CONFLICT, ECONOMIC GROWTH AND SPATIAL SPILLOVER EFFECTS IN AFRICA

Transcription

1 CONFLICT, ECONOMIC GROWTH AND SPATIAL SPILLOVER EFFECTS IN AFRICA J Paul Dunne School of Economics and SALDRU, University of Cape Town John.Dunne@uct.ac.za Nan Tian School of Economics University of Cape Town Nan.Tian@uct.ac.za October 2013 Abstract: While there is a burgeoning literature in the determinants of conflict, there has been much less attention given to the economic effects. There is a general consensus that conflict, particularly civil conflict will have negative impacts on the economy and this was supported by Collier (1999). Murdoch and Sandler (2002a) considered the likely spillover effects of conflict on growth, finding clear negative effects on primary neighbours, followed by de Groot (2010) who introduced secondary neighbour effects into the analysis and finding that there was in fact a significant positive effect for these countries. This paper updates and develops this analysis, confirming the negative sign and magnitude of the previous findings for the host country and primary neighbours, but finding no evidence of any positive effects of conflict on secondary neighbours. Considering more comprehensive measures of spillover using conflict duration gave results suggesting that the mere presence or absence of conflict, and not its duration are the significant determinants of spillover effects resulting from conflict. Keywords: Conflict; Economic growth; Spatial econometrics; Africa JEL code: C21; F51; H56; O11 Preliminary Draft: Comments and suggestions are welcome. Please do not quote without author s approval.

2 INTRODUCTION Over the past half-century and especially with the end of the Cold-War, the goal of a less conflict-ridden world has failed to emerge as civil wars have replaced international wars as the most prevalent form of large-scale violence and destruction. While civil wars have been the norm in post-cold War Africa, certain interstate wars such as the war in the Democratic Republic of Congo (DRC) do exist. It has been agreed that the types of wars fought in Africa range depending on various reasons. Intrastate wars often stem from ethnic hatreds that manifest through nationalism, separatism, or a fight for ethnic diversity (James and Goetze, 2001). In other situations, civil wars may be rooted in the greed or grievance debate (Collier and Hoeffler, 1998; 2004). Regarding interstate conflicts, these have emerged under a completely new appearance as compared to the old wars of the past. Wars are no longer being fought between two equal armies or capable states, but increasingly between a substantially stronger attacker and a fragile or collapsed defender (Hendriks, 2012). The war in the DRC can be categorised as a new war but is also fundamentally an interstate war. Following the First Congo War which ended in 1997, the second war could be argued by many as a war of opportunity. Weak and collapsed states such as the DRC became easy targets by both internal and external rebels and neighbouring states. Regardless of whether the wars fought have been interstate or intrastate, they have profound consequences on economic growth. While vast studies agree that the impact of conflict on economic growth is negative, there are opposite findings when one assesses the short and long-term effects. Research by the likes of Collier (1999) suggest conflict to have an unequivocal negative impact on growth, but Organski and Kugler (1980) argue that in the long-run, through the phoenix effect, there exists little evidence that conflict negatively affects growth. Moreover, conflict not only influences growth of home economies but also neighbours countries from factors such as diversion of foreign direct investment (FDI), disruption of trade, displacement of people, destruction of physical and human capital and the reallocation of resources towards less productive activities. This issue, however, has received little attention. The relationship between conflict in one country and economic growth in neighbouring or nearby countries has been sparingly researched; with only Murdoch and Sandler (2002a, 2002b, 2004) and De Groot (2010) addressing the spillover effects of conflict on growth in contiguous states. This paper, thus, contributes to the existing literature by continuing this investigation with a particular focus on a new empirical estimation strategy for Africa between the periods 1960 to The following section discusses the contributions to the literature by Murdoch and Sandler and De Groot and provides a potential extension to the existing studies which may yield more realistic results. Section three presents the estimation model which follows closely from that of De Groot (2010). The dataset used for empirical analysis are presented in section four, while the estimation methods and results of the analysis can be found in sections five and six, respectively. The seventh and final section provides some conclusions.

3 LITERATURE REVIEW In extending the literature on the consequences of conflict on economic growth Murdoch and Sandler (2002a) become the first study to explore the spillover effects of civil wars on neighbouring nations. Using a basic Solow growth model and adding domestic and adjacent conflicts, they argue that domestic and neighbouring conflicts are two sources that have a negative effect on the economy. They find that civil wars have significant negative influences on the steady-state level of GDP per capita of both home and neighbouring countries for a worldwide sample of 84 countries during the period 1960 to While Murdoch and Sandler believe that part of the negative civil war effect works on growth through the classical channels of capital and labour, the largest effect is found to come through the unobserved, country specific, channel. Following from Murdoch and Sandler s initial paper, two separate papers were developed, each varying in time period, sample of countries and definition of contiguity. While Murdoch and Sandler s (2002a and 2004) first and third paper used a worldwide dataset, their second paper (2002b) considers geographical heterogeneity between Africa, Asia and Latin America. The time period utilised by Murdoch and Sandler differs between papers as well, following from their initial work, their second and third paper uses a dataset that stretches from 1960 to There are further evolutions between their different papers. While their first paper uses only direct contiguity to measure regional spillover effects (whereby countries are required to share a border to in order to be classified as contiguous), the following papers utilise the Gleditsch and Ward (2001) minimum distance between nations dataset to construct contiguity matrices that take into account whether a country is within a particular distance of closest approach. In each paper, the authors analyse both the short and long-term effects of conflict on growth. Panel data techniques using five-year averages were regressed to find the short-run effects of conflict while cross sectional estimations (averaging over the entire period) were utilised for the long-run regressions. In all three papers, Murdoch and Sandler conclude that the long-run regressions yield no consistent negative growth impact, which they attribute to the phoenix effect. On the other hand, civil wars were found to have a significant negative effect on economic growth in neighbouring countries. While Murdoch and Sandler (2002b and 2004) have differing conclusions on the minimum distance over which civil conflict is felt, their overall conclusion using different datasets and contiguity definitions was the consistent negative growth effect due to civil conflict in both host and neighbouring countries. As De Groot (2010) explain, the papers by Murdoch and Sandler have significant merit since they were the first authors to address the issue of neighbouring countries suffering from spillover effects of civil conflict. It is through their work that other authors are able to further extend the literature on spillover effects and conflict. As a result, De Groot (2010) adds to the literature by continuing Murdoch and Sandler s analysis using an updated dataset on Africa for the period 1960 to De Groot also proposed theoretical changes by distinguishing between primary and secondary neighbours, which are able to estimate multi-dimensional

4 spillover effects 1. Another difference between De Groot (2010) and Murdoch and Sandler s studies is the introduction of a new type of contiguity matrix based on minimum distance weights. With these changes in mind, De Groot (2010) finds that by distinguishing between primary and secondary neighbours the spillover effects is indeed multidimensional. Conflict has a negative spillover effect on primary neighbour growth, while it is positively associated with secondary neighbour growth. Moreover, different from Murdoch and Sandler, De Groot (2010) finds this result to hold in both short and long-run. Within this emerging literature, the contribution of De Groot (2010) has provided increase substance to the importance of assessing the spillover effects resulting from conflict. However, there remain a number of issues that needs to be addressed in order to improve the results of these papers and the understanding within the literature. Firstly, as more data both quality and quantity become available and the updating of the underlying datasets till 2010, there is now more information allowing researchers to revisit earlier analysis. The increase in ten years of data could substantially change the findings of earlier papers and thus warrant a re-examination of the existing literature. Secondly, in addition to the increased data, the consideration of prolonged spillover effects resulting from conflict has been absent within the literature. Both Murdoch and Sandler (2002a, 2002b and 2004) and De Groot (2010) utilise the event of a conflict as a single shock to the host and neighbouring countries. However, as the conflict literature demonstrates, war cannot be seen as an isolated event that has a singular shock on economic growth. Rather, it must be considered as having a prolonged impact on factors such as physical and human capital, trade, FDI and institutions. All these factors will then have an effect on economic growth beyond the initial conflict shock. Another way to think about conflict is through an impulse response function, whereby the key variables of consideration are the duration, direction and magnitude of the conflict on economic growth. Although the Neoclassical growth theory predicts that an economy following a conflict recovers quickly, and converges to its steady state, this recovery is not immediate. Moreover, a report by the United Nations Development Programme (UNDP) show per capita GDP of seven 2 war-afflicted countries failing to reach pre-war GDP levels fifteen years after the official end of the conflict (UNDP, 2008). In econometric terms, the most cited number in terms of conflicts impact on growth is the 2.2 percent of GDP per war year. The issue of consecutive years of war or prolonged wars over several years is another issue that the existing literature, through the use of five-year averages, fails to take into account. In light of the identified short-comings within the literature, this represents an opportunity to fill this research gap through a theoretical improvement. Instead of using five-year averages as done in the literature, this paper plans to use annual data and using conflict duration as an additional measure to account for multiple prolonged conflicts. The use of annual data also allows for the possibility to pick up potential business 1 De Groot (2010) argued that Murdoch and Sandler s theoretical model restricted spillover effects to be unidimensional and lacked the flexibility to estimate bounce back effects that exist between contiguous states. 2 The seven war afflicted countries are Cambodia, Mozambique, Rwanda, Uganda, El Salvador, Guatemala and Nicaragua.

5 cycle effects excluded in pervious analysis. In order to deal with the drawback of historically measuring conflict as a singular shock variable, a measure of conflict duration is added to assess spillover effects of a prolonged conflict. Economic growth of a host country is often considered to have a positive impact on neighbouring economies and the same can be said during an economic downturn or one that is affected by conflict. The loss of 2.2 percent of GDP per war year or the inability after fifteen years of post-war society to reach pre-war GDP levels can most definitely be seen as having a different impact on growth than simply accounting for conflict as a one-time economic shock. THEORETICAL MODEL The basic theoretical model used within the literature is based on Solow s (1956) neoclassical growth model that has been augmented to include human capital (Mankiw et al, 1992). The growth model features a Cobb-Douglas production function for diminishing returns (i.e. decreasing marginal product) in labour (L), physical capital (K) and human capital (H). The model features constant returns to scale and along the steady state growth path, savings equals total investment in physical and human capital. In order to determine the empirical effects of conflict, the Solow model is further augmented to include conflict as a determinant of economic growth. The human capital augmented aggregate neoclassical Cobb-Douglas production function featuring Harrod-neutral technical progress can be written as: 1. ( ) ( ) ( ) ( ) ( ), where and are the elasticities of output with respect to physical and human capital, respectively. Y(t) denotes aggregate real output (income) at time t, K is real physical capital stock, L is labour, H is human capital, A(t) is the technology parameter embodied in labour and ( ) is the output elasticity of effective units of labour A(t)L(t). Labour is assumed to grow at the constant exogenous natural rate of n, technical progress grows at the exogenous rate of g and both physical and human capital depreciates at the identical rate. Given the above identified assumptions, and by dividing both sides of equation (1) by effective labour (AL), gives an expression in terms of income per effective worker (y=y/al) that is equal to: 2. ( ) ( ) ( ) where k=k/al and h=h/al are quantities per effective worker. Given equation (2), in order to complete the model one needs the transition equations for physical and human capital that accounts for the growth of capital per capita ( ) and human capital per capita ( ). By assuming that the equations of motion for labour (L), labouraugmenting technology (A), physical (K) and human capital (H) are:

6 3. ( ) ( ) ( ) ( ) 4. ( ) ( ) ( ) ( ) ( ) ( ) one can derive the transition equations 3 of physical and human capital as: and 5. ( ) ( ) ( ) ( ) 6. ( ) ( ) ( ) ( ) Having found the transition equations the expression for the steady state of physical and human capital per capita, where so that the economy converges to a steady state defined by: 7. ( ) ( ) and 8. ( ) ( ) With these steady state values for and and then substitutingthem into the production function (2), the steady state level of output per effective worker is: 9. ( ) ( ) By taking logs of equation (9) and linearising via a truncated Taylor series expansion around the steady-state, substituting in y* and setting A equal to a constant a, one is able to obtain the growth of income expressed as: 10. ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Empirically, equation (10) can be parameterised following Mankiw et al (1992) in the following way to represent the steady state long-run growth of income per capita as: 3 For full derivation see Mankiw, Romer and Weil (1992).

7 11. ( ) ( ) ( ) ( ) where ln denotes the natural logarithm, gr is the growth rate of income per capita, y0 is the initial level of income per capita, is the investment in physical capital, is level of schooling, n is the growth rate of the working age population, g is the exogenous growth rate of technical progress and is the rate of depreciation. Theoretically, growth in per capita income from the augmented Solow model- shown in equation (11) - depends positively on investment in physical and human capital. However, income per capita growth falls with increases in ( ) or higher initial level of income per capita (y0). An increase in the natural rate of labour growth (n) or labour efficiency (g) raises the denominator of the dependent variable, income per capita and thus reducing its level. Similarly, depreciation of limits income growth through reductions in growth of physical and human capital as this decline must be offset overtime by a positive accumulation. Furthermore, initial level of income per capita has a negative influence on economic growth known as conditional convergence. Barro (1991) stated that if countries similar with respect to structural parameters for preferences and technology, then poorer countries with lower initial income - tend to grow faster than richer countries. This convergence result hinges on the neoclassical growth model s theory of diminishing returns to capital. In the case of equation (11), poorer countries with low ratios of physical and human capital have higher marginal products of their respective capital as compared to richer countries and thereby grow at higher rates. In line with historical literature, the theory of convergence has been crucial in investigating the impact of conflict on economic growth. Following a conflict, a country can be considered as starting from a relatively low level of income per capita, whereby it then catches up through convergence. Organski and Kugler (1980) labelled this the Phoenix effect as countries decimated by war rise up from the ashes to record substantial growth rates in the post-conflict years. Yet, there is several more ways conflict influences growth, especially when considering the influence of conflict in neighbouring countries on economic growth. In defining the different effects resulting from conflict, this paper follows the approach set out by De Groot (2010), whereby he divides countries into three types: host nations, primary neighbours or directly contiguous neighbours and secondary neighbours or nations that are near a conflict nation but not directly contiguous. Given the above, conflict can influence economic growth at home and in neighbouring nations through similar channels theoretical channels. The first channel through which conflict may impact growth is capital. This potential influence takes different effects depending on different types of countries. The primary manner through which conflict and capital can affect growth is through the destruction of physical and human stock. This destruction effect mainly applies to host nations and through collateral damage on primary neighbours. Secondary neighbours, on the other hand, could potentially suffer from decreased

8 or no collateral damage. Essentially, the assumption is that the further a country is from the conflict origin, the less negative influence conflict has and a potential to have positive impacts. The second channel where conflict may impact economic growth is through the diversion of foreign direct investment (FDI) flowing to the region due to heightened perceived risk. Similar to the first channel, the potential negative effect primarily affects host nations and primary neighbours. The influence it has on secondary neighbours is debated since apart from the a potential decreased effect, the increased perceived risk of the host conflict region could lead to relatively more attractive investment opportunities appear in secondary neighbour countries. Channel three through which conflict may influence growth is through the displacement of labour that creates refugees within a country and the outflow of refugee migrants to neighbours. Again, the primary argument of the negative impact on growth similar to the previous two channels (capital and investment). Conflict leads to the destruction and displacement of productive labour and the reassignment of labour to less productive activities (i.e. soldering and border patrol). As a result, host countries and primary neighbours are most likely to suffer from these effects so that income per capita is reduced. On the other hand, secondary neighbours, due to no direct contiguity, may suffer no adverse effect from inflow of refugees or reassignment of labour to less productive activities. Furthermore, it can be argued that for those specific refugees that choose to cross multiple borders, they may carry with them high human capital and this is hypothesised to benefit the secondary neighbours. The fourth channel where conflict can inhibit growth is through trade. In a host country afflicted with conflict, both domestic and international trade is likely to be negatively affected, which directly harms economic growth. This could have a substantial effect on primary neighbours since diversion of trade flows with the host country will also curb economic growth. However, primary neighbours can be seen as having a choice to substitute their trade to the host countries secondary neighbours or their primary neighbour. This could potentially negate or abate the negative loss from trade with the host country and increase trade and economic growth of the secondary neighbour. The fifth channel is the use of resources to either quell local conflicts or bolster defence spending in order to defend borders for conflicts in neighbour countries. In the case of increased defence spending, resources must be diverted from productive activities and hence an opportunity cost (Dunne and Tian, 2013).The effect of reallocation of productive resources can be considered to have a negative impact on the host nation and primary neighbours, but little evidence of spillover effects to secondary neighbours 4. 4 An exception to this case would be if the host country, primary and secondary neighbours are part of a security web. While this case is interesting, it is beyond the scope of this paper and would potentially provide an interesting avenue for future research.

9 Finally, the last channel through which conflict distorts economic growth is the potential spillover effect of conflict itself (Sambanis, 2002; De Groot, 2010). While there is evidence of primary neighbours getting dragged into host country conflicts, the effect on secondary neighbours is minimal, with the exception of the Second Congo war. Overall, one would expect primary or secondary neighbours to experience either a negative or no economic effect resulting from the spillover of conflict itself. Based on the above discussion, there are various potential channels capital, FDI, labour, trade, conflict and productivity through which conflict at home or nearby can affect economic growth. While it is obvious that the host country is likely to experience a negative growth shock as a result of conflict, the impact on primary and secondary neighbours can be considered different. Primary neighbours are expected to experience either negative or no growth shock from a host country s conflict, while secondary neighbours can be deliberated to not suffer from the drawbacks of a host country s conflict but potentially reap certain spillover benefits. It can also be suggested that the above mention effects could become more pronounced once the duration of conflict is taken into account. For estimation purposes equation (11) is augmented to include three conflict elements. Firstly, conflict is added to capture host country effects. Furthermore, two conflict variables are added to pick up conflict experience in primary and secondary neighbour states. In order to account for the spatial spillover effects from neighbouring conflict, this paper follows from De Groot (2010) in setting up the four different types of contiguity matrices 5. With the creation of both the contiguity matrices and conflict variables complete, the growth equation can be written as: 12. ( ) ( ) ( ) ( ) ( ) ( ) ( ) In the above equation, and are the weights matrix of primary and secondary contiguity respectively. The weights are based on different contiguity matrices chosen within the regression. For primary neighbours, the two types of contiguity matrices used are based on direct contiguity. Firstly, dummy variables are chosen which gives a value 1 of countries share a border and 0 if they do not. Secondly, the border length shared between the countries divided by total host country border (a proportion) is used as the matrix element to construct the second type of primary neighbour contiguity matrix. For secondary neighbours, the two types on matrices are used based on the dummy variable approach as well as the minimum distance approach 6.The variable represents the three measures of conflict used within the empirical estimation and is the growth rate of host nations, primary and secondary neighbours. 5 For full description on the creation of the two primary and two secondary contiguity matrices see De Groot (2010). 6 see De Groot (2010).

10 DATA DESCRIPTION The data used to estimate the parameters of the various specifications are taken from four sources. GDP per capita, investment and population are taken from the Penn World Tables version 7.1; schooling is provided by Barro and Lee (2012) and Penn World Tables version 8.0; measures of armed conflict was provided by the UCDP/PRIO Armed Conflict Database 4, which was updated till 2010 by Themner and Wallensteen (2011); and finally, the weighted contiguity matrices was cross-referenced through personal correspondence with Olaf De Groot. The following table provides the variables used within the estimations, the description and sources. Table 1: Variable Names, Descriptions and Sources Variable Description Source gr Real GDP per capita in constant 2005 dollars Penn World Tables 7.1 pop Population in 000's Penn World Tables 7.1 invest Investment as a share of GDP Penn World Tables 7.1 school conflict intense Percentage of secondary education attained in the population older than 25 Conflict dummy, equals to 1 if conflict during the relevant sample period Conflict dummy, equals to 1 if conflict during the relevant sample period led to more than battle related deaths in a particular year Barro and Lee (2010); Penn World Tables 8.0 UCDP/PRIO Armed Conflict Database; Themner and Wallensteen (2011) UCDP/PRIO Armed Conflict Database; Themner and Wallensteen (2011) civil Conflict dummy, equals to 1 if conflict was classified as a civil war during the relevant sample period UCDP/PRIO Armed Conflict Database; Themner and Wallensteen (2011) Although pervious work has focused on dividing the empirical analysis into short and longterm, this paper uses a new approach. Instead of using five-year averages and averages over the entire sample period to measure short and long-run effects respectively, annual frequency data is used for the period 1960 to 2010 and dynamic panel analysis introduced 7.In 2010, there were a total of 53 African countries. However, four countries namely; Eritrea, Libya, Seychelles and Somalia have missing data for Real GDP per capita and investment; thus these were excluded from the sample. The final sample features 49 countries over 51 years with a maximum number of 2499 observations available. The dependent variable, gr, is defined as (ln(y1)-ln(y0))and represents annual growth. A similar approach is used for investment, which represents growth in the annual share of investment. To comply with the augmented Solow model, (g+ ) which is assumed to be equal to is added to population growth to form the term ( ). 7 Further details on the estimation method can be found in the following section titled empirical analysis. 8 The assumption that (g+δ)=0.05 follows from Mankiw et al (1992)

11 For schooling variable, this paper utilises data collected by Barro and Lee (2012), which updates their previous Barro and Lee (2001) work, and Penn World Table 8.0 s index of human capital. By amalgamating the index of human capital - which is measured annually - with the Barro and Lee education database, this paper is to create a schooling or education variable that is still in the form of education attainment in secondary schooling as a percentage of the population over 25, but measuring annually. However, as is the case with African data, the schooling variable is only available for 35 of the 49 countries. In order to check for robustness, the conflict indicator variable, from the UCDP/PRIO database, is split into three different types of conflict. The first conflict variable conflict contains all conflicts recorded in the dataset, the second variable intense, picks up only those conflicts that have at least battle related deaths per year and finally, the variable civil war includes only intrastate conflicts. Conflict duration measured in months of conflict in a calendar year is also used to assess spillover effects from prolonged conflict. To construct the different weight matrices, geographical data on common borders are needed. Using a combination of CIA World Factbook, Gleditsch and Ward (2001) s minimum distance dataset and personal correspondence with Olaf De Groot, the weight matrices are confirmed to be consistent with those used in the De Groot (2010) paper. Table 2, below, provides a summary statistic of the final dataset containing 49 Africa countries between the period 1960 to Table 2:Variable Description and Summary Statistics Variable Variable Description Mean Std. Dev gdp (1) Real GDP per capita invest (1) Investment as a share of GDP school (2)(3) Percentage of secondary education attained in the population older than pop (1) Population in 000's conflict (4) Conflict indicator (1=Conflict, 0=No conflict) intense (4) Intense Conflict indicator (1=Conflict, 0=No conflict) civil (4) Civil war indicator (1=Civil war, 0=No civil war) gdp Growth rate of real per capital GDP (log) invest Growth rate of investment as share of GDP (log) school Growth rate of education attainment (log) n+g+δ Population growth rate (clpop)+0.05 (assumed value for g+d) used in Solow-style regressions Note: (1)Penn World Table 7.1Center for International Comparisons of Production, Income and Prices (CIC), University of Pennsylvania. (2) Penn World Table 8.0Feenstra, Robert C., Robert Inklaar and Marcel P. Timmer (2013), "The Next Generation of the Penn World Table" available for download at (3) Barro, Robert and Jong-Wha Lee, "A New Data Set of Educational Attainmentin the World, " forthcoming, Journal of Development Economics. (4) UCDP/PRIO Armed Conflict Version and Themner, Lotta and Wallensteen, Peter, Armed Conflict, Journal of Peace Research.

12 EMPIRICAL ANALYSIS To study the impact of civil wars on economic growth, the approach taken in this paper consists of estimating Solow style regressions using dynamic panel data methods. The spillover effects of conflict on growth will be analysed separately and together (via interaction variables of conflict, contiguity matrix and economic growth) using different specifications that vary in measures of conflict and contiguity. In undertaking empirical analysis within the field of conflict and economic growth, data limitations are of the highest problem. While countries that have experience war are likely to have the worst data they are also the ones of interest, and although data quality has improved, issues do remain 9.Yet, in light of persistent data concerns, novel econometric techniques have been developed to help overcome certain issues. Panel data methods such as simple fixed effects, random effects and random coefficient estimators have been increasingly used; and as longer time-series data becomes available dynamic specifications have been introduced into panel data methods. The general dynamic panel takes the form: 13. By estimating the civil war growth relation through a dynamic panel, it has the added benefit of not needing to separate the empirical analysis between cross-section and panel estimates in order to measure short and long-term spillover effects of conflict. A dynamic model inherently estimates the long-run coefficients while also reporting short-run results. It must be noted that the advantage of using dynamic panel estimates does have drawbacks in the form of in the form of lagged dependent variable bias, which biases downwards, and heterogeneity bias 10, which biases estimates of upwards. To negate such issues, the choice is to either estimate each country equation individually and take the average of the estimates of introduce fixed effects to the dynamic panel. This paper follows the latter approach, which in the long-run has minimal bias due to and biases working in opposite directions and cancelling each other out (see Dunne et al, 2002). The estimated general first-order dynamic equation used for empirics takes the form: Some of the data issues relate to missing data (growth, education and investment) for periods where a country was in war (i.e. Angola and Sudan in the 1960 s) and no data for countries such as Libya and Somalia. The decision to omit countries from empirical estimation could lead to sample selection bias that will bias the results. This, of course, is not the fault of the authors but the state of data.

13 where y is GDP per capita; is investment/gdp; is secondary education attainment as a share of population over 25; is the labour force growth rate or (n+g+δ).the reparameterised general first order dynamic model has all variables are in log form; with representing the change in the independent and dependent variables (i.e. lny = change in per capita GDP) and lny1, lnk1 and lnsch1 representing the lagged level of per capita GDP, investment as a share of GDP and is secondary education attainment as a share of population over 25respectively. As stated above, and are the contiguity matrices for primary and secondary neighbours and these are interacted with country respective growth rates and conflict indicators. Finally, and are time and country effects, while is the error term. RESULTS The starting point for estimation is to regress the dynamic panel on the classical determinants of growth, presented by Mankiw et al (1992), together with the conflict experience for the host country. These initial regressions, shown in Table 3, serve as benchmarks to which further variables are added to account for spillover effects. Compared to Mankiw et al (1992) s predictions, investment and initial income is of the expected sign and statistically significant, however, differences are found regarding human capital and population growth. Firstly, human capital, measured as secondary school attainment as a proportion of population over 25, is negative and significant. Secondly, population growth plus 0.05, which theoretically should have a negative impact on per capita GDP growth is estimated to be positive and statistically significant. The insignificance of the human capital variable is not uncommon within the growth literature. Islam (1995) attributes the irregular results to the discrepancy between the theoretical variable H used in the model and the actual variable used in regressions. As is the case for Africa and many other low-income countries, the true levels of human capital have not increased much since 1960 and statistically, this leads to a negative temporal relationship between human capital used in regressions and economic growth (Islam, 1995). Moreover, it is often the case that education attainment doesn t not always translate to increase productivity, since in many African countries, education quality is still a large concern. To test the assertion that the choice of the human capital variable used in regressions matters, Table 3, Columns 4, 5 and 6 report regression results when Barro and Lee s education attainment variable is substituted by an index on the returns to human capital, found in Penn World Tables 8.0. Using a different human capital measure, the results in Table 3, Column 4, 5 and 6 provide support for Islam s claim. While, the human capital variable remains insignificant for all three specifications, the sign is now positive. It may indeed be the case that a richer specification of the production function with respect to human capital is required in order to allow the theoretical properties of the human capital variable to be better reflected in regression results. Interestingly, the choice of human capital does not change the estimation results of the other variables, investment, initial income and population growth are all significant and of the same sign.

14 Table 3: The Growth Effects of Conflict, Varying Over Different Conflict and Human Capital Types (1) (2) (3) (4) (5) (6) Conflict Type Conflict Intense Civil Conflict Intense Civil Variables ln(y) ln(y) ln(y) ln(y) ln(y) ln(y) ln(invest) 0.035*** 0.034*** 0.035*** 0.037*** 0.036*** 0.036*** (0.006) (0.006) (0.006) (0.005) (0.005) (0.06) ln(school) *** *** *** (0.040) (0.040) (0.039) ln(school) (0.244) (0.244) (0.245) n+g+δ 0.051*** 0.050*** 0.054*** 0.043*** 0.041*** 0.046*** (0.010) (0.010) (0.009) (0.009) (0.010) (0.010) ln(y0) *** *** *** *** *** *** (0.005) (0.004) (0.005) (0.005) (0.005) (0.005) ln(invest)_ *** 0.018*** 0.018*** 0.020*** 0.020*** 0.020*** (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) ln(school)_ *** *** *** * * (0.005) (0.005) (0.005) (0.024) (0.024) (0.024) Conflict *** *** ** *** *** ** (0.004) (0.007) (0.005) (0.004) (0.007) (0.005) Constant *** *** *** (0.521) (0.520) (0.521) (0.535) (0.535) (0.538) Observations R-squared Note: Dependent variable is ln(y) or growth in per captia GDP. The variable y0 represents the lagged dependent variable (i.e. ln(61)-ln(60)) and is interpreted as initial income. All regressions include a time trend variable. a Barro and Lee (2010) education attainment database b Penn World Table 8.0 human capital index Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 In terms of the relationship between population and economic growth, the existing literature remains divided and it is not uncommon within the literature that population growth actually increases income. The assertion of a population-driven economic growth hypothesis is not uncommon in low-income, developing regions such as Africa and Asia. Moreover, in many post-war economies, population growth has been positively associated with economic growth (Grier and Tullock, 1989). Thus, this paper considers the positive influence that population exerts on growth as a phenomenon of population driven growth, whereby following the end of a conflict, income and population are expected in increase. Moreover, it can be hypothesised that countries in Africa have not reached the stage whereby higher population harms per capita GDP growth.

15 The variable of interest within Table 3 is that of violent conflict. For all six specifications, violent conflict in the host country has a significant negative influence on economic growth. Not surprisingly, an intense conflict (Column 2 and 5) one which features more than battle related deaths in a year has the largest negative impact on growth. On average, an intense conflict decreases growth by between 2.3 to 2.5 percentage points, while all and civil conflicts decrease growth by 1.3 to 1.56 and 1 to 1.3 percentage points respectively. What is clear is that all types of conflict have shown to negatively affect growth. The next step is then to assess spillover effects of conflict on primary and secondary neighbouring countries. Using a method popular within the literature, in order to find spillover effects of conflict, a variety of regressions are run that is based on conflict type, varying definitions and weights of primary and secondary neighbours and its interaction with growth. Due to the large amounts of regressions ran, the results reported are those of the best fitted models, where R 2 is used to determine optimal fit 11. As shown in Table 4, of the different estimated regressions, those using border lengths and minimum distance as primary secondary neighbour contiguity matrices provided the best fit results. It must be noted that the authors of this analysis believe that potential multi-dimensional spillover effects are only possible once two sets of host country neighbours are included in the regression. Another way to think about this is to see the potential interaction between the three types of countries. Each country has at least one type of neighbour and the spillover effect can only be defined as multi-dimensional when it flows to and from at least two types of neighbours. Thus, but including secondary neighbours into the regression, spillover effect can now be consider multi-dimensional. The control variables, consistent with the results in Table 3, are estimated to show investment, population growth and initial income to positively influence economic growth, while education attainment negatively affects growth. Throughout the six specifications of different types on conflict, weighted matrix and inclusion of primary and secondary neighbour conflicts, conflict in the host country is negative and significant. The spillover effects of conflict on neighbouring countries growth rates, however, are slightly different. While conflict irrespective of type negatively affects primary neighbour growth, no such influence was found on secondary neighbours. Overall, the regressions in Table 4 appear to suggest that host-country conflict has a negative impact of 0.9 to 1.2 percentage points, for all and civil conflicts, while violent conflicts lead to a decrease of about 2.3 percentage points. For primary neighbours, the interpretation of the estimation results is slightly different to that of host nations. Since each host country has at least one neighbour, the impact of conflict must be spread over all the relevant countries. On average, a host nation has 3.47 primary neighbours, which translates to a per country influence of According to the results in Table 4, a host country conflict then translates to [0.288*-0.012*100 = -0.35] 0.35 percentage 11 The difference between the reported regressions and regressions with a lower R 2 is negligible, where all the variables are of the same sign and same significance. Moreover, the R 2 between the different regression do not vary by more than 0.02.

16 point reduction in growth. This negative effect varies depending on the type of conflict, with intense and civil conflicts are calculated to reduce primary neighbour growth by 0.5 percentage points while for all types of conflict this drop in growth is the above calculated 0.35 percentage points. This negative spillover effect from a conflict to primary neighbours is calculated to be between 22 and 39 percent of the host country effect. For comparison purposes, the current analysis is very much in line with previous studies. Murdoch and Sandler (2004) predicted civil wars in the host country to negatively influence neighbouring countries by approximately 24 percent of the host country effect, while De Groot (2010) found conflict in the host nation decreased primary neighbour growth by between 0.3 and 0.8 percentage points. Table 4: The Growth Effects of Conflict, Varying Over Different Conflict and Weight Types (1) (2) (3) (4) (5) (6) Conflict Type Conflict Intense Civil Weight Type: Pri Border Border Border Border Border Border Sec Min dist Min Dist Min Dist Variables ln(y) ln(y) ln(y) ln(y) ln(y) ln(y) ln(invest) 0.035*** 0.035*** 0.034*** 0.034*** 0.035*** 0.035*** (0.006) (0.005) (0.006) (0.006) (0.006) (0.006) ln(school) *** *** *** *** *** *** (0.039) (0.040) (0.039) (0.040) (0.040) (0.040) n+g+δ 0.054*** 0.055*** 0.053*** 0.053*** 0.057*** 0.057*** (0.010) (0.010) (0.009) (0.010) (0.010) (0.010) ln(y0) *** *** *** *** *** *** (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) ln(invest)_ *** 0.019*** 0.017*** *** 0.018*** 0.019*** (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) ln(school)_ *** *** *** *** *** *** (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) Conflict *** *** *** *** ** 0.009** (0.004) (0.004) (0.007) (0.007) (0.004) (0.004) W pri *Conflict * * ** * ** ** (0.006) (0.006) (0.008) (0.009) (0.007) (0.007) W sec *Conflict (0.009) (0.012) (0.009) Constant *** *** *** *** *** *** (0.521) (0.001) (0.521) (0.523) (0.522) (0.522) Observations R-squared Note: Dependent variable is ln(y) or growth in per captia GDP. The variable y0 represents the lagged dependent variable (i.e. ln(61)-ln(60)) and is interpreted as initial income. All regressions include a time trend variable. Standard errors in parentheses

17 *** p<0.01, ** p<0.05, * p<0.1 While the impact of conflict on host countries and primary neighbours results, presented in Table 4, are consistent within the literature, differences do emerge when analysing spillover effects on secondary neighbours. Whereas De Groot (2010) found positive spillover effects on secondary neighbours, the current analysis does not yield any similar results. Variations in conflict type, contiguity measurement and weight matrices makes little difference in the results. Possible explanations for the differences in results could be the extra ten years included in the dataset. With the number of conflicts on the decreasing over time and the difficulty in picking up singular shock effects (i.e. conflict) as distance from host nation increase, it might indeed make it difficult to pick up a significant effect. To overcome this potential shortfall, this paper now introduces an additional kind of spillover effect, which measures the duration of conflict. By interacting the contiguity matrices with duration of conflict measured in months - a new type of spillover effect can be created which is able to account for both the multi-dimensional spillover effects as an initial shock and as an impact over time. This method could potentially pick up the above mentioned theoretical spillover channels through which conflict affects growth. Table 5 below contains estimations of conflict experience, varying on conflict and contiguity weight type. As before, border length and minimum distance are the two contiguity matrices for, primary and secondary neighbours, which provided the highest regression fit respectively. Of the six different specifications, column 1, 3 and 5 include only primary neighbour spillover effect, while columns 2, 4 and 6 include the interaction of the weighted contiguity matrix of secondary neighbour s conflict duration. The rationale for this is to distinguish between unidirectional and multidirectional spillover effects 6 and whether prolonged conflicts affect a neighbouring country. Almost identical to the estimates found in Table 4, the control variables are all significant and of the same sign. Investment and population growth are positive determinants of economic growth. Education attainment, on the other hand, is a negative determinant of economic growth. The lagged dependent variable or convergence term is statistically significant and negatively affects growth. The indicator for host country conflict experience is negative and significant 12 for all six specifications. Since conflict in Table 5 is measured as months of conflict, the interpretation of the estimation results are slightly different from that in Table 4. For the host-nation, an extra month of general and civil conflict is estimated to decrease economic growth by 0.1 percentage points. Analogous to the estimates in Table 4, the impact from intense conflict estimated to be higher than general and civil conflict. In the case of Table 5, this effect is doubled at 0.2 percentage point reduction in economic growth. The interpretation of the new type of spillover effect will be grouped into specifications 1, 3 and 5, and then 2, 4 and 6. The unidimensional spillover effect on primary neighbours resulting from conflict is shown in columns 1, 3 and 5. These results are clearly different as 12 The lowest significance of 10% was found in the civil war conflict type.

18 compared to earlier estimations. While the spillover effect on primary nations is negative, it is not statistically significant. This finding is consistent throughout the three specifications. In assessing the unidimensional spillover effect from host-nation to primary neighbours, a prolonged conflict seems to have no significant spillover effect. Table 5: The Growth Effects of Conflict Duration, Varying Over Different Conflict and Weight Types (1) (2) (3) (4) (5) (6) Conflict Type Conflict Intense Civil Weight Type: Pri Border Border Border Border Border Border Sec Min dist Min Dist Min Dist Variables ln(y) ln(y) ln(y) ln(y) ln(y) ln(y) ln(invest) 0.033*** 0.033*** 0.033*** 0.032*** 0.033*** 0.033*** (0.006) (0.006) (0.006) (0.005) (0.006) (0.006) ln(school) *** *** *** *** *** *** (0.031) (0.031) (0.031) (0.031) (0.031) (0.031) n+g+δ 0.081*** 0.081*** 0.085*** 0.086*** 0.085*** 0.084*** (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) ln(y0) *** *** *** *** *** *** (0.005) (0.005) (0.005) (0.005) (0.005) (0.005) ln(invest)_ *** 0.016*** 0.016*** 0.016*** 0.016*** 0.017*** (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) ln(school)_ *** *** *** *** *** *** (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Conflict ** ** ** 0.002** * * (0.000) (0.000) (0.001) (0.001) (0.000) (0.0005) W pri *Conflict Months (0.0006) (0.0007) (0.001) (0.001) (0.0007) (0.0008) W sec *Conflict Months (0.0000) (0.001) (0.001) Constant *** *** *** *** *** (0.524) (0.524) (0.526) (0.528) (0.526) (0.526) Observations R-squared Note: Dependent variable is ln(y) or growth in per captia GDP. The variable y0 represents the lagged dependent variable (i.e. ln(61)-ln(60)) and is interpreted as initial income. All regressions include a time trend variable. Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 To consider multidimensional spillover effects columns 2, 4 and 6 in Table 5 estimates the weighted contiguity matrix of secondary neighbours with conflict duration. As before, conflict - irrespective of type - has a negative and significant impact on host country growth.