Returnees Spark More Violence: The Causal Effect of Refugee Return on Civil Conflict Intensity

Returnees Spark More Violence: The Causal Effect of Refugee Return on Civil Conflict Intensity Kara Ross Camarena Prepared for: The 70th Annual MPSA Conference International Relations Poster Session April 2012 Abstract A small literature has recently suggested that refugees flows across borders is one mechanism through which armed conflict is spread. The difficulty with this research is that civil conflict and refugee flows are often endogenous, and therefore it is difficult to isolate a causal impact. I add to this growing body of knowledge by examining the impact of refugee returns on civil conflict violence. I address some of the previous research difficulties by exploiting exogenous variation in natural disasters to isolate the impact of returnees on violence apart from refugee return responding to decreases in violence. Using the natural disaster interacted with refugee population as an instrument, I find that refugee return spurs a statistically significant, albeit small, increase in violence. When I examine how these findings fit with theory espoused in the conflict clustering literature, I find little in the way of a satisfying explanation. I draw on the literature on rebel recruitment and some work in economic development on food price shocks and labor markets to propose a causal chain through which refugee return leads to increased violence in an ongoing civil war. Namely, returnee influxes represent a food prices shock that reverberates through the labor market, making the rebel offers for fighting relatively better than working. I then test this food price-rebel recruitment mechanism. I find consistent evidence that returnee influx cause food price increases in cases where violence subsequently increases. However, on average and among countries where violence levels do not change or decrease, returnee influxes are associated with food price decreases. Kara Ross Camarena, Graduate Student, Department of Government, Harvard University, karaross@fas.harvard.edu Preliminary, please do not cite. 1

Introduction A few articles have been written recently about the regional clustering of militarized disputes, civil wars and civil strife. In the literature, refugees are thought to be one carrier of these types of conflicts. Particularly, in civil conflict, refugees are thought to spread violence through one of two channels, (1) because they are or become militarized and work with rebels in their receiving country or (2) because they cause additional economic and security burdens that lead to greater instability. (see Salehyan and Gleditsch (2006) and Salehyan (2008) for more details). While there is some empirical evidence to support the notion that refugees spread conflict, it does not elucidate through what mechanism. One logical extension of the notion that refugees spread conflict is that when they return to their country of origin there should be increased violence in the conflict they had once fled. It could be the case that refugees become more militarized while away or because when they return home they represent the same kind of demand shock to their home country that they did to the asylum country when they first arrived. Such a relationship may be difficult to unpack empirically because it likely is hidden by the more conventional understanding that refugees decide return to their country of origin when violence decreases, or as the conflict ends, because they anticipate they will be able to do so safely. To examine this relationship, I make use of the exogenous variation in natural disasters in asylum countries and instrument for refugee return. I find that while the relationship between refugee return and increased violence in civil conflicts, is negative but insignificant in an ordinary least squares regression (OLS), the two stage least squares (2SLS) regression reveals a positive relationship between refugee return and increased violence. In other words, when we account for reason the possibility that refugees may want to return home when they anticipate decreases in violence, we see refugee return is associated with an increase in violence. Furthermore, this positive relationship becomes significant when controlling for characteristics traditionally thought to explain variation in 2

civil war including ethnic fractionalization, GDP per capita, and rough terrain. It is difficult to link the existing theory to these findings on refugee return and violence. There is no reason to believe that militarized refugees would disproportionately respond to natural disasters by returning home. These findings are not consistent with a militarized refugee explanation. The second explanation in the literature is that the increased violence from refugee flows is brought about because of a shock leading to resource scarcity and instability. While such a theory remains plausible, the mechanism is not all together clear. I draw on the literature on rebel recruitment and economic development literature on economic shocks, food prices, and political violence to propose a causal mechanism through which refugee return leads to increased violence in an ongoing civil war. I then empirically test the mechanism. Exploiting the exogenous variation in natural disasters, I instrument for refugee return and compare the impact on food prices in countries where conflict intensity increases, decreases and remains the same. I find some evidence consistent with the food price-rebel recruitment mechanism. In the rest of the paper, I proceed as follows: First, I explore the relationship between refugee flows and civil war intensity and arrive at a reduced form empirical specification. Second, I propose an instrumental variable and discuss how the IV identification strategy improves upon previous research. Third, I explain my data set. Fourth, I examine my empirical findings and discuss some shortcomings. Fifth, I make use of existing literature to build a causal mechanism linking refugee return to food price increases, rebel recruitment and increases in violence. Last, I test the mechanism and conclude with some discussion of the limitation of this work and possible avenues for additional research. I: Refugee Flows and Civil War Three recent articles have been written on the regional clustering of armed conflict. Gleditsch et al. (2008) explores how civil wars cause international conflicts. In a short section, they suggest 3

that refugee flows are one mechanism for conflict spillover. Salehyan (2008) explores the refugee relationship in more detail. He finds that refugee flows are associated with sender countries initiating military action against receiving countries, as well as receiving or potential recipient countries using their military against the sender country. Salehyan and Gleditsch (2006) focus on civil war onset and find being a refugee recipient country increases the likelihood of a civil war. Collectively, the scholarship thus far argues that armed conflict is spread through refugee flows. There are essentially three explanations explored in the literature. First, refugees may be the reason for international disputes either because one country pursues its refugees across an asylum country s border or because a potential asylum country uses its military to prevent refugees from entering (Gleditsch et al., 2008). Second, refugees often present additional economic, administrative, and security burdens on an asylum country s governments, and this leads to destabilization (Salehyan and Gleditsch, 2006). Last, militarized refugees may connect with like-minded asylum country nationals, broadening the scope of a rebel movement to include an asylum country (Salehyan and Gleditsch, 2006). One problem with the empirical literature, particularly with regard to civil conflict, is that it fails to take into account the possibility that the reasons the refugees move to a particular asylum country may simultaneously be the reasons the asylum country is more prone to conflict. For example, Rwandans may flee to Burundi not just because of its proximity, but because Burundi has similar ethnic groups, language, history and way of life. It is not altogether clear whether it is these similar characteristics, or the presence of the refugees that spreads the armed conflict. It is possible that the conditions for civil war are present in two countries. One country s civil war begins first and refugees might flee to the second country. Then, the second country s civil war begins and we infer, incorrectly, a causal relationship, a spreading of the civil war. In fact, the civil war in the second country, as I have indicated above, may have been caused because of similar reasons as the first country, not as a result of spillover or contagion. 4

I aim to examine the refugee-violence relationship further to address some of these causal inference problems. To evaluate the impact that refugees have on civil conflict, I look at the effect of refugees who return to their country of origin. If refugees flows are a source of instability, then the current literature falls short in that it only evaluates refugees flows in one direction: to the asylum country. Theoretically, refugees returning to their country of origin en mass should also increase violence. Applying the notion that refugees spread conflict to returnees, suggests a simple empirical relationship: y it = β 0 + β 1 R it + ε (1) where y it is the change in the violence in the country with the civil war, and R it is the number of refugees who return to the country. 1 In addition to the work on refugees, the civil war literature more broadly has found a number of characteristics that also explain the onset of civil war. Fearon and Laitin (2003) find despite the empirical regularity that countries with ethnic majorities and significant ethnic minorities more often go to war, much of the variation in civil war onset can be explained by poverty, rough terrain and large populations. Last, one might think that distance between countries matters, for the effect of spreading civil war. All of these considerations, suggests the addition of control variables: y ijt = β 0 + β 1 R ijt + H i β 2 + T i,t 1 β 3 + M ij β 4 + ε (2) where i index the country of origin, j indexes the country of asylum, and t is years. H are the time invariant characteristics ethnic and religious fractionalization and rough terrain, and T are the time varying characteristics, population and poverty. 2 M is the distance between countries. 1 In previous research y it would instead refer to the change in violence in the asylum country, rather than the country with the civil war. Similarly, in previous research R it would refer to the number of refugees coming into the asylum country, rather than returnees coming into their country of origin. 2 Lagging these measures by one year ensures that I do not simultaneously count changes in population due to refugee flows in either the population or the GDP per capita measure. 5

II: Instrumental Variable The problem with specifications (1) and (2) is that the traditional wisdom concerning refugee return suggests the opposite causal relationship between violence and returnees. Returnees likely wait until they believe it is safe before they return to their country of origin. In general refugee repatriation occurs after hostilities have ended, and violence has decreased. Less violence, makes it more likely that the refugees will be able to return safely, settle and establish a livelihood for themselves. For this reason we would expect that there would be negative relationship between violence and refugee return, not because returnees decrease the violence in their country of origin, but rather because the returnees are responding to, or anticipating, decreased violence and coming home. These are importantly different causal mechanisms and they need to be separated. In order to separate the impact of returnees on the level of violence from the level of violence spurring returns, I use an instrumental variable. A natural disaster in an asylum country would cause refugees to return to their country of origin. At the same time, controlling for the natural disaster occurring in the country of origin, there should be no reason to think a natural disaster in one country should effect a war in another, except through population flow. 3 Therefore, natural disasters would satisfy the exclusion restriction. Because the impact of refugees returning home in response to a natural disaster will vary by the number of refugees, I interact the incidence of natural disasters with the number of refugees from each country of origin in the asylum country in the year before the disaster. This suggests the following specification: First Stage: R ijt = γ 0 + γ 1 D j P ij,t 1 + γ 2 D j + γ 3 P ij,t 1 + ν i (3) Second Stage: y it = β 0 + β 1 ˆ Rit + H i β 2 + T i,t 1 β 3 + M ij β 4 + η i (4) 3 In section IV, I provide some robustness checks for assertion. 6

where D is the incidence of a natural disaster in which more than 7,500 persons were left homeless in the country of asylum, and P is the refugee population for each country of origin in the asylum country in the year before the disaster (potential returnees). Since natural disasters are not confined by national borders, it is possible that the same natural disaster occurs in the country of origin and the country of asylum. A natural disaster in a country with a civil war may cause an increase in hostilities, suggesting the need to control for natural disasters in the country of origin also. Armed conflict in the asylum country may also cause refugees to return, so, I control for armed conflict in the asylum country, as well. Finally, you might worry that a natural disaster might increase the number of people from the asylum country fleeing to civil war country, causing more than just the returnees to arrive. I control for refugee flows in the opposite direction, as well. Including the controls from equation (2), this lead to the following specification: First Stage: R ijt = γ 0 + γ 1 D j P ij,t 1 + γ 2 D j + γ 3 P ij,t 1 + ν i (5) Second Stage: y it = β 0 + β 1 ˆ Rit + H i β 2 + T i,t 1 β 3 + β 4 C it + β 5 A jt + β 6 S ijt + β 7 M ij + η i (6) where C is an indicator variable for the incidence of natural disasters in the country of origin, A is an indicator for the incidence of war in the asylum country, and S is the presence of refugees who are asylum country nationals in the civil war origin country. III: The Data In order to study the relationship between refugee flows and civil conflict, I construct a cross country dataset with an origin-asylum-year unit of analysis from 1976 to 2005. I combine data from the United Nations High Commission for Refugees (UNHCR) on refugee flows with data from the Peace Research Institute Oslo (PRIO) Armed Conflict Dataset. For the incidence of natural disasters, I make use of the Centre for Research on the Epidemiology of Disasters (CRED) International Disaster 7

Database (EM-DAT). UNHCR Refugee Data The UNHCR counts refugees as those who are recognized under the 1951 UN Convention Relating to the Status of Refugees, its corresponding 1967 Protocol, and the OAU Convention Governing the Specific Aspects of Refugee Problems in Africa. Refugees, according to the 1951 Convention are people who [are] unable or unwilling to return to their country of origin owing to a well-founded fear of being persecuted for reasons of race, religion, nationality, membership of a particular social group,...or political opinion (United Nations, 1951). UNHCR also sometimes counts some refugees recognized by individual countries of asylum or under national laws. To be clear, all of the persons counted by UNHCR are fleeing some kind of violence or persecution. Their definition works well for me because the vast majority of UNHCR refugees who are counted are those fleeing a country with some kind of armed conflict. UNHCR s Statistical Online Population Database provides counts of refugees in each country of asylum by each country of origin. UNHCR s data is compiled from multiple sources and counts refugees primarily using registries. 4 Counts are available from 1975 to 2008 (UNHCR, 2011). This time frame is long enough to consider a variety of conflict across the world. When numbers of refugees are not available for a country-pair-year, I substitute numbers of UNHCR assisted refugees. Because the return of refugees data are available for far fewer years and countries, I define returnees based on the change from the preceding year refugee population. The refugee return variable is constructed from the year over year difference of refugee counts by origin and country of asylum from UNHCR. If the number of refugees decreased over the previous year in a country of asylum from a particular country of origin and the total number of refugees from that country of origin 4 UNHCR does not include data from most refugees in Palestinian camps as this is the domain of United Nations Relief and Works Agency for Palestinian Refugees in the Near East (UNRWA) and likely undercounts refugees not living in camps, so I exclude Palestine. For a more detailed account of definitions and limitations, see http://www.unhcr.org/45c06c662.html. 8

decreased from the previous year, the refugee return variable is the change in the number of refugees for the origin, asylum pair. If the number of refugees in the asylum country from the country of origin remained the same, increased, or the total number of refugees from the country of origin remained the same or increased, the variable is coded zero. 5 PRIO Conflict Data With the data on refugees, I pair information on internal armed conflict from PRIO (Gleditsch et al. (2002), Harbom and Wallensteen (2010)). PRIO 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 (Harbom and Eriksson, 2009). The dataset also makes a distinction (by intensity) between an armed conflict and a war by battle deaths. Over 1,000 battle deaths per year is a war, and fewer battle deaths is a conflict (Harbom and Eriksson, 2009). Specifically, I count the number of internal (type 3 and type 4) armed conflicts per country, per year and by intensity. Then I construct a change in violence variable. The variable is coded as 1 if one of two conditions is met: (1) if the number of internal armed conflicts increased, or (2) if the number of internal armed conflicts considered wars (at least 1,000 battle deaths per year) increased. If the number of internal armed conflicts and those considered wars remained the same, then the change in violence variable is coded zero. If at least one of these elements decreased and all of other elements remained the same, the change in violence variable is coded -1. Natural Disasters EM-DAT counts disasters if one of four conditions is met: (1) at least 10 deaths were reported; (2) at least 100 persons were affected; or there was (3) a declaration of national emergency or (4) a call for international assistance (EM-DAT, 2011). Natural disasters include earthquakes, 5 I include the condition of an overall decrease in refugees from the preceding year, to prevent counting refugees who moved from one asylum country to another asylum country. 9

volcanic eruptions, dry mass movement (including landslides), storms, floods, wet mass movements (including avalanches and mudslides), extreme temperature (hot and cold), droughts and wildfires. They further count the number of persons left homeless by the event. The natural disaster indicator variable is coded 1 if at least 7,500 people were left homeless because of at least one natural disaster in a given country for the given year, and zero otherwise. Controls I draw on other scholars work on civil war for control variables. I make use of Fearon and Laitin (2003) data for the percent mountainous terrain of the country. I use (Alesina et al., 2003) data for ethnic and religious fractionalization. I use Gleditsch (2002) for population and real per capita GDP. 6 I use the Mayer and Zignago (2010) for a measure of distance between country pairs. Last, the asylum country war indicator is constructed using PRIO type 2 (interstate), type 3 and type 4 (internal conflicts). 7 The origin country natural disaster indicator is constructed just like the one for the asylum country. The exchange of refugees in both directions is coded 1 if there is an UNHCR year over year change in refugees from the country pair, in the opposite direction. Summary Statistics Table 1 and 2 below provide descriptive statistics by level of violence in the civil conflict of the origin country for the data used. IV: Results Table 3 displays the results of the OLS regressions from models (1) and (2) and a revised model (2) that includes a control for institutions, Polity IV (Marshall and Jaggers 2010). 8 6 I use this data because it is compiled from multiple sources and interpolated, reducing the number of missing data. The shortcoming is the most recent beta version of the data is only available through 2004. Gleditsch (2002) for a complete discussion of the merits of the data. 7 There were no Type 1 (Colonial Wars) for the years of interest. 8 Polity IV does not provide data for a host of countries that its makers have determined are in transition or occupied. Since wars are often associated with transitions and occupations, this is less than desirable. It drops observations in violent country years, disproportionately countries of particular interest including Afghanistan, Iraq, Lebanon, and Uganda among others. 10

Table 1: Summary Statistics Full Sample Full Sample Mean SD N Refugee Return (1,000) 0.863 18.197 30,781 Origin Country Natural Disaster 0.147 0.354 30,781 Log % Mountainous Terrain 2.352 1.370 30,781 Ethnic Fractionalization 0.543 0.251 30,781 Religious Fractionalization 0.437 0.229 30,781 Polity Score (-10, 10) 0.203 6.101 29,439 Population, Lagged, (1,000) 49.9 160.9 30,781 Real GDP per Capita, Lagged (1,000) 3.980 4.848 30,781 Asylum Country War Occurring 0.131 0.338 30,781 Natural Disaster 0.089 0.285 30,781 Country Pair Refugees in Both Directions 0.066 0.248 30,781 Distance between Countries (km) 5,293 3,974 30,781 Table 2: Summary Statistics by Violence Decrease No Change Increase Mean SD Mean SD Mean SD Refugee Return (1,000) 1.616 23.280 0.690 16.000 1.299 25.368 Origin Country Natural Disaster 0.186 0.389 0.145 0.353 0.116 0.320 Log % Mountainous Terrain 2.600 1.313 2.254 1.377 2.793 1.259 Ethnic Fractionalization 0.614 0.231 0.521 0.253 0.628 0.218 Religious Fractionalization 0.446 0.218 0.435 0.234 0.444 0.204 Polity Score (-10, 10) 0.366 4.917 0.474 6.354 1.197 5.016 Population, Lagged, (1,000) 55.4 152.2 47.8 162.9 58.7 154.2 Real GDP per Capita, Lagged (1,000) 2.478 3.257 4.435 5.140 2.325 3.131 Asylum Country War Occurring 0.150 0.357 0.124 0.330 0.162 0.369 Natural Disaster 0.098 0.297 0.087 0.281 0.095 0.294 Country Pair Refugees in Both Directions 0.084 0.277 0.061 0.240 0.079 0.270 Distance between Countries (km) 5,124 3,706 5,346 4,053 5,089 3,653 N Decrease = 3,548, N No Change = 23,891, N Increase = 3,342 except for Polity for which N is 3,417, 22,978, and 3,044 for each category respectively. 11

Table 3: Ordinary Least Squares (DV - Change in Violence) Model (1) Model (2) Model (2 ) Coeff ( ˆβ) Coeff ( ˆβ) Coeff ( ˆβ) SE SE SE Refugee Return (1,000) -0.00014-0.00014-0.00002 0.00013 0.00012 0.00011 Asylum Country War 0.01080 0.00674 0.00737 0.00715 Flows in Both Directions -0.01494-0.01311 0.01185 0.01034 Significant Natural Disaster (origin) -0.07593-0.04545 0.05218 0.04568 Religious Fractionalization (origin) -0.00721 0.01929 0.03909 0.03037 Ethnic Fractionalization (origin) 0.03278 0.01984 0.03561 0.02814 Log % Mountainous Terrain 0.01222 * 0.00984 ** 0.00700 0.00461 Population, Lagged (1,000) 0.00006 0.00004 0.00004 0.00004 Real GDP per Capita (1,000) 0.00038 0.00199 ** 0.00092 0.00086 Polity Score -0.00228 0.00139 Distance Between Countries -7.41E-08-2.71E-08 6.35E-07 6.29E-07 Constant 0.00338-0.04353 * -0.05699 ** 0.00750 0.02290 0.02255 Number of Observations 44,712 30,781 29,439 Clusters 200 153 153 Standard errors are clustered at the origin country level; *** p < 0.01; ** p < 0.05; * p < 0.1 12

These reduced form results are suggestive of the traditional understanding in the literature of the relationship between refugee return and violence. Namely, refugees return to their country of origin when they perceive it is safe to do so, and refugee return is correlated with decreases in violence. Specifically the coefficient on the return relationship to violence is negative in the simplest regression (1) and insignificant. On the whole, the impact of refugee return is consistently small. The return of 100,000 people is only associated with a 1.4% decrease in violence. While the controls add some precision to the regression, the estimates from model (2) still suggest a negative but not significant relationship. The return of 100,000 people is still associated with a 1.4% decrease in violence. Adding, the polity control, does not change the sign, although the estimate drops considerably and on the whole, there is little precision in the estimates. At first glance, this would lead us to the conclusion that refugees return is associated with a decrease in violence, but ambiguously so. Table 4 displays the results of the two stage least squares regression that incorporates the natural disaster instrumental variable. This model suggests a decisively different relationship. Table 4 shows first that the incidence of natural disaster interacted with the potential return population is a valid instrument for refugee return, with a relatively high F-stat. Second, this model reveals a positive relationship between increased violence and refugee return. The last two columns of Table 4 ((5) and (6)) make use of all the controls. The controls increase our estimate of the impact of returning refugees. The impact of 100,000 returned refugees is a 3% increase in violence, and the effect is statistically significant. The difference in results between the OLS regression and the 2SLS regression are remarkable. Not only does the sign of the effect change when the endogenous relationship between conflict and refugees flow is accounted for, but the 2SLS estimates are twice the magnitude of the OLS estimates. These 2SLS estimates allows us to identify that returnees not only increase violence, but the impact of population flows is larger than the OLS regression suggests. 13

Data Problems These findings are compelling but a number of caveats should be noted. There are a few exclusions from the cross-country dataset that are worth noting. First, information about Palestinian refugees are all together excluded from the data. The UNHCR numbers on Palestinian refugees are incomplete because they do not include any of the refugees under the protection of the UNRWA. Second, UNHCR counts refugees from Western Sahara and from Tibet separately from any recognized country. Since these regions are not recognized generally in the international community as states in their own right, their unit of analysis differs from the rest of the dataset. Furthermore, there is not sufficient matching data to include them in my analysis. We have no reason to believe that the causal relationship that I have explored above would be different for these areas. However, while it may be the case that the relationship between refugee return and increased violence is the same for Palestinian, Tibetan or Sahrawi refugees, these represent some of the most long standing refugee crises outside of sub-saharan Africa. A micro study of any one of these groups of refugees, their movement and violence, would be a good compliment to the study presented here. A second consideration with my analysis is that one might think that there is something unaccounted for about natural disasters that causes increases in civil war violence in a neighboring country even in the absence of refugee flows (in other words, that natural disasters do not serve as a valid instrument for the analysis). If this were the case, natural disasters could account for the change in violence, but have nothing to do with returnees inciting the violence. To test this concern I constructed a dataset of country-pair-years in which the two countries are contiguous (Mayer and Zignago, 2010), the first country has had a civil conflict in the preceding year, and the contiguous country did not receive refugees. With this set of countries, I examine the relationship between the contiguous country having a natural disaster and the change in intensity of the civil conflict in the country of origin controlling for both war in the contiguous country and a natural disaster in the 14

Table 4: 2SLS (IV: Natural Disaster, DV: Change in Violence) (3) 1st Stage (4) 2nd Stage (5) 1st Stage (6) 2nd Stage Coeff (ˆγ) Coeff ( ˆβ) Coeff (ˆγ) Coeff ( ˆβ) SE SE SE SE Refugee Return (1,000) 0.00030 0.00030 *** 0.00028 0.00010 Nat Disaster X Potenial Returnees 0.0534 * 0.0560 * 0.0319 0.0319 Natural Disaster (Asylum) -0.7234 ** -0.8712 ** 0.3420 0.4255 Potential Returnees (1,000) 0.0781 *** 0.0760 *** 0.0278 0.0283 Flows in Both Directions 2.0072-0.0168 1.3098 0.0120 Significant Natural Disaster (origin) 0.0579-0.0760 0.2866 0.0521 Religious Fractionalization (origin) 0.5482 * -0.0072 0.3276 0.0388 Ethnic Fractionalization (origin) -0.4010 0.0326 0.4599 0.0353 Population, Lagged (1,000) -0.0004 0.0001 * 0.0003 0.0000 Real GDP per Capita (1,000) -0.0272 0.0004 0.0207 0.0009 Asylum Country War -0.0437 0.0100 0.2574 0.0072 Log % Mountainous Terrain -0.0454 0.0121 * 0.0427 0.0069 Distance Between Countries 3.56E-06-1.34E-08 1.37E-05 6.33E-07 Constant 0.1289 0.0031 0.2361-0.0436 * 0.1038 0.0021 0.2932 0.0227 F-Stat 333.38 376.68 Number of Observations 44712 30781 Clusters 153 153 Standard errors are clustered at the origin country level; *** p < 0.01; ** p < 0.05; * p < 0.1; 2SLS is calculated using the ivreg2 module (Baum et al., 2002) in STATA 11.2 15

origin country. The results of this comparison are in Table 5. Table 5: Contiguous Countries, Impact of Natural Disasters DV: Change In Violence Coefficient SE Natural Disaster (Contiguous Country) -0.0600 0.0559 Natural Disaster (Origin Country) -0.0462 0.0572 War (Contiguous Country) 0.0484 0.0328 Constant 0.1959 0.0387 *** Number of Observations 2112 Clusters 80 Standard errors are clustered at the origin country level Table 5 suggests that there is a negative, but not significant, relationship between contiguous countries natural disasters and civil conflict, when there are no refugee flows. This would suggest that my instrument, using natural disasters may be biased. However, it would be biased downward since the relationship between the contiguous country s natural disaster and the civil conflict intensity is negative and therefore makes my causal story more difficult to prove. Since I still find a positive and significant result despite this potential bias, it is not a problem for my results. Two final considerations about the analysis remain. First, since a large number of the countries have natural disasters and refugees are thought to be a contagion one might be concerned that it is the interaction between an asylum country natural disaster and an asylum country war that prompts refugees to return. I have not yet thoroughly explored this possibility. Second, my measure of changes of violence is asymmetric. It is coded as increasing when a new conflict starts (25 battle deaths) and when a conflict becomes a war (1,000 battle deaths). These are discrete, asymmetric cutoffs, and we have no reason to believe that they are qualitatively the same. Again, though I have not determined how best to overcome this problem. 16

V: Theory for Explaining why Returnees Spark More Violence The literature on how refugees spread civil war as its stands has two basic explanations for the phenomenon. (1) Refugees help to expand a rebel network and thereby spread civil war, suggesting that the subset of refugees who are militarized are largely responsible for the spreading of civil war. (2) The influx of refugees make resources more scarce. The increased scarcity of resources in already precariously situated countries yields greater instability and violence. The findings here are certainly not consistent with the first explanation of militarized refugees. There is no reason to believe that militarized refugees would disproportionately respond to natural disasters by returning home. If anything, militarized refugees should be more strategic, waiting until they have an operational advantage to return home as rebels and continue their fight. The second explanation, that refugee flows lead to increased scarcity of resources and presents additional burdens to the government, thereby cause instability and violence, squares better with the findings here. Still though the mechanism is not clear. How do we get from refugees returning home to more violence? I proceed by drawing on literature on price shocks and violence and literature on rebel recruitment, to examine the causal chain. Price Shocks There is a good bit of empirical economics and political science literature linking growth and political stability. Causal identification is problematic though because in the absence of well defined theory, it is not clear whether it is economic growth that causes political stability or if political stability provides the foundation for economic growth. In order to to identify causation economists and political scientists have made use of exogenous shocks to establish direction. It has become clear that positive economic shocks (primary commodity exports price increases) are followed by political stability and peace (Deaton, 1999). Meanwhile negative economic shocks (primary commodity export price decreases, insufficient rainfall) predict political instability and political exit (coups, 17

forced resignations - Deaton and Miller (1995) and Dell et al. (2008)) and civil war onset (Miguel et al., 2004). Recruitment Dube and Vargas (2008) and Nillesen and Verwimp (2009) embed their investigation of economic shocks and civil war in the context of rebel recruitment. Dube and Vargas find that in Colombia when prices of coffee - a labor intensive sector - decrease, violence increases. When prices of oil a capital intensive sector - increase, violence also increases. They argue that the negative coffee price shock - violence relationship is evidence of rebel recruitment mechanism. Specifically, when coffee prices decrease, the wages in the coffee industry go down. The Colombian who is choosing between fighting and working, has a relatively better offer to fight. Those at the margin choose to fight rather than work in the coffee industry. With more fighters, violence increases. Nillesen and Verwimp investigate rebel recruitment activities in Burundian villages in response to shocks. They find a similar relationship between recruitment activities and negative weather shocks. Following insufficient rains in Burundian villages, recruitment activities increase. That is, in the absence of agricultural work, rebel recruitment increases. The Returnee Shock Mechanism Drawing on the theoretical foundation on the rebel recruitment in Dube and Vargas, I investigate a rebel recruitment mechanism set in motion by an influx of returnees to their country of origin. Thinking of refugee return as the shock of interest, I embed the choice to become a rebel in a labor market framework. An influx of returnees represent a food demand shock. In response to the increased demand for food, food prices increase. In the absence of a corresponding wage increase, the increase of food prices are tantamount to a real wage decrease. Assume that a rebel organization benefits from economies of scale or alternatively, that the rebel organization can benefit from food appropriation. In the midst of the food price increase, the net benefits of working have gone down, but the net benefit of fighting has gone down less. The marginal worker becomes a fighter. The rebel organization is able to expand and violence increases. 18

The mechanism critically hinges on the assumption that the rebel group can provide sustenance to its recruits at a lower cost than individual workers can at market prices. This make sense if we think about organized groups that are structured like a military. The rebel group provide for food, shelter and other basic needs of their recruits at a base of operation. If the rebel group is well organized and sufficiently large, they will be able to take advantage of wholesale rather than retail prices, reducing the cost of providing sustenance by some factor. The other plausible interpretation of this assumption is that the rebels (either as a group or individually) can appropriate some food, rather than purchasing it all directly (see Azam (2006) for theory and discussion of rebel groups and appropriation). In this case, the cost of food to the rebels is reduced by some factor. In the simplest of models, before the returnee shock, each person has the choice between working where the utility of labor is wages received less the cost of sustenance (u l = w l c i ) and fighting, where the utility of fighting is u f = b f. Modelling the rebel offer as one with direct provision of food, b f can be decomposed into some wages, w f and the cost to the rebel organization of providing sustenance for 1 person, c r = αc i, where 0 < α < 1, is economies of scale of the parameter. In equilibrium, the marginal person is just indifferent between working and fighting. That is, at food price p b (b for before influx), w m p b = w f or alternatively, w m = w f + αp b. The simple model gives us the basic intuition we need for understanding how refugee return could increase rebel forces. The influx of returnees will shift the demand curve up shifting the equilibrium price from p b to p a (a for after), with p a > p b. Then, since w m p a < w m p b = w f and w m = w f + αp b < w f + αp a, the marginal person who was working before will choose to fight. The bigger the shift from p b to p a the greater the number of people who will choose to fight. Furthermore, the smaller α is, the greater the number of people who will choose to fight. When is the Influx of Returnees a Shock? Initially, it is not clear exactly why we should expect that the influx of returnees should represent a food price shock, shock being the operative part. If returnees arrive in sufficiently small numbers, 19

there may be essentially no change in the price of food. If returnees arrive slowly, we should expect that the gradual increase in demand for food to be met with a matching gradual increase in the price of food. While this may correspond to gradual increase in the number of rebels, it is not clear that a small change in rebel forces should be linked to remarkable changes in violence. Even more to the point, if there is planning for the repatriation of returnees, there may even be food aid. In these cases, we would expect to see little rise in food prices, and perhaps even a decrease in food prices if food aid is delivered not only to returnees but those who continued to live in their homes in the midst of civil war. Rather, there seem to be two possibilities for how returnee influx can cause a price shock that results in a notable increase in rebel recruitment and violence. The first is relatively straight forward - a particularly large influx of returnees in a short amount of time for which planning was not possible or not sufficient. This seems most plausible when refugees have to travel a relatively short distance to return home, and when the institutions and infrastructure to facilitate planning for the refugees are relatively poor. It is in this case the returnee influx might best mirror the refugee flows that are the subject of the civil war contagion literature (see also Gleditsch (2007) and Gleditsch et al. (2008)). The other possibility for how returnee influx can cause a price shock and rebel recruitment is elucidated when we embed the refugee return into a poverty development labor economic model. Dasgupta and Ray (1986) first consider the implications of nutrition on labor in the developing economies to investigate inequality. I borrow this framework, which has two critical insights (see Ray (1998), chapters 8 and 13 for a more developed explanation of the theory). The first is an S-shaped capacity-income curve. The second is noting how the S-shaped capacity-income curve, translates into and a disjoint labor supply curve, thereby allowing for the possibility of involuntary unemployment in the labor market. Involuntary unemployment in the midst of an expanding pool for labor, in our case, translates into an even larger pool of recruits for the rebels. 20

Specifically, the nutritional requirements of doing labor are such that at low levels of income, people are unable to work, without tradeoffs on physical capacity to do the work. After a certain threshold, income barely improves a persons nutritional capacity to work (490). What this means for much of the developing world is that there is a range of very low wages at which only people with other sources of income are able to accept a wage offer. After a certain threshold of wages, nearly everyone is able to work. This makes for a disjoint labor supply curve that is directly related to the cost of providing sustenance for labor (492). The disjoint labor supply curve, allows for the possibility that the market does not clear, and forces individuals into involuntary unemployment. That is, the equilibrium wage is a wage at which some volume of people would work, but the demand at that wage is less. The lucky people work, and the unlucky people are unemployed. The involuntary unemployment dynamic in labor markets in particularly poor countries, provides another avenue for the influx of returnees to represent a shock to food prices that translates into large rebel recruitment. Suppose before the influx, the labor market is clearing. The influx of returnees shifts the demand for food up, causing an increase in the price of food. Without a corresponding shift in supply, the increase in food prices is effectively a decrease in real wages. The problem is that the vast majority of people are incapable of working at lower wages. Across the board, they must demand higher wages. That is, the decrease in real wages translates into an upward shift in the labor supply curve. Since the rebel organization can provide food directly and at lower cost this differentially improves the rebel offer. Furthermore, if the shift in the labor supply curve differs enough from the fighter supply curve shift, it could mean that the market for labor does not clear, but market for fighting does. Some workers are forced into unemployment. In this scenario, since the rebel group pays less for food, and provides food directly to its soldiers, the rebels are able to recruit some segment of the now unemployed, providing an even larger recruitment pool. With more rebel fighters, violence increases. It is worth noting that for the food price-rebel recruitment mechanism to be operative, it is not 21

necessary that the returning refugees enter the labor market, rather it is just necessary that they enter the food market as consumers. While it is plausible that some refugees returning home may enter the labor market, the food price-rebel recruitment mechanism operates through the already existent labor market. It is not necessary that the rebels are recruiting from the returning refugees. Rather since most refugees tend to be elderly, women and children (?), they are more likely recruiting from the workers, already in the country, who take the relatively better offer to fight over a market job. Or in a subsistence economy case, the rebels are recruiting from the labor force that has been forced into unemployment because with the food price shock, the labor market is not clearing at the efficient wage. While the role of a nutritional constraints, and the disjoint supply curve seem particularly salient conditions for many countries in which there is civil war, I have found it difficult to identify when this more nuanced theory is in fact at work, empirically. It seems most likely to be applicable in developing countries, with subsistence economies, with relatively undifferentiated labor markets. In the absence of data on variation in these factors, I focus on testing the impact of a large flow of returnees home in countries in which the World Food Programme operates. In using countries in which the WFP operates, I make use of the subset of countries in which the subsistence economy, undifferentiated labor market assumption are most likely to hold. VI: Food Prices -Rebel Recruitment, an Empirical Test Empirical Model If refugee return increases violence in countries with civil wars through the the food price - rebel recruitment mechanism, the first link in the causal chain to test is whether refugee return is associated with rising food prices. Since we are particularly interested in food prices, but not a general rise in prices, we are interested in food price increases net of inflation. This suggests a fairly simple model: p = β 0 + β 1 R + β 3 I + ε, (7) 22

where p is the change in food prices, R is the number refugees who returned, and I is the inflation rate. There are two problems with this simple model. First, the empirical literature on violence and economic decline is rather unequivocal on the problem of endogeneity, and food price increases net of inflation are probably indicative of economic decline. The behavior of refugees and migrants more generally is likely endogenously determined too. If refugees anticipate that it is safe to return home, they are more likely to do so. On the converse, if refugees anticipate that if they return home they will be unable to provide for their basic needs once returning home, they are likely to stay in an asylum country. Without some source of exogeneity, it is impossible to identify which of these dynamics is occurring in the data. The second problem with the simple model is more pragmatic. Repatriation is often organized by international organizations and accompanied with development aid. (UNHCR explicitly includes providing the most urgent of needs for returnees and coordinating with international and community organizations on development activities in its Repatriation Handbook (UNHCR (1996), section 6.4 ). These voluntary organized repatriation activities in anticipation of decreased violence are not the returnee shocks theoretically modelled and are likely not what causes renewed violence. Rather, it is larger spontaneous returnee events that increase violence. If there is renewed violence because of food price increases and rebel recruitment, it should occur in these unplanned events, and particularly in the midst of poor institutions that cannot accommodate these changes quickly. To deal with the problems of endogeneity, I introduce exogeneity in the form of a natural disaster in asylum countries. To model the returnee shock I instrument for refugee return with an interaction of natural disasters in the country of asylum (an indicator variable) and the refugee population from the country of origin in the country of asylum. Since natural disasters are not confined by country borders, we might think that a natural disaster in the country of origin caused food prices to increase (rather than refugee return) and then the natural disaster is the reason for violence. 23

I include a control, an indicator for a natural disaster occurring in the country of origin. Finally, since the literature contends that refugees (not returnees) are the source of spreading conflict, I include a control for when the origin country of interest is also an asylum country. Having accounted for refugee flows into the country of origin, the presence of natural disasters, in addition to inflation and institutions, I argue we have good reason to think that the exclusion restriction is satisfied. This gives us the following Two Stage Least Squares (2SLS) model: First-Stage: R it = γ 0 + γ 1 P it 1 + γ 2 D jt + γ 3 D jt P it 1 + η (8) Second-Stage: p it = β 0 + β 1 ˆ Rit + β 2 I it + β 3 D it + β 4 S ijt + β 5 E it + ε i (9) where i is the index for the country of origin, j indexes the asylum country and t is the index for the year. P is the count of refugees from a country of origin in a country of asylum. D is an indicator for a natural disasters occurring in a country. S is an indicator for when refugees from the asylum country are in the country of origin, and E is a measure of the institutions environment. Thinking about the influx of refugees as not increasing food prices gradually, but constituting a shock, suggests two other likely considerations. First, we would expect the distance the refugees travel to lessen a shock. The more distance they travel, the more likely their arrival is gradual. With this concern in mind, I also look at sub-national food price data in a number of countries and control for distance by road from the closest border of the asylum country border to the food market. For this purpose I investigate a market town-asylum country-year unit of analysis in the following 2SLS model: First-Stage: R kt = γ 0 + γ 1 P it 1 + γ 2 D jt + γ 3 N jt P it 1 + η (10) 24

Second-Stage: p kt = β 0 + β 1 K kj + β 2 ˆ Rit + β 3 I it + β 4 D it + β 5 F ijt + β 6 E it + ε k, (11) where k indexes the market-town in the country of origin, and K is the distance by road from the market town to the closest border of the asylum country. To connect the rise of prices to violence, I investigate heterogeneous effects of refugee return on food prices. Namely, the theory would suggest that there be a positive relationship between returnees and prices when violence subsequently increases, and no relationship or a negative relationship when violence does not change or decreases. I connect the price increase to violence by splitting the sample into three categories, countries in which violence increases the following year (year t+1), countries in which violence remains the same, and countries in which the violence decreases. Additional Data To test the food price rebel recruitment mechanism, I analyze food price with the context of a subset of my returnee violence dataset. Adding food prices for where available from 1990 to 2008. Because I also have food prices at the subnational level, I build a second dataset analyzes food price changes on an market town - asylum country - year unit of analysis. I use data data from World Food Programme (WFP) (2011) and Food and Agricultural Organization of the United Nations (FAO) (2011) for food prices, and add make use of relevant controls. The WFP and the FAO collect data on monthly food prices at various markets throughout the world. For the countries of interest, the food prices are for staple goods like rice, corn, wheat, and sorghum. Prices are reported in local currency. I have selected the staple good for each market for which the most information is available. If data was not available from the World Food Program, I supplement with data from FAO. When retail prices were not available, I make use of wholesale prices. Since in many of the places of interest, food prices vary seasonally, I construct an average food price for each 25