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 Sub-Saharan 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 Sub-Saharan 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 Sub-Saharan Africa. The authors through which such spillovers can spread--most define neighborhoods both in geographic terms and importantly infrastructure endowments. by membership in the same regional trade association. This paper--a product of the Environment and Energy Team, Development Research Group--is 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 E-mail: mr10013@cam.ac.uk Uwe Deichmann Development Research Group The World Bank 1818 H Street, NW Washington, DC 20433, USA E-mail: 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 6-7th 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 contagion-like 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 Sub-Saharan 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 cross-country 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 so-called 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 well-defined geographic regions. With localized cross-country 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 136-137). 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 cross-country 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 intra-regional trade2, trade between SSA countries is negligible: about one-half 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 take-off, 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 Intra-regional 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 1990-2000. 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 cross-country spillover effects. Due to data limitations we employ two samples of countries. The first, "long sample" consists of 131 countries for the period 1970-2000. 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 time-invariant 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 sub-samples 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 1992-2000. 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 sample-period, we need to rely on cross-sectional, rather than panel, techniques. 4 Also, in the absence of well-developed 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 cross-country 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 cross-country spillovers using both the full sample of countries for the period 1970-2000 and various geographically and income based sub-samples. 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 1992-2000 sample. Section 7 presents our hypothetical simulation exercise. Finally, section 8 concludes. 2. Literature on Cross-Country Growth Spillovers A large literature models growth between sub-national 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 cross-sectional 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 steady-state 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 so-called 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 cross-sectional growth regression when all other explanatory variables are excluded (included) (ibid., p. 405). Using data for 1965-1989, they find evidence that both types of spillover are extremely strong. Other studies which use cross-sectional 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 two-sector 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 cross-sectional data, they use average growth rates for three 7 pooled 10-year 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 136-137). All of the above studies are concerned with long-run growth spillovers. There also exist several studies which instead focus on short-run 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 time-invariant 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 cross-country spillover effects might help to explain SSA's poor growth performance over recent decades, Collier and O'Connell offer a competing hypothesis. Namely, that cross-country 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 longer-run 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 cross-country 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 non-parametric 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 row-standardizing 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 sample-period, rather than before it, such an agreement usually embodies long-standing trading relationships which predate both the formal RTA and the beginning of the sample-period. 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, pre-1991 and pre-1981 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 1970-2000 sample-period. 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 1962-1979. 18 A related endogeneity concern is the existence of countries which withdrew from a RTA during the sample-period 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 rule-out the possibility of the dissolution of RTAs and of decisions to withdraw from RTAs having been endogenous to the growth process over the sample-period. Again, however, in many cases where RTAs have been dissolved over the sample-period, 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 sample-period 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 Cross-Country 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 sample-period of 1970-2000 on which the regression analysis of section 5 focuses, splitting this into 5-yearly cross-sections. To provide a general indication of whether or not there is evidence of significant clustering of growth rates for each 5-year 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 sub-periods, with the exception of 1995-2000. 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 1995-2000 might rest with the impacts of the 1997 Asian financial crisis. 22 For a discussion of different methods of inference see Anselin (1992, p 133-135). In implementing the permutation approach, we used 999 permutations. 15 1 % level for all sub-periods 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 fast-growing neighbors (HH); slow growing countries with slow-growing neighbors (LL); fast-growing countries with slow-growing neighbors (HL); and slow- growing countries with fast-growing neighbors (LH). Mapping these categories then provides a visual impression of the possible spatial heterogeneity in growth clustering.23 Figure 2, which relates to 1990-1995, 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 USA-Canada, 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 slow-growing 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, 1990-1995; 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 1990-1995, 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 USA-Canada, 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 1970-1975 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 sub-periods. 5. Cross-Country 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 effects--or their absence in the case of SSA--or the existence of cross-country 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 start-off with similar initial levels of income per capita. In this section, we test whether any cross-country correlation of growth rates remains after controlling for observable determinants of growth, as well as for unobservable time-invariant determinants, using data for 1970-2000.25 This allows us to examine whether there is evidence of spatial heterogeneity in the strength of cross-country 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 cross-sections 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 country-specific time-invariant 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 cross-country growth spillovers. By controlling for country specific, time-invariant, 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 cross-sectional 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 5-year 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 cross-country spillover effects. Overall, our exact cross-sectional 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 long-run 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 sub-samples unreliable. For SSA, we also experimented with additional controls relating to resource richness and the number of years in each 5-year period a country had been free of the various policy syndromes discussed in Collier and O'Connell (2007). Inclusion of these co-variates did not materially affect any of the results presented below. Finally, we experimented with a sample period of 1960-2000. Again, because it dramatically reduced sample sizes, this made estimation for our various sub-samples 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 estimators--a pooled OLS estimator which excludes all country FEs; a standard within-group (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 cross-country 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 cross-sectional 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 1960-1990. 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 cross-country 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 income-defined sub-samples. 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 sub-samples are presented in table 3. For completeness, we also report results for various other geographically and income-defined sub-samples 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 cross-country ^ 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 two-tailed test in which the alternative hypothesis makes no distinction between positive and negative spillovers. In a one-tailed 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 sub-samples, it is also possible to calculate the size of the associated neighborhood multipliers. If we consider any one of the six 5-year 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 = 1970-75, ..., 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 sub-samples. 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 taking-off. In this section, we investigate the interrelationships between the strength of longer-run 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 un-weighted mean of the non-missing 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 start-of-period (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 cross-sectional sample which covers the period 1992-2000. 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 sample-period (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 1a-5a), whereas our second excludes Equatorial Guinea (specifications 1b-5b). 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 sample-period, 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 second-highest DFFITS score in the sample is -0.3831 (Uzbekistan). 28 Concentrating, therefore, on the results for specifications (1b-5b), 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 cross-country 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 country-specific 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 cross-sectional 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 1984-1992. 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 Sala-i-Martin et al. (2004; see table 2, p. 284) report as "significantly related to growth" for the period 1960-1996. 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 co-efficient 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 re-estimate 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 non-landlocked countries for which | i | 1 . Thus, while faster growth of these countries' neighbors is amplified when ^ spilling-over 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 Sub-Saharan 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 1970-2000 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 Switzerland--with all its domestic human and physical capital--from 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 cross-country 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 sub-sample. 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 1970-2000.39 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 1970-2000. 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 1970-1975, OECD is the corresponding k 1 column vector of ^ parameters estimated using the OECD sub-sample, 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 sub-sample, and u1970,OECD is the estimated residual for 75 Switzerland for 1970-1975 using the OECD sub-sample. 39 With the exception of log(GDP per capita)initial (see below). 35 Having calculated yCHE for 1970-1975, in step 2 we update the growth rates for all of the other countries in SSA for 1970-1975. 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 sub-period 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 1975-1980, ..., 1995-2000). 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 ever-widening 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 sub-Saharan 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 cross-country 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 cross-country spillovers of long-term growth both globally and in various geographically and income defined sub-samples. 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 1970-2000. In particular, while such cross-country 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 effective--as compared to pro forma-- integration observed within SSA. Furthermore, our cross-sectional 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 long-term 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. Non-spatial 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,t-T)]/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 intra-state 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 1970-2000 sample-period. 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 sample-period 1992-2000 (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 Free-Trade Area (ArFTA), Australia New Zealand Closer Economic Agreement (ANZCERTA), Asia-Pacific Economic Cooperation (APEC), Baltic Free-Trade Area (BAFTA), Bangkok agreement (BANGKOK), Bay of Bengal Initiative for Multi-Sectoral 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 Free-T 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 Free-Trade 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 TI2004-079/3. 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), 62-87. 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), 127-148 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 S-92-1, 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 Per-Capita 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 cross-sectional 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, 1950-1990", Papers in Regional Science, 74(2): 143-152. 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): 463-477. Bosker, Maartin 2007, "Growth, agglomeration and convergence, a space-time analysis for European regions", Spatial Economic Analysis, 2(1): 91-100. 43 Bosker, Maartin and Harry Garretsen, 2008, "Economic geography and economic development in sub-Saharan 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 sub-Saharan Africa ", Journal of African Economics. Caselli, Francesco, Gerardo Esquivel and Fernando Lefort, 1996, "Reopening the convergence debate: a new look at cross-country growth empirics", Journal of Economic Growth, 1(3): 1381-4338. Coe, David T. and Elhanan Helpman, 1995, "International R&D spillovers", European Economic Review, 39(5): 859-887. 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 cross-country spillovers", Journal of Economic Growth, 7(2): 1573-7020. Easterly, William, 2009, "The Indomitable in Pursuit of the Inexplicable: The World Development Reports' Failure to Comprehend Economic Growth Despite Determined Attempts, 1978-2008" 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): 120-142. Elhorst, J. Paul, 2003, "Specification and Estimation of Spatial Panel Data Models", International Regional Science Review, 26(3): 244-268. 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): 1033-1062. 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): 1127-1170. Limão, Nuno and Anthony J. Venables, 2001, "Infrastructure, Geographical Disadvantage, Transport Costs and Trade", World Bank Economic Review, 15(3): 451-479; 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: 243-51. Moreno, Ramon and Bharat Trehan, 1997, "Location and the Growth of Nations", Journal of Economic Growth, 2(4): 399-418. Rey, Sergio J. and Brett D. Montouri, 1998, "US regional income convergence: a spatial econometrics perspective", Regional Studies, 33(2): 143-156. Roberts, Mark, 2004, "The Growth Performances of the GB Counties: Some New Empirical Evidence for 1977-1993", Regional Studies, 38(2): 149-165. Sala-i-Martin, Xavier, Gernot Doppelhofer and Ronald I. Miller, 2004, "Determinants of Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach", American Economic Review, 94(4): 813-835 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):58-66. 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): 369-82. 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 = row-standardized RTA matrix) Period Mor an's I Mean Z-value (prob) 1970-1975 0.2969 0.2641 1.3418 (0.1797) 1975-1980 0.2492*** 0.1747 2.7219 (0.0065) 1980-1985 0.1261*** 0.0169 3.8288 (0.0001) 1985-1990 0.2190*** 0.0899 4.7558 (0.0000) 1990-1995 0.0904*** 0.0068 3.1678 (0.0015) 1995-2000 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 133-135). 999 permutations were used. 46 Table 2: Results from estimation of spillover model for global sample: Spatial versus non-spatial methods, 1970-2000 (W3 = row-standardized RTA matrix) Variable Pooled OLS Spatial FE Spatial-ML 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 (within-group) estimator; spatial-ML 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 t-statistics; 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, 1970-2000 (W3 = row-standardized 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 country-specific time invariant fixed effects and were estimated using the Elhorst (2003) spatial fixed effects panel data estimator; Figures in brackets are asymptotic t-statistics; 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 cross-sectional spillover model, 1992-2000 (W3 = row-standardized RTA matrix) Full-sample (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 (p-value) (0.4228) (0.8728) (0.6682) (0.6682) (0.8272) (0.9651) (0.3607) (0.5749) (0.5804) (0.3918) Figures in brackets are t-statistics based on White heteroscedasticity consistent standard errors; * indicates significance at the 10 % level, ** significance at the 5 % level, *** significance at the 1 % level 51