Patterns of Conflicts and Effectiveness of Aid

Patterns of Conflicts and Effectiveness of Aid Arcangelo Dimico * Queen s University of Belfast This Version: 13/05/2012 Abstract The effect of aid on civil war is one of the most debated in economics. Using proxies for the evolution of conflict we show that aid significantly affects the escalation and de-escalation of conflicts and both measures are highly sensitive to the rate of change in the share of aid over the course of the conflict. We also show that the incidence of conflict does not have any economic impact once we control for the evolution of conflict while the latter significantly affects economic growth. The effect of the evolution of conflict on economic growth provides evidence of a sort of indirect effect of aid on economic growth in countries experiencing a civil war. * Corresponding author: arcangelo_dimico@hotmail.it 1

1) Introduction Countries experiencing conflicts are among the largest recipient of aid benefiting of an average share of Official Development Assistance (ODA) equal to almost 6 percent of the GDP compared to an average 5.6 percent for countries at peace. The share of ODA is particularly high for countries experiencing an extensive conflict (almost 12 percent of GDP) and for countries in which several groups are involved in the conflict (above 7 percent). The rate of growth of the share of ODA/GDP for conflict economies is also much larger than the share for countries at peace with an average growth equal to 0.2 percent compared to a mere 0.07 percent. It is not surprising therefore to wonder whether aid to conflict economies has any effect in breaking the spell of conflicts which they are experiencing. And if aid is effective, in which way does it affect the evolution of the conflict? The economic literature, both from the empirical and theoretical point of view, has for long tried to provide an answer to these questions often with puzzling results. Sometime aid is found to have a positive effect on conflict, sometime the effect is negative, and other times there is no effect at all. This pattern of results depends on the fact that it is still difficult to understand the main channel through which aid may affect the conflict. If we consider aid as a potential prize then it is possible that aid may affect opportunities shrinking the Potential Settlement Region (Hirshleifer, 1995, 2001) and therefore increasing the probability of conflict. Aid may represent a rent affecting the incentive for rebel groups to engage in a conflict as well as the incentive for the government to defend its rent (Grossman 1991, 1992, 1999). On the other hand aid may spur economic growth, creating opportunities in terms of employment, structural, and distributional policies which then can have an effect on the cost opportunity of citizens to enrol for rebel groups. The higher opportunity cost then makes the conflict a less attractive option (Collier and Hoeffler, 2002). In addition, because aid is fungible (Feyzioglu et al., 1998; McGillivray and Morrissey, 2004; Pack and Pack, 1990; Devarajan, 2

1999) it may be diverted to military expenditure providing the government with a clear military advantage which in turn affects the opportunity for rebel groups to engage in a conflict (Arcand and Chauvet, 2001). From an empirical point of view the evidence is also not so convincing. De Ree and Nielsen (2010) look at a direct effect of aid on the probability of conflict using a 2SLS for a sample of 39 African countries over a 19-years period and they find a significant and direct effect of aid on the probability of conflict. A 10 percent change in the share of aid decreases the probability by 6-8 percent per year. Savun and Tirone (2011) find that democracy aid considerably reduces the risk of conflict by reducing the commitment problem and uncertainty due to a democratic transition. Candland et al. (2011) look at aid shocks and they find a positive effect of negative aid shocks on the risk of conflict. On the other hand Collier and Hoeffler (2004) and Nunn and Qian (2011) seem to be less optimistic. Collier and Hoeffler (2004) find that aid increases military expenditure and that this does not have any effect on the probability of conflict. Nunn and Qian (2011) focus on the U.S. food aid programme and they show that a change in the amount of aid increases the incidence, onset, and duration of civil conflicts in recipient countries. The main problem with evaluating the effect of aid on conflict is determined by the fact that it is difficult to isolate the effect of aid in a regression where the latter changes with other important policy variables. As a result the effectiveness of aid may depend on the specification of the model. For this reason it may be more appropriate to look at the effect of aid on conflict in a context where the national government has no room for manoeuvring while aid keeps changing because its allocation is decide by foreign countries. The evolution of the conflict may therefore represent the ideal set-up to evaluate the effect of externally determined variables like aid given that over the course of the conflict national policies variables are likely to have a small variance (because of limited power of the government) while external imposed shocks are likely to account for a large 3

part of the variation in the conflict. There are also other more practical reasons why it may be worth to focus on the evolution rather than on the incidence, onset, and duration. Most of the times, besides the importance of evaluating the effect of aid on the outbreak of civil war we also want to know whether the international community can help the government in restoring peace in a situation in which the financial involvement of donors becomes more intense. This can only be done looking at the evolution of the conflict given that different from the incidence the evolution permits to evaluate what happens to the conflict when donors increase the share of aid (conditional on the probability of conflict). Because it is a conditional probability this also decreases biases due to the endogeneity of the onset and incidence (see Blatman and Miguel, 2010) and to the use of small samples typical of duration models (see Bleaney and Dimico, 2011). The evolution is also more insightful than the duration given that it is possible to distinguish between the escalation and the deescalation/attenuation of the conflict providing useful insights from a policy point of view. It also may be considered as a primitive of the duration given that the latter will depend on whether the conflict either escalates or deescalates 1. There are several possible candidates that one can use in order to proxy the evolution. The extent of conflict, the number of groups in conflict with the government, and the number of fatalities are all possible viable options. However while the evolution of the geographical extent and of the number of groups in conflict have both important strategic implications, the same it is not particularly true for the evolution of fatalities which of course has important humanitarian consequences. Focusing on the first two proxies of the evolution of conflict above, we find that aid significantly deters the escalation of conflicts both in terms of number of groups and in terms of geographical extent. However, what is even more important for the evolution of the conflict is the rate of change of the share of ODA/GDP over the course of the conflict because it may represent a sort of positive shock 1 Cunningham (2006) and Cunningham et al. (2009) show that the duration of conflict largely depends on the number of groups in conflict. 4

in a static environment. We find that a positive rate of growth of aid over the course of the conflict is associated with a deterrent effect in terms of number of groups involved and geographical extent. The rate of growth of ODA/GDP also affects the probability of groups retreating, significantly shortening the duration of the conflict. These effects are extremely significant from an economic point of view given that the entry of new groups in the conflict decreases the rate of growth of GDP per capita by almost 4 percent, while the withdrawal of any contesting group increases the rate of growth by almost 3 percent. On the other hand the escalation of the conflict in terms of geographical extent decreases GDP growth by almost 5 percent. Besides the economic importance of the evolution/escalation of conflicts there are other important policy implications which need to be discussed. First, increasing the share of aid over the course of the conflict seems to be particularly useful in ethnic fractionalized countries where the entry of new groups seems to be more likely. Given that aid on the one hand permits to prevent the escalation of conflicts and on the other hand increases the probability of contesting rebel groups to retreat, then a change in the share of aid in ethnic fractionalized countries may effectively shorten the duration of conflict. Second, given the preventive effect of aid both in terms of geographical extent and numbers of rebel groups involved in the conflict it seems reasonable to expect that the earlier donors increase the share of aid the lower will be the probability of the escalation of the conflict, shortening the duration, and containing its negative effect on development. The paper is structured as follows. In the next Section we describe dependent variables, explaining the approach we follow in order to build our variables for the evolution of conflict. Section 3 presents empirical results for the effect of levels of ODA on the evolution. In Section 4 we focus on 5

the effect of the rate of change in the share of ODA/GDP and in Section 5 we show the economic impact of the evolution of conflict using a simple growth model. The paper ends with short conclusions and policy implications. 2) Data and Model Despite the number of viable options one can have in order to proxy the evolution of conflicts we decide to focus on the number of groups involved in the conflict and on the geographical extent of the conflict. These two dimensions of the conflict seems to be extremely interesting from a strategic point of view given that duration and intensity of conflict are highly likely to depend on these features. Although geographical extent and numbers of groups are likely to be positively correlated in our dataset the correlation is far from perfect. The correlation between the geographical extent and the number of groups in conflict is only 0.58 and the correlation between a change in these two different dimensions is only 0.28. Given the low correlation it is a good idea to use both measures in order to check the robustness of results. It is also important to evaluate possible differences and resulting policy implications which may refer to different set of countries. For example, highly fractionalized countries may experience several conflicts at the same time (i.e. Democratic Republic of Congo) even though the share of the country affected by the conflict is not large. On the other hand, countries which are not fractionalized may experience extensive conflict even though the number of groups is not large (i.e. Cambodia, and Egypt). The UCDP/PRIO Dyadic Dataset on Armed Conflicts and the Political Instability Task Force (PITF) are the two sources we use to construct our dependent variables. The UCDP/PRIO Dyadic Dataset is used in order to construct a count variable which records the progressive number of 6

groups involved in a conflict against the government for each country-year episode over the 1960-2005 period. On the other hand the PITF provides figures for the geographical extent of the conflict 2. The extent is coded using an ordinal variable between 0 and 4 with 0 denoting episodes with no conflict and 4 denoting cases in which the conflict affects more than a half of the country s surface area 3. Table 1 shows the distribution of the number of groups and geographical extent of conflicts in our dataset. Over the 1960-05 period the PITF records a total of 703 annual episodes of conflict providing an average probability of conflict equal to 11.4 percent. The majority of these conflicts have an extent equal two, which means that in the majority of the episodes a share of the country between 10 and 25 percent is affected by the conflict. On the other hand, the number of conflicts coded by the UCDP/PRIO is equal to 846 4 providing an average probability of conflict equal to 13.7 percent 5. In almost 9 percent of the cases the conflict only involves one rebel group (and the government). However, there is a 3.2 percent of episodes in which the conflict involves two rebel groups and a 1.4 percent of episodes in which the conflict involves three or more different rebel groups. This implies that in almost one third of the episodes with conflict there is more than one rebel group involved in the conflict. 2 3 7 The UCDP/PRIO also provides figures for the radius of the conflict. However this source has potential problems compared to the PITF. First the annual variation in the extent of the conflict is much smaller. Second, for some countries the conflict extends over an unreasonable share of the country. If we divide the area of the conflict by the area of the country we find that the maximal extent in the country is 3,400 percent which seems hard to believe. The PITF coding rule is: 1 if the conflict affects less than 10% of the surface area; 2 if the conflict affects an area between 10 and 25 percent of the surface area; 3 if the surface area affected by the conflict is in between 25 and 50 percent; 4 if it larger than 50 percent. 4 The correlation between the incidence in the PITF and in the UCDP/PRIO is 0.72 5 Differences in the number of conflicts between the two sources are the result of the different coding rule applied. The UCDP/PRIO Dyadic Dataset on Armed Conflicts defines an armed conflict as: a contested incompatibility that concerns government and/or territory where the use of armed force between two parties, of which at least one is the government of a state, results in at least 25 battle-related deaths in a year 5. On the other hand the PITF defines a conflict as an episode of violent conflict between the government and a politically organized group where each party mobilizes 1000 or more people (armed agents, demonstrators, troops), and resulting in at least 1000 direct conflict-related deaths over the full course of the armed conflict and at least one year when the annual conflict-related death toll exceeds 100 fatalities.

Table 1: Descriptive Statistics for the Extent and Number of Groups PITF UCDP/PRIO Extent Freq. Percent Groups Freq. Percent 0 5,465 88.6 0 5,322 86.28 Extent < 10% 217 3.52 1 557 9.03 10%< Extent < 25% 251 4.07 2 197 3.19 25%< Extent < 50% 125 2.03 3 58 0.94 Extent > 50% 110 1.78 4 14 0.23 5 7 0.11 6 4 0.06 7 3 0.05 8 6 0.1 Total 6,168 100 Total 6,168 100 Table 2 shows the distribution of ODA/GDP given the geographical extent and the number of groups involved in the conflict. Countries at peace receive an average share of ODA equal to 5.6 percent of the GDP. There is a massive 12.1 percent of ODA (in terms of GDP) received by countries experiencing a conflict affecting more than 50 percent of the surface area. The share of ODA is also above the mean for countries where the government faces either one or five rebel groups. The share is equal to 6.1 percent for countries in which there is only one rebel group, while the share of ODA for countries where there are five rebel groups is equal to 7.7 percent. ON the other hand, the share of ODA becomes relatively small (below 1 percent) if there are more than 5 groups. 8

Table 2: ODA/GDP per Extent and Nr of Groups in Conflict Extent of Conflict ODA/GDP Std. Dev. Nr Groups ODA/GDP Std. Dev. 0 0.0564877 0.1023408 0 0.0567246 0.1027441 Extent < 10% 0.0403742 0.0701157 1 0.0610886 0.0937362 10%< Extent < 25% 0.0572187 0.0903296 2 0.0535041 0.1186654 25%< Extent < 50% 0.043343 0.0557456 3 0.0520643 0.0785595 Extent > 50% 0.1215668 0.1681074 4 0.0433767 0.0498396 5 0.0774599 0.1080709 6 0.0047759 0.0017071 7 0.0053173 0.0043445 8 0.003406 0.001382 Total 0.0568468 0.1021111 Total 0.0568468 0.1021111 Using the distribution for the geographical extent and number of groups in Table 1 we construct new variables which proxy the evolution (escalation/deescalation) of the conflict. The evolution is computed using first differences and it is summarized in Table 3. A negative change denotes cases in which the conflict deescalates either in terms of geographical extent or number of groups involved. On the other hand, positive changes denote an escalation of the conflict. With regard to the geographical extent there are almost 258 country-year episodes in which the geographical extent changes with respect to the extent in the previous year. Positive and negative changes for the extent are almost equally distributed. In addition the extent of conflicts seems to be relatively stable over the course given that in the majority of the cases the conflict only escalates/deescalates by one unit 6. On the other hand, the number of episodes in which the number of groups changes over the course 6 A one unit change in Table 3 can also represent the onset/outset of conflict given that the distribution is not conditional on conflict. 9

is 462 and even in this case in the majority of the episodes there is only a single group which stepsin or retreat. Table 3: Descriptive Statistics for the Evolution of Conflict (Not Conditional on Conflict) PITF UCDP/PRIO Extent Growth Freq. Percent Nr Groups Growth Freq. Percent -4 13 0.21-3 11 0.18-3 21 0.35-2 21 0.35-2 37 0.61-1 191 3.14-1 61 1 0 5,613 92.4 0 5,817 95.75 1 203 3.34 1 64 1.05 2 29 0.48 2 40 0.66 3 7 0.12 3 12 0.2 4 10 0.16 Total 6,075 100 Total 6,075 100 However, changes in the tail of the distribution in the Table above are relatively small causing potential biases to the analysis. To deal with such a problem we summarize positive and negative changes using four different dependent dummy variables. The first dummy variable is equal to one if there has been a positive change in the geographical extent of the conflict. The second dummy is coded one if there has been a negative change in the geographical extent. The third variable is positive if any group steps-in the conflict. Finally the fourth dummy is coded one if any of the rebel groups in conflict retreats. Table 4 shows the conditional distribution of these four dummy dependent variables. Out of 699 episodes of conflict for which we have data for the geographical extent there are a total of 126 cases in which the conflict escalates and a total of 49 episodes in which the conflict deescalates. On the 10

other hand, there are 239 cases in which new groups step-in over the course of the conflict, and 72 cases in which groups withdrawal (out of a total of 840 episodes of conflict for which we have information about the number of groups). Table 4: Dummy Dependent Variables for the Evolution of Conflict (Conditional on Conflict) Positive Change in the Extent Positive Change in the Nr of Groups Increase in the Extent Freq. Percent Cum. Groups Stepping In Freq. Percent Cum. 0 573 81.97 81.97 0 601 71.55 71.55 1 126 18.03 100 1 239 28.45 100 Total 699 100 Total 840 100 Negative Change in the Extent Negative Change in the Nr of Groups Decrease in the Extent Freq. Percent Cum. Groups Retreating Freq. Percent Cum. 0 650 92.99 92.99 0 768 91.43 91.43 1 49 7.01 100 1 72 8.57 100 Total 699 100 Total 840 100 While positive changes (both for the extent and number of groups) are fairly large, the number of negative changes is relatively small which then can underestimated the probability if estimated using simple probit/logit models given that the capacity of the model to determine a cut-off point is biased in the direction of favoring zeros at the expense of ones (King and Zeng, 2001a, 2001b). A possible alternative to these models is to use a rare-event logit/probit. However because the evolution is conditional on the probability of conflict this would require to estimate a two-step Heckman model (Heckman, 1979) which in turn is not an efficient estimator. Heckman-Probit models can also be estimated using a Maximum Likelihood Estimator which provides efficient estimates. On the negative side the MLE Heckman Probit does not allow to control for rare events underestimating the probability. Given the impossibility to find a good trade-off between efficiency and consistency we decide to use a Maximum Likelihood Heckman-Probit Model in order to gain efficiency in the estimates. 11

The Heckman-Probit we use in order to estimate the effect of ODA on the evolution of conflict can be written as follow: Incidence i,t * = Φ(λ 1 ODA/GDP i,t-1 + λ 2 Incidence i,t-1 + λ 3 Discrimination i,t-1 + λ 4 W i,t-1 ) (1) Equation (1) represents the selection equation which drives the probability of conflict. The probability of conflict depends on the Incidence (t 1), on the share of ODA (t 1), and a set of control variables W i,t-1 which includes GDP per capita (t 1), population (t 1), ethnic fractionalization, share of mountainous terrains, a dummy for oil producers, and government consumption (t 1) 7. The variable Discrimination i,t-1 denotes the share of the population which is politically discriminated and which in our model represents the exclusion restriction. The reason to choose this exclusion is because the share of the population which is discriminated may represent a sort of opportunity cost to engage in a civil war. However, once the war has started the effect of the latter on the evolution will depend on how discrimination is distributed across groups and across regions. This cross-group and cross-regional discrimination is likely to be captured by the ethnic fractionalization of the country and because of that the share of the population which is discriminated should not affect the evolution. Once the probability of conflict is computed the conditional probability of either a positive or a negative change will be determined by the following outcome equation: (Evolution i,t Incidence=1) = Φ(β 1 ODA/GDP i,t-1 + β 2Level i,t-1 + β 3 X i,t-1 ) (2) where Evolution i,t is our proxy for the evolution of conflict, Level i,t-1 is the level of conflict (t -1), X i,t-1 is the same set of control variables as in the selection equation, and ODA/GDP i,t-1 is the share of ODA (t 1). 7 Government consumption is used to control for a possible diversion of aid to current expenditure and/or military expenditure. 12

Data for the share of ODA at current US dollars, countries GDP at current US dollars, and real GDP per capita are collected from the World Bank (WDI). The Penn World Table 7 is used to collect data on government consumption and population; Fearon and Laitin (2003) is the source for data on ethnic fractionalization, and mountainous terrains; the Polity IV dataset is the source for democratic variables, while the Ethnic Power Relations (2010) dataset provides data on the share of the population which is discriminated. 3) Evolutions of Conflict and Levels of ODA We start our analysis by looking at the effect of levels of ODA (the share of ODA/GDP) on the evolution of conflict. Table 5 reports estimates for the selection and the outcome equation of the Heckman Probit. Panel A (on the top) shows estimates for the outcome equation where the dependent variable is the evolution of conflict. Panel B (on the bottom) shows the selection equation where the dependent variable is the incidence of conflict which is also coded using data from the PITF. Model 1 and Model 2 in Panel A report estimates for a negative and a positive change in the geographical extent of the conflict respectively. On the other hand, Model 3 and Model 4 in Panel B report estimates for a negative and a positive change in the number of groups involved in the conflict, respectively. Given that data on number of groups involved in the conflict is provided by the UCDP/PRIO, we use data on the incidence from this other source to estimate the incidence for these other two models (Model 3 and Model 4 in Panel B). Selection equations are quite standard. When we use data from the PITF (Model 1 and Model 2 in Panel B) the probability of conflict increases with the share of the population which is discriminated 13

and with possible conflicts in the previous period. GDP per capita, and the share of ODA/GDP have both a negative effect on the probability of conflict. On the other hand, when we use data from the UCDP/PRIO (Model 3 and Model 4 in Panel B) the share of ODA/GDP becomes not significant while ethnic fractionalization and government consumption are both significant. This difference between the first and the last two models is the result of different coding rules across datasets. It is well known that the UCDP/PRIO tends to over-estimate the effect of ethnic fractionalization (see Bleaney and Dimico, 2011; and Hegre and Sambanis, 2004), on the other hand the PITF seems to over-estimate the effect of ODA on the incidence of conflict. Regarding the evolution of conflict ODA/GDP seems to have a significant and negative effect on a positive change in the geographical extent of the conflict (Model 2) which decreases the probability of the conflict escalating in terms of geographical spread. On the other hand, ODA/GDP does not seem to have any effect on the probability of the conflict deescalating in terms of territorial diffusion (Model 1). This effect implies that ODA/GDP can prevent the territorial diffusion of the conflict, but once the conflict has already spread its effect on a possible de-escalation of the conflict is not statistically significant. The same it is true for the number of groups involved in the conflict. The share of ODA/GDP does not affect the probability of groups retreating (Model 3) but it significantly affects the probability of new groups entering in the conflict (Model 4) confirming a sort of preventive effect of ODA/GDP on the diffusion of the conflict. This preventive effect in turns can explain the significance of ODA/GDP on the probability of conflict in Model 1 and Model 2 (Panel B). Apart from the share of ODA/GDP there are no other variables which significantly prevent the diffusion of the conflict either in terms of geographical extent or number of groups involved. This 14

may be related to the fact that the share of ODA/GDP is the only policy variable outside the control of government and it is likely to be one of the few tools which can be used to prevent the diffusion of conflict in a situation of political stalemate. Because of limited legitimacy and resources, national governments have a small room for manoeuvring and because of that policy variables under the control of the government may have a smaller variance explaining the insignificance of policycontrolled variables. 15 Table 5: MLE Heckman Probit Evolution and Incidence of Conflict Panel A: Second Step (Outcome Equation): The Evolution of Conflict Model 1 Model 2 Model 3 Model 4 Dependent Variables Δ- Extent Δ+ Extent Δ- Groups Δ+ Groups Log Population (t -1) 0.118-0.0898-0.187*** 0.0588 (1.29) (-1.44) (-3.53) (1.37) Log real GDP per cap. (t - 1) 0.107-0.0997 0.0116-0.0944 (1.33) (-1.31) (0.21) (-1.48) Oil producers Dummy 0.0603-0.268-0.261** -0.0327 (0.25) (-1.38) (-2.30) (-0.20) Ethnic Fractionalization 0.0398 0.309-0.236 0.420 (0.13) (1.00) (-1.20) (1.52) ODA/GDP (t - 1) 1.406-5.721*** 0.276-1.244** (1.57) (-6.46) (0.53) (-2.32) Government Consumption (t - 1) 0.00483-0.0897-0.0738 0.0712 (0.04) (-0.77) (-0.92) (0.66) Extent (t - 1) 0.478*** -0.215*** (2.87) (-2.83) Nr Groups (t - 1) 0.412*** 0.0214 (7.61) (0.52) Constant -5.066** 1.627 2.162* -2.007* (-2.29) (1.14) (1.82) (-1.85) LR test of indep. eqns. (rho = 0) 0.299 0.000*** 0.000*** 0.000*** Observations 630 630 751 751 Panel B: First Step (Selection Equation): Incidence of Conflict Model 1 Model 2 Model 3 Model 4 Dependent Variable Sources PITF PITF UCDP/PRIO UCDP/PRIO Log Population (t -1) 0.0825** 0.0805** 0.146*** 0.129*** (2.40) (2.35) (5.55) (5.06) Log real GDP per cap. (t - 1) -0.199*** -0.198*** -0.116*** -0.117*** (-4.95) (-4.98) (-3.96) (-4.03) Oil producers Dummy 0.144 0.153 0.220*** 0.218*** (1.35) (1.46) (2.70) (2.66) Ethnic Fractionalization -0.0400-0.0521 0.330** 0.338** (-0.22) (-0.29) (2.27) (2.34) ODA/GDP (t - 1) -1.512*** -1.554*** -0.652-0.733 (-2.69) (-2.76) (-1.37) (-1.56) Government Consumption (t - 1) 0.0556 0.0543 0.124** 0.139** (0.74) (0.73) (2.11) (2.32)

Discriminated Population Share 0.779*** 0.768*** 0.794*** 0.674*** (3.31) (3.87) (4.11) (4.20) Incidence (t - 1) 3.173*** 3.179*** 2.470*** 2.515*** (36.17) (36.58) (36.53) (37.45) Constant -2.258*** -2.225*** -3.901*** -3.653*** (-3.04) (-3.00) (-6.82) (-6.52) Observations 4,451 4,451 4,451 4,451 z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 4) Evolution of Conflict and Change in ODA The problem with results in the previous section relates to the fact that the evolution of the conflict is more likely to change as a response to changes in the share of ODA rather than to levels. Because of that in Table 6 we also look at the effect of changing the share of ODA on the escalation/deescalation of conflicts. As in the previous table, Panel A (on the top) shows estimates for the outcome equation where the dependent variable is the evolution of conflict. Panel B (on the bottom) shows the selection equation where the dependent variable is the incidence of Conflict. Starting with the selection equation (Panel B) where the dependent variable is the incidence of conflict we find that the rate of growth of the share of ODA/GDP has a significant and negative effect on the probability of conflict in all Models decreasing the probability of conflict by an average 20 percent per a one percent in the change of the share of ODA/GDP (Model 1 to 4 in Panel B). This seems a huge effect but it needs to be evaluated with respect to the mean rate of change of ODA/GDP which is equal to 0.1. From this point of view increasing the share of ODA/GDP by one percent represents a sort of positive shock which is likely to have a robust effect on the probability of conflict. The effect of ODA/GDP on the probability of conflict seems to be the result of a significant effect of its rate of growth on the evolution of conflict. Increasing by one percent the rate 16

of growth of aid increases the probability of the conflict deescalating in terms of geographical extent by almost 12 percent (Model 1, Panel A). This is likely to be the result of a clear military advantage the government can achieve by diverting more resources to the settlement of the conflict. On the other hand the effect of ODA/GDP on the probability of the conflict escalating in terms of geographical diffusion is not significant. With regard to the effect of the rate of change ODA/GDP on number of groups involved in the conflict, increasing the rate of growth of ODA/GDP by one percent increases the probability of groups exiting from the conflict by almost 30 percent and decreases the probability of new groups entering in the conflict by almost 32 percent. 17 Table 6: MLE Heckman Evolution and Incidence of Conflict Panel A (Outcome Equation): Evolution of Conflict Model 1 Model 2 Model 3 Model 4 Dependent Variables Δ- Extent Δ+ Extent Δ- Groups Δ+ Groups Log Population (t -1) 0.101-0.0262 0.0624-0.176*** (1.05) (-0.42) (1.32) (-3.38) Log real GDP per cap. (t - 1) 0.0849 0.0357-0.0709 0.00957 (1.08) (0.46) (-1.17) (0.24) Oil producers Dummy 0.000530-0.184 0.0342-0.281*** (0.00) (-0.93) (0.20) (-2.63) Ethnic Fractionalization 0.0421 0.285 0.494* -0.212 (0.14) (0.83) (1.78) (-1.06) ΔODA/GDP (t - 1) 2.921** -1.208-2.366*** 1.364* (1.97) (-0.53) (-3.48) (1.85) Government Consumption (t - 1) 0.0216-0.153 0.0310-0.0479 (0.18) (-1.20) (0.30) (-0.70) Extent (t - 1) 0.477** -0.213** (2.56) (-2.41) Nr Groups (t - 1) 0.0278 0.406*** (0.70) (8.47) Constant -4.569** -0.492-2.279** 1.949** (-1.99) (-0.34) (-2.01) (1.97) LR test of indep. eqns. (rho = 0) 0.355 0.000*** 0.000*** 0.000*** Observations 625 625 742 742 Panel B (Selection Equation): Incidence of Conflict Model 1 Model 2 Model 3 Model 4 Dependent Variable Sources: PITF PITF UCDP/PRIO UCDP/PRIO Log Population (t -1) 0.0991*** 0.0972*** 0.132*** 0.149*** (2.92) (2.88) (5.08) (5.64)

Log real GDP per cap. (t - 1) -0.171*** -0.169*** -0.107*** -0.107*** (-4.56) (-4.53) (-3.95) (-3.98) Oil producers Dummy 0.208* 0.216** 0.244*** 0.243*** (1.93) (2.06) (2.99) (3.00) Ethnic Fractionalization -0.0770-0.0840 0.310** 0.306** (-0.43) (-0.46) (2.11) (2.11) ODA/GDP (t - 1) -2.394*** -2.339*** -1.898*** -1.747*** (-3.36) (-3.29) (-2.86) (-2.61) Government Consumption (t - 1) 0.0276 0.0256 0.121** 0.109* (0.36) (0.34) (2.02) (1.90) Discriminated Population Share 0.848*** 0.839*** 0.695*** 0.805*** (3.52) (3.57) (4.45) (4.16) Incidence (t - 1) 3.173*** 3.176*** 2.525*** 2.478*** (35.86) (35.98) (37.65) (36.40) Constant -2.748*** -2.728*** -3.774*** -4.013*** (-3.87) (-3.86) (-6.84) (-7.36) Observations 4,370 4,370 4,370 4,370 z-statistics in parentheses *** p<0.01, ** p<0.05, * p<0.1 4) The Economic Impact of the Evolution of Conflict Of course economists are interested in the economic impact of policies. From this point of view allocating aid to countries experiencing a conflict in order to affect the evolution would be justified if the latter has a significant effect on economic growth and development. For this reason in this section we evaluate the economic impact of the evolution of conflict. In order analyse the effect of the evolution on development we construct five-year episodes of growth. Five-year episodes of growth are quite common in the literature on conflict (Chen et al., 2008; Collier and Hoeffler, 2004; and Elbadawi et al., 2008) even though it may be argued that economic growth needs to be evaluated over a long-term horizon. The reason why we need to look at a shorter period is because it would be difficult to proxy the evolution of conflict over a longer time-horizon given its annual variation. Using a longer time-horizon may therefore cause a loss of useful information related to this variation. Using five-year episodes on the other hand represents a fair compromise between the need to give to the evolution of conflict the necessary variation and the need of evaluating a sort of 18

intermediate effect in terms of economic growth. Data on real GDP per capita growth, real GDP per capita, and ODA/GDP are collected from the WDI. We use the Penn World Table 7 to collect data on government consumption, trade/gdp, investment/gdp, and population growth. Variables for the evolution of conflict are the same as the ones we used in previous sections. Therefore, we use four dummy variables to proxy whether in the five year there has been an escalation or a de-escalation of the conflict either in terms of geographical extent or in terms of number of groups involved. Table 7 shows growth-estimates using a 2-step GMM estimators in order to deal with a possible endogeneity of the incidence and evolution of conflict. In Model 1 we test a possible growth effect of the escalation/de-escalation of conflict in terms of rebel groups involved in the conflict. While the incidence of conflict is found to be not significant, the evolution of conflict does have a robust effect on economic growth. The rate of growth of GDP decreases by almost 3.7 percent if any new group step-in the conflict over the five-year period. This effect is significant at a 1 percent level. On the other hand, the rate of growth increases by almost 2.8 percent if any of the existing rebel groups retreat. In Model 2 we look at the escalation/de-escalation of conflict in terms of geographical extent which also has a significant growth effect. Same as in Model 1, the incidence of conflict is not significant which means that the conflict does not have any economic impact if it remains stable over time. However, if the conflict escalates in terms of geographical extent than the rate of growth of output decreases by an average 4.7 percent over the five-year period. Given that ODA is the only variable which affect the evolution of conflict this implies that there is a significant indirect effect of ODA on economic growth in conflict economies. 19

Table 7: Economic Impact of the Evolution Dependent Variable: 5-year Average real per capita GDP growth Estimation Method: 2-step GMM Model 1 Model 2 Log Real GDP per Capita, Lagged -1.545** -1.368** (-2.56) (-2.07) Log (1 + Investment/GDP) 2.823*** 2.657*** (2.94) (2.59) Log (1 + Government Consumption) -2.103** -0.805 (-2.28) (-1.03) Log (1 + Trade/GDP) -0.215-0.276 (-0.34) (-0.45) Population Growth 27.01 27.17 (0.79) (0.75) Aid Recipient Dummy -5.436*** -4.248** (-2.70) (-2.06) ODA/GDP 2.741 0.203 (0.62) (0.05) Incidence -0.438 1.106 (-0.54) (1.36) Δ+ Number of Groups -3.674*** (-2.92) Δ- Number of Groups 2.831*** (2.99) Δ+ Extent -4.734*** (-3.27) Δ- Extent -0.317 (-0.42) Ethnic Fractionalization -1.267-1.703 (-0.94) (-1.41) Asia dummy -0.871 0.621 (-0.71) (0.48) SSafrica Dummy -2.402-1.703 (-1.58) (-1.23) Constant 16.53* 12.10 (1.92) (1.41) AR (1) Test (p-values) 0.012 0.009 20

AR (2) Test (p-values) 0.791 0.691 Hansen J-Statistics (p-values) 0.868 0.742 Observations 882 882 Number of Countries 142 142 Robust z-statistics in parentheses (Windmeijer; 2005): *** p<0.01, ** p<0.05, * p<0.1 GMM Instruments: Lag (1/2) Xt or Lag (2/3) Xt depending on whether the variable is endogenous or predetermined IV Instruments: Time dummies 5) Conclusions Whether aid has a positive or negative effect on the probability of conflict and the economic impact of aid in conflict economies are two of the most important issues that international donors face when they have to decide whether to allocate aid to countries experiencing a conflict. Using proxies for the evolution of conflict we show that aid has a significant effect on the probability that the conflict will either escalate or deescalate and that by affecting the evolution of conflict aid also affects the incidence of conflict. In addition we find that the evolution of conflict is what matters for economic growth. Therefore by affecting the evolution of conflict aid has also a significant and indirect effect on economic growth. We also find that policy variables under the control of the government are hardly significant in explaining the evolution of conflict which is likely to be the result of the small room national policy makers have for manoeuvring. For this reason the support of the international community is essential in countries experiencing a conflict given that international donors are the only one who can provide enough variation in national budgets. 21

References Arcand, Jean Louis & Lisa Chauvet, 2001. Foreign Aid, Rent-Seeking Behavior, and Civil War, mimeo Blattman, Christopher, and Edward Miguel, 2010. "Civil War." Journal of Economic Literature, 48(1): 3 57. Bleaney, M.F. & Arcangelo Dimico, 2010. Incidence, Onset and Duration of Civil War: A Review of the Evidence, forthcoming in the Journal of Peace Research Burnside, Craig & David Dollar, 2000. "Aid, Policies, and Growth," American Economic Review, American Economic Association, vol. 90(4), pages 847-868, Candland, T., Davis, Z. S., Findley, M. G., Nielsen, R. A. and Nielson, D. L., 2008 "Can Peace Be Bought? The Effects of Foreign Aid on Civil War Onset" Paper presented at the annual meeting of the ISA's 49th annual convention Cederman, Lars-Erik, Brian Min, Andreas Wimmer, 2009. "Ethnic Power Relations Dataset", http://hdl.handle.net/1902.1/11796 Chen, Siyan, Norman V. Loayza, and Marta Reynal-Querol, 2008. The aftermath of civil conflict, World Bank Economic Review 22(1): 63-85. Collier, Paul & Anke Hoeffler, 2002. Aid, Policy, and Peace, Defence and Peace Economics, Vol 13, Issue 6 2002, pages 435-450 Collier, Paul & Anke Hoeffler, 2004. Greed and Grievance in Civil War, Oxford Economic Papers 56, 563 95. Collier, Paul & Anke Hoeffler, 2007. `Unintended Consequences? Does Aid Fuel Arms Races?', Oxford Bulletin of Economics and Statistics 69(1): 1-27 Cunningham, David E., 2006. Veto Players and Civil War Duration, American Journal of Political Science, 50(4): 875-92 Cunningham, David E., Kristian S. Gleditsch, & Idean Salehyan, 2009. It Takes Two: A Dyadic Analysis of Civil War Duration and Outcome, Journal of Conflict Resolution, 53(4): 570-97 De Ree, Joppe and Eleonara Nillesen. 2009. Aiding Violence or Peace? The Impact of Foreign Aid on the Risk of Civil Conflict in Sub-Saharan Africa. Journal of Development Economics 88 (2): 301-313 Devarajan, S., Rajkumar, A. S. and Swaroop, V., 1999, What Does Aid to Africa Finance? World Bank Policy Research Working Paper No. 2092. Elbadawi, I. & N. Sambanis, 2002. How Much Civil War Will We See? Explaining the Prevalence of Civil War, Journal of Conflict Resolution 46(2): 307-334. 22

Fearon, J. & D. Laitin, 2003. Ethnicity, Insurgency, and Civil War, American Political Science Review, 97(1):75 90. Feyzioglu, T., Swaroop, V. and Zhu, M., 1998. A Panel Data Analysis of the Fungibility of Foreign Aid, World Bank Economic Review, 12(1), pp. 29-58. Grossman, H. I. (1991). A general equilibrium model of insurrections, American Economic Review 81(4), 912-921. Grossman, H. I. (1992). Foreign aid and insurrection, Defence Economics 3, pp. 275-288. Grossman, H. I. (1999). Kleptocracy and revolutions, Oxford Economic Papers 51, 267-283. Harbom, Lotta, Erik Melander & Peter Wallensteen, 2008. Dyadic Dimension of Armed Conflict, 1947 2007, Journal of Peace Research 45(5): 697 710. Heckman, James. 1979. Sample Selection Bias as a Specification Error. Econometrica 47:153 161 Hegre, Håvard & Nicholas Sambanis, 2006. Sensitivity Analysis of the Empirical Literature on Civil War Onset, Journal of Conflict Resolution 50(4), 508 535. Hirshleifer, Jack. 1995. Theorizing About Conflict. In K. Hartley and T. Sandler, eds. The Handbook of. Defense Economics 1 Hirshleifer, Jack, 2001. The Dark Side of the Force: Economic Foundations of Conflict Theory, Cambridge University Press King, Gary & Langche Zeng. 2001a. Explaining Rare Events in International Relations." International Organization 55:693{715. King, Gary & Langche Zeng. 2001b. Logistic Regression in Rare Events Data." Political Analysis 12:137{163. McGillivray, M., and O. Morrissey (2004), 'Fiscal Effects of Aid' in T. Addison and A. Roe (eds.),fiscal Policy for Development, Basingstoke: Palgrave Macmillan/WIDER, pp. 72-96 Nunn, Nathan & Nancy Qian, 2012. "Aiding Conflict: The Impact of U.S. Food Aid on Civil War," NBER Working Papers 17794. Pack, H. and J.R. Pack. 1990. Is foreign aid fungible? The case of Indonesia. Economic Journal, vol. 100 no. 3: 188 94. Political Instability Task Force, (2009). Internal Wars and Failures of Governance, 1955-2007. Center for Global Policy in the School of Public Policy at George Mason University. Savun, B. and Tirone, D. C., 2011. Foreign Aid, Democratization, and Civil Conflict: How Does Democracy Aid Affect Civil Conflict?. American Journal of Political Science, 55(2),: 233 246 23

Windmeijer, Frank (2005). A Finite Sample Correction for the Variance of Linear Efficient Two- Step Estimators, Journal of Econometrics 126(1): 25-51 24