Central Asia, excluding Western
Europe), LAC (Latin America & the Caribbean), MENA (Middle East & North Africa), OHIE (Other High
Income Economies, i.e. excluding the OECD countries) and SAS (South Asia).
29
This is based on a twotailed test in which the alternative hypothesis makes no distinction between
positive and negative spillovers. In a onetailed test in which the null is 0 and the alternative > 0, the
estimated spillover effect would be significant at the 5 % level.
22
The above results are consistent with Collier and O'Connell's (2007) hypothesis that
spillovers of growth are likely to be absent between SSA countries on account of the
region's lack of integration. By contrast, while we do not find evidence of significant
spillovers using our global sample, significant spillovers do exist between OECD
countries, which are highly integrated both with each other and within their own
geographic regions. The lack of spillover effects in SSA comes despite the region's
"spaghetti bowl" of RTAs, which are incorporated into W3. Our analysis suggests that,
as they stand, these agreements have proved ineffective at promoting growth spillovers
within the region; this likely stems not only from deficiencies in these agreements and
their application, but also from a lack of regional integration emanating from inadequate
levels of transport and telecommunications infrastructure development.
Given our estimates of for different subsamples, it is also possible to calculate the size
of the associated neighborhood multipliers. If we consider any one of the six 5year cross
sections in our sample, then, ignoring the country FEs for notational convenience:
gt = Xt + Wgt + t [3]
where gt is a N 1 vector of observed growth rates for period t (t = 197075, ..., 1995
2000), Xt is a N k matrix of observations on the k controls for period t, and t is the
corresponding N 1 vector of disturbance terms.
23
Providing 0 and 1/ is not an eigenvalue of W, it follows:
gt = (IN  W) 1(Xt) + (IN  W) 1t [4]
where IN is a N N identity matrix.
Equation [4] tells us that, given 0, a country's growth rate not only depends on the
observed values of the control variables for the country itself, but also on those of all
other countries. Likewise, not only do domestic innovations in the disturbance term
matter for growth, but so too do innovations in all other countries. All of these effects are
captured by the inverse N N matrix (IN  W)1. This is a matrix of neighbor multiplier
effects. If we think of a set of policy changes which are pursued in tandem and succeed in
directly raising the growth rate of each country by 1 % then, from this, it follows that,
provided > 0, the final increase in the growth rate of each country will be given by 1/(1
 ) % > 1 %. It follows that M 1 /(1 ) gives the estimated neighborhood multiplier.
^ ^
Table 3 reports this for both our global sample and each of our subsamples. Whereas,
for the OECD, M 1.245 , which implies that coordinated policy actions to raise growth
^
in the OECD will be amplified by 25 % through the feedback effects associated with
growth spillovers, M 1 for SSA. Therefore, contra Easterly and Levine, our results
^
suggest that SSA countries cannot obtain enhanced growth benefits from acting in unison
relative to acting alone. Rather, to obtain such benefits, they first need to cultivate
appropriate channels for spillover effects through pursuing policies to promote more
meaningful integration.
24
To conclude this section, we note how varies when we use the neighborhood weights
^
matrices W1 and W2 instead of W3 (table 4). The results for W2 are very similar to those
for W3, except that, for the OECD countries, is now significant at the 5 % level. By
^
contrast, using W1 (shared borders) dramatically reduces for the OECD countries to
^
0.1170, which is statistically insignificant at conventional levels. This is not too
surprising. After all, membership of the OECD is based not on a country's geographical
region, but on its level of development. Based on these results we can hypothesize that
the primary mechanisms driving spillovers between the OECD countries are likely to be
related to trade.
[table 4 about here]
6. The Role of Infrastructure and Geographic Location
Earlier in this paper we outlined Collier and O'Connell's (2007) finding, based on short
run growth data, that, whereas globally, resource poor landlocked countries are more
dependent on the growth of their neighbors, the opposite is true for such countries in
SSA. This is significant because, according to Collier and O'Connell, spillovers of growth
from their neighbors represent the best hope for development for SSA's landlocked
countries, assuming that these neighbors eventually succeed in takingoff. In this section,
we investigate the interrelationships between the strength of longerrun growth spillovers
experienced by a country, whether or not the country is landlocked, and the country's
25
level of transport and telecommunications infrastructure development.30 This analysis is
motivated by the hypothesis that a country's effective integration into both its own region
and the wider world economy will depend not only on its participation in formal trade
agreements, but also on its accumulated level of investment in infrastructure that
facilitates trade and other interaction. Although this applies for all economies, this is
likely to be especially true for landlocked countries.
Infrastructure data comes from the World Bank's Development Data Platform (DDP).
Following Limão and Venables (2001), we use four indicators of a country's level of
infrastructure development: (1) the density of roads (i.e. number of km of road per km2 of
land area); (2) the density of paved roads (km of paved road per km2 of land area); (3) the
density of railways (km of rail per km2 of land area); and (4) the number of telephone
main lines per capita.31 We combine these four indicators into a single measure of a
country's infrastructure development by first standardizing the observations on each
indicator32 and then taking the simple unweighted mean of the nonmissing observations
across the four indicators for each country.33 Negative values of the resultant index are
associated with levels of infrastructure which are low by global standards, reflecting,
inter alia, the existence of limited road and rail networks. By contrast, positive values
30
For brevity, we simply refer to transport and telecommunications infrastructure as infrastructure in the
remainder of this section.
31
For each country, these four indicators are themselves measured by their mean values over the sample
period (we ignore missing values in the calculation of means). This raises some endogeneity concerns as a
result of possible reverse causation from a country's rate of growth to its level of infrastructure. However,
the results that we report in table 5 below remain essentially unchanged if we instead use startofperiod
(i.e. 1992) values for the four indicators in the construction of Infra.
32
For each observation i on the infrastructure indicator I, we standardize by applying the formula Si = (Ii
M)/s where S denotes the standardized value, M the sample mean across observations on that indicator, and
s the corresponding sample standard deviation.
33
This is equivalent to assuming that the four different types of infrastructure enter as perfect substitutes to
a transport services production function (Limão and Venables, 2001, p 472).
26
reflect levels of infrastructure which are high by global standards. Because
comprehensive coverage of infrastructure data in the DDP only starts in the late
1980s/early 1990s, our analysis is restricted to a crosssectional sample which covers the
period 19922000. Although this rules out the use of panel data techniques, it does have
the advantage of allowing us to further expand our sample, for W3, to 143 countries.
Notably, we are now able to include the majority of nations which comprise the former
Soviet Union. Many of these countries are landlocked. Indeed, while SSA has the greatest
number of landlocked countries of all World Bank regions, ECA has the highest
proportion (World Bank, 2008, p 101).
The regressions which we estimate take a similar form to those in section 5. In particular,
we regress the annual average logarithmic growth rate of real GDP per capita on our
neighbor growth variable (Wy) and a set of controls, again focusing on W3. This set of
controls includes not only those that were considered in section 5, but also dummy
variables for whether or not a country is landlocked (LL), whether or not a country could
be classified as resource rich in 1992 (RR92) and whether or not a country not classified as
resource rich in 1992 became resource rich during the sampleperiod (RRnew). In addition,
we include our measure of infrastructure development (Infra) as a control, both by itself
and interacted with LL. However, with the exception of Infra, our primary interest is not
so much with these extra controls, as with the various interaction effects with Wy. Thus,
we interact Wy with LL and/or Infra in various specifications.
27
Table 5 reports our results for two samples of countries. The first is our full sample of
142 countries (specifications 1a5a), whereas our second excludes Equatorial Guinea
(specifications 1b5b). With Equatorial Guinea included, there is no evidence of
significant interaction effects involving Wy. By contrast, excluding Equatorial Guinea
does yield significant interaction effects in several of our specifications. We prefer the
results excluding Equatorial Guinea. This is because Equatorial Guinea is an outlier and
exhibits extreme leverage on the relationship between Wy and y (where y is the vector of
growth rates). In this relationship, not only does Equatorial Guinea have a value of
Cook's d statistic of 1.329134, but it also has by far the highest DFFITS score (1.7422) in
the sample.35 During the sampleperiod, Equatorial Guinea experienced an extremely
high average annual growth rate of real GDP per capita (almost 15 %), while several of
the countries (namely, the Republic of Congo, Gabon and Chad) with which it shares an
RTA experienced absolute declines in real GDP per capita. Equatorial Guinea's fast
growth, however, was unrelated to the decline of these countries. Rather, it was a
consequence of extremely large discoveries of oil reserves in 1996. Although the
inclusion of RRnew as a control was intended to capture the average impact on growth of
resource discoveries, in the case of Equatorial Guinea, the impact was so large as to
warrant the country's exclusion from the sample.
[table 5 about here]
34
^
Cook's d statistic measures the normalized change in the vector of fitted values, y , attributable to the
deletion of the corresponding observation. Values of d > 1 are normally considered extreme. Equatorial
Guinea is the only country in the sample for which d > 1.
35
The secondhighest DFFITS score in the sample is 0.3831 (Uzbekistan).
28
Concentrating, therefore, on the results for specifications (1b5b), we see, first of all, that
Infra has no statistically significant direct role in determining a country's growth rate and
this is true for both coastal and landlocked countries (1b). Likewise, without allowing for
interaction effects, there is no evidence of significant crosscountry spillovers of growth
(2b). However, this changes in 3b when we interact Wy with LL. According to this
specification, for coastal countries, a 1 % increase in the weighted average growth rate of
their RTA neighbors generates a 0.72 % increase in the domestic growth rate, an effect
which is significant at the 10 % level. By contrast, for landlocked countries, this effect is
more than offset by the negative estimated coefficient on LL*Wy. Indeed, for these
countries, the implied estimate of the spillover coefficient, , is negative.
Simply interacting LL with Wy, however, allows for no distinction between landlocked
countries in SSA, and, to a lesser extent, Central Asia, which have very poorly developed
transportation and telecommunications networks, and landlocked countries such as
Austria and Switzerland, which are in the heart of Europe and which have excellent
networks. Specification 4b, therefore, interacts Wy with both LL and Infra. Estimation
of this specification replicates the result of a growth spillover effect for coastal countries
which is significant at the 10 % level. Specifically, for such countries, a 1 % increase in
Wy is now estimated to increase the domestic growth rate by 0.65 %. However, the
positive, and significant at the 5 % level, estimated coefficient on LL*Infra*Wy indicates
that landlocked countries whose levels of infrastructure are higher (lower) than the global
average, experience a stronger (weaker) growth spillover effect than this. Indeed, from
the results of 4b, we can derive an estimated spillover coefficient, i , for each
^
29
landlocked country in our sample. Figure 3 plots these estimated coefficients as a
function of Infra. It shows very high i for Luxembourg (LUX), Switzerland (CHE) and
^
Austria (AUT) on account of their sophisticated transportation and telecommunications
networks. Hungary (HUN) and Slovakia (SVK), landlocked countries which have both
recently joined the EU, and, in the case of Slovakia, the Eurozone, also have i in excess
^
of that estimated for coastal countries (i.e. i > 0.65). Macedonia (MKD), Moldova
^
(MDA) and Uzbekistan (UZB), by contrast, have i which are somewhat below that
^
estimated for coastal countries. Finally, the i for SSA's landlocked countries, which are
^
characterized by very poor transportation and telecommunications networks, are all
clustered around zero. Interestingly, the interaction between infrastructure and spillovers
is conditional on controlling for whether or not a country is landlocked. Without this
distinction, Infra has no significant influence on the strength of spillovers (5b).
30
Figure 3: Estimated countryspecific spillover coefficients as a function of the level of
transport and telecommunications infrastructure development, landlocked countries only
Note: figure based on results from col. 4b of table 5
The above results show that the importance of infrastructure lies not in its direct
contribution to economic growth, but in the benefits it brings to landlocked countries in
their ability to experience and absorb beneficial growth spillovers from neighboring
countries. It is, therefore, investment in such infrastructure that, along with more
formalized trading agreements, has helped to integrate countries such as Switzerland and
Austria into their neighborhoods and the global economy, and which differentiates them
from the landlocked countries of, in particular, SSA. These results are consistent with
Collier and O'Connell's (2007) hypothesis that, globally, landlocked countries can be
expected to be more dependent on the growth of their neighbors than coastal countries,
with the exception of SSA where regional integration is low.
31
On a note of caution, however, the results reported in table 5 are based on OLS
estimation. This is problematic given the inherent endogeneity of Wy. Ideally, we would
have adopted an estimation approach which explicitly controls for such endogeneity.
However, standard crosssectional ML estimators which allow for spatial effects
(Anselin, 1988), are unable to allow for interaction effects involving Wy. Likewise,
although we experimented with the use of various instruments for the variables in
specifications (2a/b)(5a/b) which involve Wy, these experiments proved unsatisfactory.
In particular, we experimented with using "spatial lags" of the control variables (i.e. WX,
where X is the matrix of observations on the controls) and their interactions with LL and
Infra as instruments, as well as with using the values of Wy from 19841992. However,
the resultant instruments proved to be very weak. In the case of the instruments based on
WX, this was because the controls themselves have disappointing explanatory power
(see, for example, the R2 values for specifications (1b)(5b) in table 5).
We also experimented with using an expanded set of controls, at the expense of sample
size, but this did not improve matters because the expanded set did not much improve the
fit of our regressions.36 Meanwhile, in the case of the instruments based on the temporal
lag of Wy, their weakness can be explained by the fact that, globally, medium to long
term growth rates contain a strong transitory element (Easterly, 2009, p 122) which
36
In particular, we experimented extensively with an expanded set of controls including all of the variables
which SalaiMartin et al. (2004; see table 2, p. 284) report as "significantly related to growth" for the
period 19601996. That this expanded set of controls did not help to improve the fit of our regressions and,
therefore, the strength of our instruments based on WX, is demonstrated by the fact that the adjusted R2 in a
regression of growth on just the controls was actually lower for this expanded set of variables (0.1828) than
for the equivalent specifications reported in columns 1a and 1b of table 5. This seemingly paradoxical
result can be explained by the reduction in the sample size to 99 countries caused by the use of the
expanded set of controls. The full set of results is available upon request from the authors.
32
implies that growth rates in the 1980s are poor predictors of growth rates in the 1990s,
thereby also making the temporal lag of Wy a poor predictor of Wy. Not only did the
instruments that we experimented with prove to be unsatisfactory on account of their
weakness, but also because they sometimes led to theoretically implausible estimates of
, the coefficient on Wy. In particular, a spatially stable growth process requires  < 1.
However, values of 1 were obtained for some of the specifications when using IV
^
estimation.37
Notwithstanding the fact the above results are based on OLS, we are reasonably confident
that our main conclusions are not driven by endogeneity of Wy. This is so for several
reasons. Firstly, as Behar (2008, p 7) argues, Wy is only endogenous under the
hypothesis of growth spillovers. Therefore, tests of the rejection of the null of no
spillovers based on OLS estimation retain some validity. Second, as noted above, when
entered in our specifications by itself without any interaction effects, the coefficient on
Wy is not significantly different from zero (specification 2a, table 5). It is only when Wy
enters in more subtle forms that we detect significant spillover effects. Third, and finally,
when we reestimate the simple spillover specification, 2b, with no interaction effects
using a ML estimator which does explicitly take into account the simultaneity of y and
Wy, we find that the estimated spillover coefficient is actually larger than that reported in
37
Our estimates of i for Luxembourg, Switzerland, Austria and Hungary implied by specification 4b in
table 5 also fall outside of this range (see figure 3). However, this does not necessarily imply a spatially
unstable growth process for these countries because they have amongst their neighbors nonlandlocked
countries for which  i  1 . Thus, while faster growth of these countries' neighbors is amplified when
^
spillingover to Luxembourg, Switzerland, Austria and Hungary, the reverse feedback to the neighbors is
then damped. This leads to the possibility of a spatially convergent growth process, even if it appears
locally unstable in places.
33
table 5. Therefore, in this instance, it seems that, if anything, the use of OLS leads us to
under, rather over, estimate the strength of growth spillover effects.38
7. The Costs of Fragmentation for SubSaharan Africa's Landlocked Countries
Having provided evidence of heterogeneous spillover effects across landlocked countries
and related these to differences in the strength of integration, in this section we present
the results of a simulation exercise which is designed to be suggestive of the welfare
losses associated with a lack of integration for such countries. This exercise answers the
hypothetical question: What would have been the cumulative loss in real GDP over the
period 19702000 had Switzerland, a landlocked country which is fully integrated with
both its immediate neighborhood and the world economy, been subject to spillovers of
the strength that the Central African Republic experienced? Thus, our exercise is akin to
the thought experiment of relocating Switzerlandwith all its domestic human and
physical capitalfrom the heart of Europe so that it takes the Central African Republic's
place in the heart of SSA.
To implement this exercise, we draw on our results from section 5 and make a highly
conservative set of assumptions. Hence, we assume that the only parameter which
changes from Switzerland's viewpoint is , i.e. the strength of the crosscountry spillover
effect. From the results of table 3, therefore, we assume that Switzerland shared the
estimated value of of 0.0430 with the rest of SSA rather than the value of 0.2350
estimated for it as part of the OECD subsample. Apart from this, however, we assume
38
The results from the application of this ML estimator are available upon request from the authors.
34
that everything else for Switzerland remains unchanged. Thus, the change in
neighborhood is assumed not to impact on any of Switzerland's observable or
unobservable determinants of growth over the period 19702000.39 Furthermore, we
assume that the impacts of the observed determinants of growth for Switzerland remain
as estimated for the OECD sample. Finally, we assume no change in the underlying
pattern of shocks experienced by Switzerland over the period 19702000.
More concretely, our simulation methodology comprises of five steps. In step 1 we
calculate Switzerland's new growth rate of real GDP per capita, yCHE, for the period 1970
1975 given its change in neighborhood. In particular:
nSSA
yCHE , SSA
1970 75 CHE ,OECD x197075 OECD SSA wCHEj y1970,j75 u1970,OECD
^ CHE ^ ^ SSA CHE
75 [5]
j 1
where CHE,OECD is the size of Switzerland's FE as estimated using the OECD sub
^
sample, x197075 is the 1 k row vector of observations on the control variables for
CHE
Switzerland for 19701975, OECD is the corresponding k 1 column vector of
^
parameters estimated using the OECD subsample, w
CHEj y1970,j75 is the weighted
SSA
average growth rate for Switzerland's neighbors in SSA now that it has taken the place of
the Central African Republic in W3, SSA is the estimated value of the spillover
^
CHE
parameter from our original SSA subsample, and u1970,OECD is the estimated residual for
75
Switzerland for 19701975 using the OECD subsample.
39
With the exception of log(GDP per capita)initial (see below).
35
Having calculated yCHE for 19701975, in step 2 we update the growth rates for all of the
other countries in SSA for 19701975. This is necessary because these countries are now,
either directly or indirectly, connected to Switzerland, instead of the Central African
Republic, through W3. Step 3 then involves iterating steps 1 and 2 until convergence
between successive iterations in each element of the vector of SSA country growth rates,
including Switzerland in the place of the Central African Republic, is achieved.40 In step
4 we calculate log(GDP per capita)1975 for both Switzerland and all other countries in
SSA. This is required because the value of this control variable in one subperiod is
endogenous to growth in the previous period. Finally, in step 5, we repeat steps 1 4 for
all subsequent time periods (i.e. for 19751980, ..., 19952000).
Figure 4 shows the results. In 1970, Switzerland's real GDP per capita in the
counterfactual simulation is the same as its actual observed level. However, as time
progresses, an everwidening shortfall of simulated GDP per capita below its observed
level emerges as a result of the weaker spillover effects. By 2000, Switzerland's real GDP
per capita is 9.3 % lower under the counterfactual. Cumulating the losses over 1970
2000 gives an aggregate real GDP loss of $334 billion (2000 international dollars), which
was the equivalent of 162 % of Switzerland's real GDP in 2000.
40
Convergence is assumed to have occurred when the absolute difference between each country's growth
rate in successive iterations is less than 0.001 %.
36
Figure 4: Simulating the impact on Switzerland's real GDP per capita of weaker growth
spillovers associated with its hypothetical relocation to the centre of subSaharan Africa
29
Actual
28 Counterfactual
GDP per capita (thous; cnst 2000 international dollars)
27
26
25
24
23
22
21
20
1970 1975 1980 1985 1990 1995 2000
Year
Although the welfare gains for the Central African Republic from greater spillovers
would obviously be lower in absolute terms than Switzerland's simulated losses, it is
clear that the welfare losses associated with weak crosscountry spillovers stemming
from a lack of integration are very large for landlocked countries. Indeed, if anything,
our simple exercise considerably underestimates the losses. This is because of the highly
conservative assumptions which underpin the exercise.
8. Conclusion
In this paper, we have examined the strength of crosscountry spillovers of longterm
growth both globally and in various geographically and income defined subsamples.
Our objective was to find evidence for the possible spatial heterogeneity of such effects
which can be linked to differences in the integration of countries, both with their
immediate neighborhoods and globally. We have further investigated the relationship
37
between the strength of any growth spillover effect, landlockedness, and the level of
development of a country's transport and telecommunications networks. In doing so, we
were motivated by the observation that a country's ability to benefit from spillovers is
likely to depend not only on its participation in formal trade agreements, but also on the
level of development of such networks and this is especially true for landlocked
countries. Indeed, in the case of SSA, the development of such infrastructure is likely to
be a critical prerequisite for cultivating beneficial growth spillovers. This is because there
already exists a "spaghetti bowl" of RTAs which, in themselves, have proved to be
largely ineffective.
Overall, our panel data results provide moderate evidence in favor of the existence of
heterogeneous growth spillover effects for the period 19702000. In particular, while
such crosscountry spillovers have been a significant determinant of growth for OECD
countries at the 10 % level, we cannot reject the hypothesis that spillovers are absent in
SSA countries. This seems consistent with the high level of integration that exists
between the OECD countries and the lack of effectiveas compared to pro forma
integration observed within SSA. Furthermore, our crosssectional analysis for 1992
2000 suggests that, globally, coastal economies experience, on average, stronger growth
spillover effects than landlocked countries. This result, however, is attributable to the
fact that most landlocked countries are located in SSA and, as such, are characterized by
very poorly developed transport and telecommunications networks. Once we allow the
level of development of such networks to interact with our neighboring growth variable
for landlocked countries, we uncover a dichotomy of experiences.
38
On the one hand, landlocked countries such as Luxembourg, Switzerland, Austria and
Hungary, which are in the heart of Europe, experience stronger spillovers of growth from
their neighbors than the average coastal country on account of their high levels of
transport and telecommunications infrastructure. On the other hand, the landlocked
countries of SSA, not to mention Central Asia, experience essentially no beneficial
growth spillovers from their neighbors. This is because of, inter alia, inadequate
investments in transport and telecommunications infrastructure accumulated over long
periods of time. Our hypothetical simulation exercise of allowing Switzerland to take the
place of the Central African Republic in SSA demonstrates that the welfare losses
associated with missing out on such beneficial spillovers are substantial.
The above conclusions support and extend previous arguments and findings made by
Collier and O'Connell (2007). They are less consistent with those of Easterly and Levine
(1998) who have partly attributed SSA's growth failure to reinforcing spillovers of slow
growth. Such arguments seem inconsistent not only with our findings, but also with the
fact that some countries in the region, such as Botswana, have experienced fast growth,
while growth in neighboring countries has floundered. This casts doubt on the notion that
a coordinated stimulus across SSA might benefit from a multiplier effect such that the
overall impact on longterm economic growth far outweighs the direct initial impact on
each country's domestic growth. Rather, our results suggest that more effective
integration involving, in particular, investments in transport and telecommunications are
39
first required to generate the transmission mechanism for such a multiplier effect. This is
particularly true for the region's landlocked countries.
40
Data appendix
This appendix details the different sources of data used to construct the various variables
used in the analysis of this paper.
Nonspatial variables
Real GDP per capita, Pop. Growth, Aver(I/Y), Aver(G/Y), Openness
The underlying data is from the Penn World Table v 6.2 (Heston et al., 2006) and was
downloaded from http://pwt.econ.upenn.edu/php_site/pwt_index.php. Pop. growth is the
(natural) logarithmic growth rate of population and was calculated as [ln(Pi,t)  ln(Pi,tT)]/T
where Pi,t denotes country i's population level in year t and T is the number of years over
which the population growth rate is calculated. Aver(I/Y), Aver(G/Y) and Openness are
all measured as averages over the period of interest, with the level of Openness in any
one year being given by (Xi,t + Mi,t)/Yi,t where Xi,t denotes country i's level of exports in
year t, Mi,t its level of imports and Yi,t its level of real GDP. Real GDP is measured in
2000 constant international dollars.
Civil war
Measured as the number of years in a given time period for which a country was
characterized by civil war, where civil war is itself defined as an intrastate conflict
involving 1000+ annual battle deaths. The underlying data is from the International
Peace Research Institute, Oslo, and was downloaded from http://www.prio.no/CSCW/
Datasets/.
LL
A dummy variable indicating whether or not a country can be classified as landlocked (1
= landlocked, 0 = not landlocked). Both the Democratic Republic of Congo and Sudan
are classified as landlocked, despite the fact that they have coasts. This is on the basis of
their lack of access to their coast lines. Ethiopia is classified as being landlocked
throughout the 19702000 sampleperiod. The data is from the data appendix of Collier
and O'Connell (2007), supplemented with information from Wikipedia.
RR92, RRnew
Dummy variables indicating respectively whether or not a country could be classified as
resource rich in 1992 (1 = resource rich, 0 = not resource rich) and whether or not a
country for which RR92 = 1 became resource rich during the sampleperiod 19922000 (1
= became resource rich, 0 = did not become resource rich). The variables are constructed
from data contained within the data appendix of Collier and O'Connell (2007) which
gives the first year in which they classify a country as being resource rich based on
several criteria, supplemented, in some cases, with judgemental adjustments.
Infra
The construction of this variable is explained in section 6 and follows Limão and
Venables (2001), as well as Bosker and Garretson (2008). The underlying data is from
the World Bank's Development Data Platform.
41
Spatial variables
As explained in detail in section 3, three different neighborhood weights matrices
(referred to in the main text as W1, W2 and W3) were used to construct variables
measuring the weighted average growth rate of a country's neighbors. The contiguity
data used to construct W1 is from the Correlates of War (COW) v 3.1 direct contiguity
dataset and was downloaded from www.correlatesofwar.org/COW2%20Data/dDDirect
Contiguity/DCV3desc.htm. Meanwhile, the weights in W2 are based on the weighted
average of the bilateral distances between cities of population greater than 100,000 in
countries i and j. The population data used to identify these cities is from the Global
Rural Urban Mapping Project (GRUMP) database, Center for International Earth Science
Information Network (CIESIN), the Earth Institute, Columbia University. This data was
downloaded from http://sedac.ciesin.columbia.edu/gpw/. Distances between cities are
measured using the great circle method. Finally, the regional trade agreement (RTA)
data used to construct W3 was kindly supplied by Souylemane Coulibaly and was
originally sourced from the World Trade Organization. The RTAs included in W3 are
listed in table A1 below.
Table A1: List of Regional Trade Agreements (RTAs) included in the definition of W3
Association of Southeast Asian Nations (ASEAN), Arab Maghreb Union (AMU), Arab FreeTrade
Area (ArFTA), Australia New Zealand Closer Economic Agreement (ANZCERTA), AsiaPacific
Economic Cooperation (APEC), Baltic FreeTrade Area (BAFTA), Bangkok agreement (BANGKOK),
Bay of Bengal Initiative for MultiSectoral Technical and Economic Cooperation (BIMSTEC),
Cooperation Council for the Arab States of the Gul f (C CASG), Central Ame rican C ommon
Market (C ACM), Andean Communi ty (CAN), Caribbean Community and Common Market
(CARICOM), Central Europe an FreeT rade Agreem ent (CEFTA) , Economic and Monetar y
Community o f Cen tral A frica (CE MAC), Commonwealth of Independent States (CIS), Common
Market for Eastern and So uthern Africa (COME SA), East African Community (EAC), Eurasian
Economic Community (EAEC), European Union (EU), European Cooperation Organisation (ECO),
Economic Community of West A frican States (ECOWAS), European Eco nomic Area (EEA),
European Free Trade Association (EFTA), General System of Trade Preferences among Developing
Countries (GSTP), Latin American Integration Association (LAIA), Southern Common Marke t
(MERCOSUR), Melanesian Spearhead Group (MSG), North Amer ican FreeTrade Agree ment
(NAFTA), Overseas Countries and Territories (OCT), Agreement on Trade and Commercial Relations
between the Government of Australia and the Government of Papua New Guinea (PATCRA), Protocol
Relating to Trade Negotiations among Developing Countries (PTN), South Asian Association for
Regional Cooperation (SAARC), Southern Africa Customs Union (SACU), Southern African
Development Com munity (SADC) , SAARC Pre ferential Trad ing Agree ment (SAPTA) , South
Pacific Regional Trade and Economic Cooper ation Agree ment (SP ARTECA), Tripartite
Agreement (TRIPARTITE), West African Economic and Monetary Union (WAEMU)
The agreements listed correspond to those which have been notified to GATT/the WTO and which
were in force as of 13 October 2003. RTAs which are considered to be more prominent are in bold.
42
References
Abreu, Maria, Henri L.F. de Groot and Raymond J.G.M. Florax, 2004, "Spatial Patterns
of Technology Diffusion", Tinbergen Institute Discussion Paper TI2004079/3.
Ades, Alberto and Hak B. Chua, 1997, "Thy Neighbor's Curse: Regional Instability and
Economic Growth", Journal of Economic Growth, 2(3), 279: 304.
Angeriz, Alvaro, John S.L. McCombie and Mark Roberts, 2008, "New Estimates of
Returns to Scale and Spatial Spillovers for EU Regional Manufacturing, 1986
2002", International Regional Science Review, 31(1), 6287.
Angeriz, Alvaro, John S.L. McCombie and Mark Roberts, 2009, " Increasing Returns and
the Growth of Industries in the EU Regions: Paradoxes and Conumdrums",
Spatial Economic Analysis, 4(2), 127148
Anselin, Luc, 1988, Spatial Econometrics: Methods and Models, Dordrecht: Kluwer
Academic Publishers.
Anselin, Luc, 1992, SpaceStat Tutorial: A Workbook for Using SpaceStat in the Analysis
of Spatial Data , Technical Software Series S921, National Center for
Geographic Information and Analysis (NCGIA), University of California, Santa
Barbara, 1992.
Anselin, Luc, 1995, "Local Indicators of Spatial Association  LISA", Geographical
Analysis, 27(2): 93115.
Anselin, Luc, 1996, "The Moran Scatterplot as an ESDA Tool to Assess Local Instability
in Spatial Association" in Fischer, Manfred, Henk J. Scholten and David Unwin
(eds.), Spatial Analytical Perspectives on GIS, London: Taylor & Francis.
Arbia, Giuseppe and Gianfranco Piras, 2005, "Convergence in PerCapita GDP across
European Regions using Panel Data Models Extended to Spatial Autocorrelation
Effects", Instituto di Studi e Analisi Economica Working Paper No. 51.
Armstrong, Harvey W., 1995a, "An appraisal of the evidence from crosssectional
analysis of the regional growth process within the European Union", in Harvey
W. Armstrong and Roger W. Vickerman (eds), Convergence and Divergence
Among European Regions, vol. 5, London: Pion.
Armstrong, Harvey W., 1995b, "Convergence Among Regions of the European Union,
19501990", Papers in Regional Science, 74(2): 143152.
Behar, Alberto, 2008, "Neighborhood growth effects: an annual panel data approach",
background paper prepared for the 2009 World Development Report, Reshaping
Economic Geography, Centre for the Study of African Economies, University of
Oxford, Oxford.
Bernat, Andrew J., 1996, "Does manufacturing matter? A spatial econometric view of
Kaldor's laws", Journal of Regional Science, 36(3): 463477.
Bosker, Maartin 2007, "Growth, agglomeration and convergence, a spacetime analysis
for European regions", Spatial Economic Analysis, 2(1): 91100.
43
Bosker, Maartin and Harry Garretsen, 2008, "Economic geography and economic
development in subSaharan Africa", mimeo.
Brakman, Steven, Harry Garretsen and Charles Van Marrewjik, 2009, The New
Introduction to Geographical Economics, Cambridge: Cambridge University
Press.
Buys, Piet, Uwe Deichmann and David Wheeler, forthcoming, "Road network upgrading
and overland trade expansion in subSaharan Africa ", Journal of African
Economics.
Caselli, Francesco, Gerardo Esquivel and Fernando Lefort, 1996, "Reopening the
convergence debate: a new look at crosscountry growth empirics", Journal of
Economic Growth, 1(3): 13814338.
Coe, David T. and Elhanan Helpman, 1995, "International R&D spillovers", European
Economic Review, 39(5): 859887.
Collier, Paul and Stephen A. O'Connell, 2007, "Growth and opportunities", in Benno J.
Ndulu, Robert H. Bates, Paul Collier and Charles C. Soludo (eds.), The Political
Economy of Economic Growth in Africa, 19602000, Cambridge: Cambridge
University Press
Conley, T.G. and Ethan Ligon, 2002, "Economic distance and crosscountry spillovers",
Journal of Economic Growth, 7(2): 15737020.
Easterly, William, 2009, "The Indomitable in Pursuit of the Inexplicable: The World
Development Reports' Failure to Comprehend Economic Growth Despite
Determined Attempts, 19782008" in Shahid Yusuf (ed.), Development economics
through the decades: a critical look at 30 years of the world development report,
Washington, DC: The World Bank.
Easterly, William and Ross Levine, 1998, Troubles with the Neighbors: Africa's
Problem, Africa's Opportunity, Journal of African Economies, 7(1): 120142.
Elhorst, J. Paul, 2003, "Specification and Estimation of Spatial Panel Data Models",
International Regional Science Review, 26(3): 244268.
Ertur, Wilfried and Cem Koch, 2005, "The Role of Human Capital and Technological
Interdependence in Growth and Convergence Processes: International Evidence",
Laboratoire d'Economie et de Gestion.
Ertur, Wilfried and Cem Koch, 2007, "Growth, Technological Interdependence and
Spatial Externalities: Theory and Evidence", Journal of Applied Econometrics,
22(6): 10331062.
Fingleton, Bernard and John S.L. McCombie, 1998, "Increasing returns and economic
growth: some evidence for manufacturing from the European Union regions",
Oxford Economic Papers, 50(1): 89105.
Grossman, Gene M. and Elhanan Helpman, 1991, Innovation and Growth in the Global
Economy, Cambridge, Mass.: MIT Press.
44
Heston, Alan, Robert Summers and Bettina Aten, 2006, Penn World Table Version 6.2,
Center for International Comparisons of Production, Income and Prices at the
University of Pennsylvania.
Islam, Nazrul, 1995, "Growth Empirics: A Panel Data Approach", Quarterly Journal of
Economics, 110(4): 11271170.
Limão, Nuno and Anthony J. Venables, 2001, "Infrastructure, Geographical
Disadvantage, Transport Costs and Trade", World Bank Economic Review, 15(3):
451479;
Mankiw, N. Gregory, David Romer and David N. Weil, 1992, "A Contribution to the
Empirics of Economic Growth", Quarterly Journal of Economics, 107(2): 407
437.
Moran, P.A.P., 1948, "The interpretation of statistical maps", Journal of the Royal
Statistical Society B, 10: 24351.
Moreno, Ramon and Bharat Trehan, 1997, "Location and the Growth of Nations",
Journal of Economic Growth, 2(4): 399418.
Rey, Sergio J. and Brett D. Montouri, 1998, "US regional income convergence: a spatial
econometrics perspective", Regional Studies, 33(2): 143156.
Roberts, Mark, 2004, "The Growth Performances of the GB Counties: Some New
Empirical Evidence for 19771993", Regional Studies, 38(2): 149165.
SalaiMartin, Xavier, Gernot Doppelhofer and Ronald I. Miller, 2004, "Determinants of
LongTerm Growth: A Bayesian Averaging of Classical Estimates (BACE)
Approach", American Economic Review, 94(4): 813835
Stinnett, Douglas M., Jaroslav Tir, Philip Schafer, Paul F. Diehl, and Charles Gochman,
2002, "The Correlates of War Project Direct Contiguity Data, Version 3", Conflict
Mangagement and Peace Science, 19(2):5866.
Solow, Robert M., 1956, "A contribution to the theory of economic growth", Quarterly
Journal of Economics, 70(1): 6594.
Weinhold, Diana, 2002, "The Importance of Trade and Geography in the Pattern of
Spatial Dependence of Growth Rates", Review of Development Economics, 6(3):
36982.
World Bank, 2008, World Development Report 2009: Reshaping Economic Geography,
Washington, DC: The World Bank.
45
Tables
Table 1: Moran's I test results for global spatial autocorrelation (W3 = rowstandardized RTA matrix)
Period Mor an's I Mean Zvalue (prob)
19701975 0.2969 0.2641 1.3418 (0.1797)
19751980 0.2492*** 0.1747 2.7219 (0.0065)
19801985 0.1261*** 0.0169 3.8288 (0.0001)
19851990 0.2190*** 0.0899 4.7558 (0.0000)
19901995 0.0904*** 0.0068 3.1678 (0.0015)
19952000 0.1538 0.1877 1.3435 (0.1791)
*** indicates significance at the 1 % level
Inference is based on a permutation approach to inference (see Anselin, 1992, p 133135). 999
permutations were used.
46
Table 2: Results from estimation of spillover model for global sample: Spatial versus
nonspatial methods, 19702000 (W3 = rowstandardized RTA matrix)
Variable Pooled OLS Spatial FE SpatialML
log(GDP per 0.0079 0.0630 0.0646
capita)initial (5.6129) (12.3311) (13.5756)
Pop. growth 0.1287 0.0173 0.0163
(1.2821) (0.1277) (0.1325)
Aver(I/Y) 0.1398 0.0735 0.0729
(7.6365) (2.3576) (2.5538)
Aver(G/Y) 0.0160 0.0499 0.0594
(1.08) (1.5466) (2.0589)
Openness 0.0130 0.0458 0.0465
(3.9492) (5.9707) (6.6667)
Civil war 0.0026 0.0044 0.0045
(2.8101) (3.1309) (3.5582)
Wy 0.4569 0.2083 0.0700
(5.4217) (2.3494) (0.9624)
Spatial multiplier 1.8414 1.2631 1.0309
^
naive
R2 0.1683 0.2637 0.4429
R2 0.1608  0.3252
N 131 131 131
NT 786 786 786
Spatial FE corresponds to fixed effects (withingroup) estimator; spatialML to Elhorst's
(2003) spatial fixed effects panel data estimator; in the case of Spatial FE, R2 excludes effect of
spatial fixed effects on model fit; Figures in brackets are asymptotic tstatistics; Bold indicates
significance at the 10 % level; bold and italics significance at the 5 % level
47
Table 3: Results from estimation of dynamic panel spillover model, 19702000 (W3 = rowstandardized RTA matrix)
Variable Gl obal OECD SSA RoW Other groupings
EAP EC A LAC MENA OHIE SAS
log(GDP per 0.0646 0.03988 0.0828 0.0673 0.0532 0.1327 0.1066 0.1163 0.0994 0.0116
capita)initial (13.5756) (6.59240) (7.6604) (10.6408) (4.2122) (7.9345) (8.3078) (4.9193) (7.1910) (1.1045)
Pop. growth 0.0163 0.42442 0.5213 0.3602 0.5068 2.7633 1.2089 0.6508 0.0434 0.9924
(0.1325) (1.38112) (2.2824) (2.1760) (0.4816) (2.2235) (2.0104) (0.6697) (0.1755) (2.0890)
Aver(I/Y) 0.0729 0.057831 0.0842 0.0155 0.1928 0.2794 0.2422 0.2958 0.2350 0.3219
(2.5538) (1.73223) (1.8270) (0.3044) (1.5486) (3.3451) (2.7986) (2.2091) (1.7595) (1.6522)
Aver(G/Y) 0.0594 0.12698 0.0369 0.1018 0.0233 0.0091 0.0702 0.0892 0.1538 0.0179
(2.0589) (1.88543) (0.5946) (2.7449) (0.2348) (0.0663) (1.7085) (0.8488) (1.5709) (0.1191)
Openness 0.0465 0.07472 0.0494 0.0385 0.0203 0.0712 0.0469 0.0658 0.05160 0.1014
(6.6667) (7.043378) (3.6360) (4.0360) (1.0083) (3.0688) (2.6144) (2.3431) (2.6909) (2.3676)
Civil war 0.0045  0.0083 0.0009 0.0010 0.0019 0.0037 0.0017  0.0014
(3.5582) (3.6251) (0.5348) (0.2277) (0.5810) (1.4879) (0.5081) (0.5476)
Wy 0.0700 0.20000 0.0170 0.0940 0.0481 0.1619 0.0269 0.0300 0.0711 0.0169
(0.9624) (1.79340) (0.1278) (1.0355) (0.2674) (1.6039) (0.2261) (0.1593) (0.7166) (0.0823)
Spatial multiplier 1.0309 1.24500 1.0173 1.1038 1.0505 0.8607 0.9738 1.0309 1.0765 0.9834
^
naive
3.5914 2.62280 4.1595 3.6840 3.1798 5.3511 4.7815 5.0056 4.6063 0.9940
R2 0.4429 0.6277 0.4616 0.4526 0.5528 0.8775 0.5175 0.4864 0.5582 0.7563
R2 0.3252 0.5330 0.3343 0.3304 0.3944 0.7834 0.3845 0.2686 0.4271 0.6432
log(LIK) 1594.8341 459.9361 484.2164 780.881 144.5984 75.3253 297.7167 99.5591 150.1743 100.9331
2
^ 0.0010 0.0001 0.0013 0.0011 0.0007 0.0001 0.0005 0.0009 0.0016 0.0005
N 131 24 42 65 11 4 21 8 14 7
NT 786 144 252 390 66 24 126 48 84 42
All specifications include countryspecific time invariant fixed effects and were estimated using the Elhorst (2003) spatial fixed effects panel data estimator;
Figures in brackets are asymptotic tstatistics; Bold indicates significance at the 10 % level; bold and italics significance at the 5 % level
48
Table 4: Comparison of the estimated spillover coefficient across different W matrices
W3 (RTA) W2 (Distance) W1 (Contiguity)
Global 0.0700 0.0480 0.0510
OECD 0.20000 0.2200 0.1171
SSA 0.0170 0.0430 0.0380
RoW 0.0940 0.0330 0.0641
Bold indicates significance at the 10 % level; bold and italics significance at the 5 % level
49
Table 5: OLS results from estimation of hybrid crosssectional spillover model, 19922000 (W3 = rowstandardized RTA matrix)
Fullsample (N = 142) Excluding Equatorial Guinea (N = 141)
1a 2a 3a 4a 5a 1b 2b 3b 4b 5b
Constant 0.0720*** 0.0726*** 0.0718*** 0.0723*** 0.0734*** 0.0730*** 0.0726*** 0.0677*** 0.0683*** 0.0764***
(2.8694) (2.9266) (2.9200) (2.9208) (2.8556) (2.8819) (2.8253) (2.6219) (2.6211) (2.9316)
LL  0.0151*** 0.0133 0.0158** 0.0150*** 0.0139*** 0.0135** 0.0062 0.0209*** 0.0129**
0.0151*** (3.0959) (0.7171) (2.4016) (3.0979) (2.7616) (2.5965) (0.5825) (3.5548) (2.4309)
(3.0607)
RR92 0.0036 0.0033 0.0035 0.0036 0.0035 0.0035 0.0032 0.0039 0.0035
0.0034 (0.5665) (0.5057) (0.5548) (0.5681) (0.5797) (0.5272) (0.4928) (0.5892) (0.5370)
(0.5173)
RRnew 0.0666*** 0.0667*** 0.0663*** 0.0673*** 0.0396*** 0.0413*** 0.0360*** 0.0335*** 0.0425***
0.0751*** (2.8785) (2.7296) (2.7988) (2.8218) (4.6853) (4.6289) (4.1401) (3.8277) (4.6276)
(3.1754)
L(GDP pc)92 0.0088*** 0.0090*** 0.0089*** 0.0088*** 0.0094*** 0.0097*** 0.0104*** 0.0103*** 0.0098***
 (3.0537) (3.0610) (3.0341) (3.0323) (3.2647) (3.3929) (3.5893) (3.5385) (3.4372)
0.0095***
Pop. growth (3.3784) 0.2549 0.2448 0.2528 0.2525 0.2826 0.2870 0.2758 0.2968 0.2735
(1.1121) (1.0609) (1.0952) (1.1011) (1.2545) (1.2884) (1.2584) (1.3546) (1.2089)
0.2681
Aver(I/Y) (1.1930) 0.1884*** 0.1898*** 0.1818*** 0.1889*** 0.1691*** 0.1590*** 0.1531*** 0.1505*** 0.1618***
(5.3994) (5.3903) (5.3923) (5.3865) (4.8859) (4.3015) (3.9972) (3.9149) (4.3807)
0.1788***
Aver(G/Y) (5.2914) 0.0178 0.0187 0.0182 0.0182 0.0162 0.0140 0.0162 0.0171 0.0155
(0.6834) (0.7348) (0.7088) (0.6878) (0.6251) (0.5390) (0.6550) (0.6729) (0.5821)
0.0185
Openness (0.7035) 0.0043 0.0044 0.0044 0.0042 0.0049 0.0060 0.0073 0.0073 0.0052
(0.7389) (0.7560) (0.7436) (0.7142) (0.8849) (1.0649) (1.3139) (1.3055) (0.9050)
0.0055
Civil war (0.9846) 0.0008 0.0008 0.0008 0.0008 0.0007 0.0007 0.0009 0.0008 0.0008
(0.9453) (1.0283) (0.9704) (0.9551) (0.8652) (0.8999) (1.1241) (1.0226) (0.9428)
0.0008
Infra (1.0129) 0.0038 0.0036 0.0037 0.0056 0.0038 0.0033 0.0017 0.0020 0.0114
(1.1265) (1.0461) (1.1034) (0.4478) (1.1826) (0.9961) (0.5075) (0.5991) (0.9604)
0.0031
LL*Infra (0.9546) 0.0085 0.0103 0.0051 0.0086 0.0065 0.0053 0.0138** 0.0300 0.0059
50
(1.3524) (1.5024) (0.1710) (1.3536) (1.0373) (0.8110) (2.2290) (1.6084) (0.8959)
0.0079
Wy (1.2592) 0.3358 0.2647 0.3182 0.3564  0.2333 0.72 01* 0.6503* 0.0994
(0.8793) (0.5883) (0.7397) (0.8917) (0.8103) (1.8318) (1.6761) (0.3065)

LL*Wy  0.2453      1.0373**  
(0.3874) (2.3285)

LL*Infra*Wy   0.1367     1.4640** 
(0.1177) (2.1174)

Infra*Wy 0.0675 0.3035
(0.1523) (0.7329)
naive
^ (% pa) 0.9156 0.8516 0.8405 0.8405 0.8547 0.9034 0.9382 0.9990 0.9870 0.9444
R2 0.3758 0.3811 0.3811 0.3811 0.3811 0.2843 0.2891 0.3161 0.3102 0.2878
0.3230 0.3235 0.3183 0.3183 0.3126 0.2232 0.2224 0.2461 0.2396 0.2149
R2
2
^ 0.0006 0.0006 0.0006 0.0006 0.0006 0.0006 0.0006 0.0006 0.0006 0.0006
Moran I stat. 0.8016 0.1601 0.4286 0.4286 0.2184 0.0438 0.9141 0.5608 0.5528 0.8564
(pvalue) (0.4228) (0.8728) (0.6682) (0.6682) (0.8272) (0.9651) (0.3607) (0.5749) (0.5804) (0.3918)
Figures in brackets are tstatistics based on White heteroscedasticity consistent standard errors; * indicates significance at the 10 % level, ** significance at the 5
% level, *** significance at the 1 % level
51