Do Refugees Spread or Reduce Conflict?

Transcription

1 Do Refugees Spread or Reduce Conflict? Andrew Shaver and Yang-Yang Zhou July 11, 2017 Abstract As the number of forcibly displaced individuals around the world reaches record highs, understanding how their presence affects conflict is a major outstanding academic and policy question. Using new geo-coded data on refugee sites and conflict data at the subnational level from 1989 to 2008, we find no support for claims that refugee settlement increases the likelihood of conflict onset or incidence. Instead, we find that, conditional on there existing no other refugee sites present in the same country for a given year, provinces hosting refugee sites experience substantively large decreases in their likelihood of both conflict onset and incidence. We confirm these findings when examining heterogeneous effects of formal refugee camps and informal settlements. To exclude the possibility of unobserved confounders, we use placebo tests to show that there are no effects of future refugee sites on past conflict outcomes. To explore whether an influx of humanitarian and government assistance and associated infrastructural development to refugee sites may reduce the risk of conflict for these settlement areas, we carry out a nighttime lights analysis. Keywords: refugees, migration, civil war, conflict, subnational, GIS We thank Lamis Abdelaaty, Alexander Betts, David Carter, Miguel Centeno, Rafaela Dancygier, Benjamin Fifield, Kosuke Imai, Robert Keohane, Kabir Khanna, Matthew Kocher, Jacob Shapiro, the Imai Research Group, and participants from ISA 2014, APSA 2015, and the 2017 Harvard-MIT-Yale Political Violence Conference for helpful comments. We also thank Tsering Wangyal Shawa for providing GIS and data processing support. We are especially grateful to the United Nations High Commissioner for Refugees for providing data and expertise. Shaver and Zhou acknowledge funding from Princeton University s Institute for International and Regional Studies. Zhou also acknowledges support from the National Science Foundation (SES ). All errors are ours. Pre-Doctoral Fellow, Kennedy School of Government, Harvard University. ashaver@princeton.edu, URL: PhD Candidate, Department of Politics, Princeton University. Mailing address: 026 Corwin Hall, Princeton NJ yz3@princeton.edu, URL: yangyangzhou.com

2 1 Introduction Do refugees spread conflict to host countries and communities? This question has gained prominence as millions of refugees and migrants have crossed into Europe, largely driven by Syria s civil war. The fear that refugees spread conflict has been central to the political rhetoric of Far Right parties in Europe, the Brexit referendum, and Donald Trump s candidacy and presidency in the United States. These debates will likely intensify given that levels of forced migration have reached record highs every year since Currently, more than 65.3 million individuals around the world have been displaced involuntarily, a population so massive it would constitute the world s 24th largest country. Within this nation of the displaced, 21.3 million are refugees who have fled their countries of nationality (UNHCR, 2016). Although they are a population of policy concern and academic research, it is still unclear how they affect the security of host communities. Given refugees growing demographic size and political salience, understanding this dynamic is not simply a matter of academic inquiry but is critical to policymakers around the world facing the humanitarian implications of closing borders and building walls. This article examines whether the presence of refugee sites increases the likelihood of civil conflict onset or incidence within the areas, specifically subnational provinces, in which they are located. 2 While previous research links increased numbers of refugees to greater likelihood of civil conflict onset in the hosting country (e.g. Salehyan and Gleditsch, 2006), we re-examine this relationship at the subnational level using placebo tests to exclude the possibility that selection effects bias our results. Using new global data on geo-coded refugee sites provided by the United Nations High Commissioner for Refugees (UNHCR) and civil conflict events at the subnational A refugee is someone who owing to a well-founded fear of being persecuted for reasons of race, religion, nationality, membership of a particular social group or political opinion, is outside the country of her or his nationality and is unable, or due to such fear, unwilling to avail him- or herself of the protection of that country. (Article 1, The 1951 Convention Relating to the Status of Refugees). 1

3 level from 1989 to 2008, we find no evidence that refugees directly incite or sustain conflict. Instead, we find that conditional on no other refugee sites being present in the same countryyear, provinces that host refugees see their conflict onset likelihood cut by nearly half and their conflict incidence likelihood lessened by almost 5 percentage points. We call this the conditional risk reduction effect. We carry out a variety of robustness checks, which include examining heterogeneous effects of refugee camps versus settlements. In most cases, the conditional risk reduction effect we observe persists. In all cases, we find no evidence that refugee communities generate or sustain conflict. Because our findings contrast with much of the existing quantitative research on this topic, we also offer a theory to explain the conditional risk reduction effect; specifically, we ask whether the influx of humanitarian and host state assistance facilitates infrastructural development in and around settlement areas, ultimately reducing the risk of conflict within these affected areas. We use satellite night-lights data collected by the U.S. Air Force Weather Agency as a proxy for increased infrastructure in the area to test our proposed mechanism for the conditional risk reduction effect. Our results are inconclusive. We observe no effect of refugee communities on levels of local economic activity, either positive or negative. We conclude by suggesting additional avenues for future qualitative case study research; rather than maintain focus on cases where refugees have spread conflict, we identify those countries which experienced the largest conditional risk reduction effects during the study period. For clarity, the term refugee sites in this article refers to the geographic locations where refugees settle; it encompasses all formal refugee camps as well as informal refugee settlements (even those in which the UNHCR is not actively working or has a mandate); unless there is a concentrated site within a city, these do not include major cities where dispersed refugees may 2

4 have settled. By conflict, we are referring to instances of armed civil conflicts with a threshold of least 25 battle deaths (Gleditsch, 2002a); conflict onset refers whether a new conflict begins in a given province, which precludes subsequent years of continued conflict. conflict incidence indicates whether conflict is occurring in a province, which includes both the year of onset and subsequent years of continued conflict. Our geographic units of interest are provinces, the first order administrative units within each country. This study provides the first quantitative evidence at a global, subnational scale of the general effect of refugee communities on conflict likelihood within host countries and addresses both the direct and indirect effects of refugee presence this outcome. Our findings respond to a growing body of research seeking to identify the effects of migrants, and specifically refugees, on host countries and communities (e.g. Landau, 2008; Maystadt and Verwimp, 2014; Taylor et al., 2016; Zhou, 2017). A subset of this research focuses on conflict and violence outcomes. Beginning with Zolberg, Suhrke and Aguayo (1989), this literature highlights cases of violent conflict attributable to militarized refugees such as the Banyarwanda refugees in Eastern Congo in the 1960s and South African refugees during Apartheid (e.g. Lischer, 2006; Muggah, 2006). Generalizing these cases into a global phenomenon, scholars including Loescher (1992), Salehyan and Gleditsch (2006), Salehyan (2007), Salehyan (2009), and Choi and Salehyan (2013) find that refugees increases the likelihood of civil conflict and other forms of sub-state violence across international borders. They theorize that refugees do so by exacerbating ethnic tensions, intensifying economic competition with locals, and expanding insurgent social networks by transporting weapons and using camps to recruit and harbor combatants. However, these theorized mechanisms have not been directly tested. This paper proceeds as follows. Section 2 discusses the current debate in the literature re- 3

5 Refugee sites Onset 0 Onset.25 Onset.5 Onset.75 Onset 1 Figure 1: This map shows the propensity of civil conflict onset averaged over the timespan of our panel by province (red shading) as well as all refugee sites (blue points) in our data which were open at one point during that time. This map shows that refugee sites are concentrated near regions of conflict, which is unsurprising because refugees are almost always displaced due to conflict. Data source: UNHCR and Uppsala/PRIO. garding whether refugees affect conflict, our theoretical expectations, and testable implications. Section 3 introduces our novel data of geo-coded displacement sites and subnational conflict data. There, we introduce our primary testing strategies, including the role of placebo tests as part of our identification strategy. Section 4 presents the main results relating refugee sites to conflict onset and incidence and discusses robustness checks and heterogeneous effects of formal camps versus informal settlements. Section 5 explores our theory of how this conditional risk reduction effect might occur through the mechanism of increased infrastructure. Finally, section 6 concludes. 2 Theories about Refugees and Conflict 4

6 The primary aim of this article is to engage with recent quantitative and case study research linking refugees to the spread of civil conflict, using new global, subnational panel data. Our data casts doubt on the conventional wisdom and is generally inconsistent with much of the academic literature on refugees and conflict. We find no evidence that refugees either directly cause conflict where they settle or diffuse conflict to other areas of the country. Instead, we find, consistent with a subset of academic literature on refugees and conflict, that when refugees are concentrated in only one province, all else equal, that province experiences a sizable decrease in its likelihood of conflict onset and incidence. Here we briefly discuss this previous research, explain how our approach differs from previous studies, and lay out our hypotheses as testable implications. 2.1 Quantitative and Case Study Research Linking Refugees to Conflict. Security studies research finds that countries whose neighbors experience civil conflict are themselves significantly more likely to experience civil conflict. Such regional conflict diffusion suggests that incidents of sub-state conflict are not independent events with purely domestic causes (Gleditsch, 2002a; Hegre and Sambanis, 2006; Gleditsch, 2007). Several studies empirically show that the likelihood of conflict and terrorism increase in countries that host more migrants, and specifically refugees from nearby war-torn countries (Salehyan and Gleditsch, 2006; Salehyan, 2007, 2009; Choi and Salehyan, 2013; Milton, Spencer and Findley, 2013; Bove and Böhmelt, 2016). Notably, Salehyan and Gleditsch (2006) find that larger populations of refugees from neighboring counties increase the likelihood of civil war onset in the host country. At first glance, our own subnational province-year panel data show a similar association; figure 1 overlays refugees sites and propensity for conflict onset by province during the timespan of our study Refugee sites are clearly geographically concentrated in regions in which conflict is endemic. This is unsurprising insofar as refugees are most frequently displaced 5

7 due to conflict and tend to settle in neighboring countries that may also be affected by regional instability. Thus, our findings that areas hosting refugees are no more at risk of conflict, and in some cases, are less likely to experience conflict than non-hosting areas, seem to contradict these findings of Salehyan and Gleditsch (2006). However, upon further investigation, an extension of Salehyan and Gleditsch (2006) s analysis, which is discussed in more detail in section 4, suggests that this difference may be due to Cold-War versus post-cold War conflict dynamics. The recent quantitative research linking refugees to conflict builds upon rich case studies of groups of refugees spreading conflict, which migration scholars note are exceptional and extreme cases (Onoma, 2013). Several examples feature prominently: militarized refugees such as the Banyarwanda refugees in Eastern Congo in the 1960s; refugee flows from Liberia s two civil wars from 1989 through 2003 contributing to the insecurity of neighboring countries of Sierra Leone, Guinea, and the Ivory Coast; forced migration leading to conflict in several Balkan states throughout the 1990s; refugees from the Rwandan genocide in 1994 who were later involved in conflicts in the Democratic Republic of the Congo; and members of Al-Shabaab operating within the large Somali refugee camps in Kenya (Matthews, 1972; Zolberg, Suhrke and Aguayo, 1989; Loescher, 1992; Lischer, 2006; Whitaker, 2002; Muggah, 2006). While these studies identify individual cases of refugees causing new conflict, it is important to note that their goal is to provide explanations and context for how refugees can produce or otherwise facilitate conflict when they have, as opposed to determining whether refugees generally do. 2.2 Proposed mechanisms for how refugees might spread conflict Are these cases exceptional or a part of a larger pattern? Scholars are not unanimous in the position that refugee spread conflict generally. There is ample debate surrounding the various explanations for and against this statement. 6

8 Militarized refugees. First and most directly, refugee populations might harbor militants or become militarized themselves. Under this narrative, refugee camps serve as potential recruitment sites and rebel bases. Humanitarian assistance to these militarized refugees could facilitate conflict and inadvertently legitimize their cause (Lischer, 2006). Even if refugees themselves are not combatants, they can serve as auxiliary forces, assisting in the transfer of mobilization resources, such as munitions, to domestic opposition groups with whom they share an ethnic or ideological affinity (Zolberg, Suhrke and Aguayo, 1989; Lischer, 2006). Additionally, refugee populations may change ethnic balances within the host country, with this demographic shift provoking conflict. For example, Loescher (1992) points to refugee influxes shifting the ethnic composition of host countries, increasing the likelihood of conflict in societies with a precarious ethnic balance or pre-existing ethnic rivalries. Many migration scholars contend that refugee populations generally do not incite conflict, in large part because they consist overwhelmingly of civilian noncombatants who avoid or are physically incapable of engaging in many of the activities described above (Matthews, 1972; Whitaker, 2002; Onoma, 2013). Lischer (2006) discusses the specific circumstances for refugeedriven conflict; the types of refugee groups, along with how they are received and the role of external actors like the UNHCR, will determine whether they militarize. For example, situational refugees who flee due to general destruction from civil war rather than targeted persecution are unlikely to militarize. State-in-exile refugees, on the other hand, are a rare case of refugee populations who contain exiled former political and military leaders defeated in the conflict that displaced them; these refugee groups have the highest propensity for political violence. Drawing on empathy-driven altruism theories in psychology, Hartman and Morse (2017) find that during the Ivorian refugee crisis, Liberian hosts who were previously affected by violence were 7

9 willing to host more refugees, especially greater proportions of out-group refugees. Similarly, Hazlett (2013) finds that with Darfurian refugees in Chad, exposure to violence is associated with war-weariness and pro-peace attitudes rather than support for retribution and continued violence. Economic competition and aid resentment. Second, there also may be economic competition between refugees and natives for land, jobs, housing, and resources. Additionally refugees often live in settlements that are served by humanitarian organizations, providing assistance such as water and food that nearby native citizens might not have access to, which may lead to resentment and conflict. Some scholars find that refugees economic impacts are usually negligible, temporary, and/or ambiguous, depending on access to land and initial occupation (Chambers, 1986; Whitaker, 2002; Cortes, 2004; Landau, 2004) However, recent experimental and quasiexperimental studies find that refugees can bring human and physical capital and improve local host economies (Maystadt and Verwimp, 2014; Taylor et al., 2016). Challenging state legitimacy. Next, refugees may challenge state capacity and even legitimacy. Whitaker (2002) argues that the 1994 influx of Rwandan refugees sparked conflict in the Democratic Republic of the Congo but not in Tanzania because the Congolese government lacked political legitimacy and various elites used refugees to further their ethnic and political conflicts. Similarly, Onoma (2013) finds that in Africa, cases of local, host populations attacking refugees occur only when the host state is experiencing political instability and actively encourages these attacks. Nevertheless, Onoma (2013) contends that these cases of conflicts are rare, and when violence does occur, refugees are more often the victims of attack. Furthermore, refugee camps are often monitored by host states, the UNHCR, and other humanitarian agencies. Even weak states can to administer and run refugee camps by deploying 8

10 a mix of civilian officials such as police and gendarmes as well as military personnel (Onoma, 2013). To dissuade the possibility of rebel activities, the deployment of security forces to refugee settlements would increase security in the region (e.g. Malkki, 1995; Jacobsen, 2005). An accompanying influx of humanitarian organizations, personnel and their security, and aid might also facilitate infrastructural development of the area, ultimately rendering it more secure (Jacobsen, 1996; Bariagaber, 2006). Indirect insecurity. Finally, the mechanisms linking refugees to conflict might be indirect, meaning rather than refugees directly engaging in actions that increase conflict likelihood, their mere presence could affect factors elsewhere within the country that destabilize those areas. Weiner (1992) notes that refugees may destabilize the host country indirectly by imposing a heavy economic burden and straining the host country s social services, infrastructure, and ecological resources. Similarly, an influx of refugees might force a state to reallocate limited security resources to refugee-receiving areas, leaving those areas that experience the resource loss more susceptible to instability. Empirically, this suggests that the area hosting refugees might not itself bd more prone to conflict, but due to the refugees presence, conflict might develop elsewhere within the country. Thus, if the presence of refugees within a particular area of a country leads to either risk reduction or increase, this effect may be conditional on whether refugee populations exist elsewhere within the same country. 2.3 Hypotheses While we recognize that there exist several cases in which refugees directly or indirectly spread conflict to host communities for the specific reasons outlined by Loescher (1992); Lischer (2006) and others, we aim to determine whether these cases are exceptional or indicative of global 9

11 patterns. In response to research supporting both sides of the argument, we seek foremost to make an empirical contribution that provides an answer to this question. We test the following observable implications of refugees having a direct and/or indirect effects on conflict in host communities and countries: Hypothesis 1a (H1a): Provinces with refugee sites are more likely to experience conflict onset than those without refugees sites. (Direct Effect on Conflict Onset) Hypothesis 1b (H1b): Provinces with refugee sites are more likely to experience conflict incidence than those without refugees sites. (Direct Effect on Conflict Incidence) Hypothesis 2a (H2a): Non-hosting provinces in countries hosting refugee sites are more likely to experience conflict onset than those without refugees sites. (Indirect Effect on Conflict Onset). Hypothesis 2b (H2b): Non-hosting provinces in countries hosting refugee sites are more likely to experience conflict incidence than those without refugees sites. (Indirect Effect on Conflict Incidence) To preview our findings, we find that provinces hosting refugee sites are no more likely and, in some cases, much less likely to experience civil conflict than comparable provinces. As an initial exploration of this conditional risk reduction effect, rather than investigate the many mechanisms proposed above to explain how refugees might increase conflict, we probe one prominent mechanism for how refugees might lessen the likelihood of conflict. We theorize that the presence of refugees can prompt humanitarian organizations and host-country governments to increase their levels of assistance, security, and infrastructure to affected areas (Jacobsen, 1996, 2005; Bariagaber, 2006), and we test the following hypothesis: Hypothesis 3 (H3): Refugee-hosting provinces, conditional on there being no other refugee hosting provinces in the same country-year, receive more infrastructure (proxied by increased night-time lighting) compared to non-hosting provinces. (Effect on Infrastructure) 10

12 Proportion of Onset across Provinces Proportion of Incidence across Provinces Year Year All provinces Subset provinces Figure 2: The first figure shows the propensity of conflict onset over years for the full dataset of all countries (dashed line) and the subset of countries who have ever hosted refugees (dotted line). Averaged over time, the likelihood of conflict onset is about 3.9% for the full data and 7.58% for the subset. The second figure shows the propensity of conflict incidence over years again for the Full and Subset data. Averaged over time, the likelihood of conflict incidence is about 13.95% for the full data and 23.22% for the subset. 3 Data and Methods Our analysis uses a global panel dataset of 2597 provinces, formally, first-order administrative units (typically, states, provinces, governorates, etc., hereafter province for simplicity) from Thus, our units of analysis are province-years. To construct this data, we use a new geo-referenced dataset of the universe of displacement sites, which covers refugees and internally displaced persons (IDPs), provided by the UNHCR s Population and Geographic Data Department. This data encompasses all displacement sites, including those for/in which the UNHCR has no mandate/active presence. Table 2 in the Appendix shows, by region, the number of refugee sites broken down by camps and settlements. We also use the UCDP/PRIO Conflict dataset, 11

13 and various terrain and economic data described in further detail below. 3 While the Full Data covers all provinces in every country, we recognize that not all non-refugee hosting provinces may be appropriate counterfactuals. Since we aim to compare provinces that host refugee sites (treatment group) with provinces that could plausibly host refugee sites (control group), it may not be appropriate to include provinces in countries or island nations such as Japan, Luxembourg, and Turks and Caicos that are unlikely to ever host refugee sites, given that they have never done so. Thus, we also repeat all analyses with our Subset Data, which only includes provinces in the 67 countries that have at some point in our panel hosted a refugee site, subsetting out the provinces in 169 never-hosting countries. 3.1 Treatment variables: Presence of Refugee Site or Sites (binary) For each displacement site, our geo-referenced data includes the latitude/longitude coordinates, its creation date, its closing date (if closed), and whether the site is a formal camp or informal settlement. According to the UNHCR, the distinction between refugee camp and settlement varies by host-country. In general, however, settlements are informal refugee communities whereas camps receive humanitarian assistance, UNHCR presence, and host-country oversight. Unfortunately, for our study period, the UNHCR did not collect information on population, demographic or ethnic composition, reason for displacement, nationality of displaced persons, levels of aid or specific UNHCR involvement. We construct an indicator RefugeeSites for whether a province-year was hosting at least one refugee site. Again, refugee sites encompasses all formal refugee camps as well as informal refugee settlements; these do not include major cities even though some refugees were dispersed across 3 The dataset of georeferenced refugee and IDPs sites is not yet publicly available. However, all replication materials, including R code and data, will be made available via Harvard University s Dataverse. 12

14 some of these cities. Table 2 in the Appendix show the number of sites, broken down into camps and settlements, in each region. Most refugee-hosting provinces only have one site in any given year (850 province-year observations), while 562 province-years host more than one. However, because we do not have information on the population sizes of these sites, we do not believe using number of sites is preferable to an indicator of presence of at least one site. Table 3 in the Appendix show the number of refugee sites by number of province-year observations. There are several outlier province-years with many sites listed. Because UNHCR country offices classify sites differently for their own logistical purposes, the number of listed sites is not readily comparable between countries. Additionally, Lischer (2006) cautions against the assumption that larger refugee groups are more prone to conflict, as the presence of exiled military and political leaders are more likely to stir conflict. Thus, the findings of our subnational analyses speak only to the presence versus absence of refugees and not the number of refugees. To account for the possibility that refugees can produce conflict within provinces outside of their host province, we also construct OtherProvRefugeeSites, which indicates whether there exists at least another province in the same country-year hosting refugee sites. Our main treatment variables are these two variables as well as the interaction between RefugeeSites and OtherProvRefugeeSites. 3.2 Outcome variables: Civil Conflict Onset and Incidence (binary) Our outcome variables are binary indicators of whether a province-year experiences Conflict Onset or Conflict Incidence. To construct these variables, we use the UCDP/PRIO Conflict data, in which conflict is defined as at least 25 battlefield deaths (Gleditsch, 2002a). For Conflict Onset, which indicates the existence of new conflict, we code 1 for the year that a province-year 13

15 experiences new conflict, NA for subsequent years of ongoing conflict so these observations are dropped, and 0 for years of no conflict. For Conflict Incidence, which indicates whether there is conflict regardless of whether it is a new conflict, we code 1 when the province-year experiences conflict and 0 when it does not. Figure 2 shows the averages of our outcome variables across provinces by year. 3.3 Confounders First, to account for past outcomes, we include a one year lagged dependent variable Lagged Conflict Onset or Lagged Conflict Incidence, depending on the outcome of interest. Second, following the conflict diffusion literature, we construct War Border Province which is a binary indicator for whether, for each year, a given province borders another province at war regardless of whether the bordering province is of the same or different country. We also include War Border Country, a binary indicator for whether, for each year, a given province borders another country at war. Third, we include logged Population 4 and logged GDP per capita 5 of the province-year from Tollefsen, Strand and Buhaug (2012b) because population density and the size of the economy of a province may affect both inward migration and conflict likelihood. Not all years have data for these two indicators; when data is missing, we linearly impute. 4 Data on province-level population were similarly extracted from PRIO GRID, which, derived it grid-level population measures from CIESIN (2004), which is produced by U.S. National Aeronautics and Space Administration (NASA) Socioeconomic Data and Applications Center. This data similarly cover the years 1990, 1995, 2000, and The CIESEN effort consists of combining these dataset with particular models to estimate population levels at a sub-national grid level. 5 Data on province-level gross GDPpc were extracted from Tollefsen, Strand and Buhaug (2012a), which in turn derives its data from Yale University s G-Econ Project (Nordhaus et al., 2006), a global [geophysically scaled economic] data set on economic activity for [27,500] terrestrial grid cells [for the years] 1990, 1995, 2000, and covering virtually all terrestrial regions of the world. Specifically, gross domestic product per capita data are produced by scholars using a variety of data on Regional gross product, Regional income by industry, Regional employment by industry, and Regional urban and rural population or employment along with sectoral data on agricultural and non-agricultural incomes. 14

16 Because the ethnic makeup (and diversity) of an area may affect both conflict likelihood and refugee settlement patterns, we use the GeoEPR dataset to include the number of Ethnic Groups in the province. The 2014 version we use codes the settlement patterns of politically relevant ethnic groups in independent states with more than 500,000 inhabitants from based on the group list in the Ethnic Power Relations dataset version 2014 (Wucherpfennig et al., 2011). This data is expressed in shapefiles (which represent ethnic groups individual settlement patterns); thus specific time-varying boundaries of ethnic groups can be matched to the province boundaries that we adopt in this analysis. TerrainSD takes into account the level of terrain ruggedness; specifically it denotes the standard deviation in the absolute elevation change of every one-kilometer grid relative to all contiguous one-kilometer grids within each province (Shaver, Carter and Shawa, 2016). 6 Research links conflict to ruggedness (e.g. Fearon and Laitin, 2003). Also, a main factor of refugee settlement is considerations of terrain and distance to border (Jacobsen, 1996; Jacobson, 1996; Bariagaber, 2006). We also include ProvSize (sq km) to control for the size of the province. The number of ethnic groups, terrain ruggedness, and province size all vary by province but do not change over time. Next, we include BorderDistance, the distance (km) between the province s centroid to the international border. We do so because refugees tend to settle around international borders and such areas are likely to be more susceptible to conflict diffusion from neighboring countries. We also include CapitalDistance, the distance (km) between the province s centroid to the country s capital, because capitals are typically centers of state capacity, which also can affect 6 Shaver, Carter and Shawa (2016) use absolute elevation data generated with satellite technology at the 30 arc second (1x1 km) square level to calculate 30 arc second ruggedness measures. Specifically, they measure the average elevation [change] of [a given 30 arc second] area differs from all those of neighboring 1 km 2 areas. Because they use satellite data, they are able to measure all terrestrial areas of the planet. 15

17 refugee settlement and conflict propensity. These variables are also time invariant. Furthermore, our UNHCR dataset of displacement sites also indicates where internally displaced persons (IDPs) settle. We, therefore, include a binary indicator of whether a province-year hosts one or more IDP sites, which includes both formal IDP camps and IDPS living outside of camps in informal settlements. Finally, we include country and year fixed effects. Table 1 in the Appendix shows the summary statistics of our data. 3.4 Estimation Formally, we use the following logistic regression model for our main analysis to test whether provinces hosting refugee sites have an increased likelihood of conflict onset or incidence: P r(y i,t = 1) =logit 1 (β 0 + β 1 RefugeeSites i,t + β 2 OtherP rovrefugeesites i,t + β 3 RefugeeSites i,t OtherP rovrefugeesites i,t + β 4 Y i,t 1 + β 5 W arborderp rovince i,t + β 6 W arbordercountry i,t + β 7 P opulation i,t + β 8 GDP pc i,t + β 9 EthnicGroups i,t + β 10 T errainsd i + β 11 P rovsize i + β 12 BorderDistance i + β 13 CapitalDistance i + β 14 IDP Site i,t + γ c[i] + λ t + ɛ i,t ) in which i denotes provinces, t denotes year, and c[i] denotes countries. Y i,t is the binary outcome of either conflict onset or incidence; RefugeeSites i,t is the binary treatment indicator for whether the province-year is hosting refugee sites; OtherP rovrefugeesites i,t is the binary treatment indicator for whether there is another province in the same country-year (not counting provinceyear i, t) hosting refugee sites; RefugeeSites i,t OtherP rovrefugeesites i,t is the interaction term; the lagged dependent variable Y i,t 1 ; our control variables W arborderp rovince i,t a binary indicator for whether the province-year borders a province at war, W arbordercountry i,t a binary indicator for whether the province-year borders another country at war, the logged P opulation i,t 16

18 and the logged GDP pc i,t of that province-year, the number of EthnicGroups i,t, T errainsd i which is the standard deviation in terrain ruggedness, P rovsize i the size of the province in square kilometers, BorderDistance i the distance in kilometers between the province s centroid to the international border, CapitalDistance i the distance in kilometers between the province s centroid to the capital, and IDP Site i,t a binary indicator for whether the province-year is hosting an IDP sites; γ c[i] denotes country fixed effects; λ t denotes year fixed effects; and ɛ i the provinceyear error term. 3.5 Placebo Tests using Future Refugee Hosting A major obstacle inherent to research on migrant-host relations is possible selection bias introduced by where migrants are located vis à vis natives. Voluntary migrants can choose to live in neighborhoods that are more accepting of them and/or whose residents share cultural and ethnic ties (Massey and Denton, 1987, 1988). Unlike voluntary migrants, however, refugees often have considerably less agency and time to choose where they settle since they are forcibly displaced. Nevertheless, we use a placebo test strategy to directly address selection bias concerns with the location of refugee sites. How are locations for refugee sites chosen? First, no standard international procedure exists to regulate how refugees are distributed within a host country or where refugee sites should be located. Migration research and qualitative interviews with refugee officials identify the key factors that determine where refugee sites are located, all of which we control for in our analyses. Patterns of refugee movements and settlements are foremost determined by the exogenous shocks of conflict and other exigent crises in neighboring countries. Within the host country, refugee sites are generally situated in areas where a quorum of refugees congregate based on terrain and proximity to the border (Jacobsen, 1996; Jacobson, 1996; Bariagaber, 2006). Once settled, host 17

19 governments are bound by international humanitarian law to the principle of non-refoulement the obligation to not send refugees away (Article 33, The 1951 Convention Relating to the Status of Refugees). Interviews with UNHCR officials confirm that while a few refugee camps, such as those in Turkey, are planned and controlled by the state, refugee sites in general are almost always established ad hoc or unplanned. 7 We consider it highly unlikely that refugees are typically aware of local social and political dynamics or have the resources to carefully determine optimal settlement locations. Instead, refugees tend to settle on an ad hoc basis. Humanitarian organizations like the UNHCR and state security forces then often respond by deploying resources to these sites as they are being established or sometime thereafter. We control for the variables that researchers and experts believe affect refugee settlement and conflict likelihood. Nevertheless, there may still be unobserved factors inherent to the location of the refugee sites that also affect conflict. To exclude the possibility of unobserved confounders, we use placebo tests to show that there is no relationship between conflict and whether a provinceyear will host refugee sites in the future. Because these provinces eventually do host refugee sites in the future, they are perfect candidates for refugee sites, embodying the selection characteristics (e.g. geographic, sociological, legal, etc.) that would determine where a potential site could be located. Similar to how Lee (2001) shows that future electoral outcomes do not affect past elections, our placebo strategy tests whether future refugee sites affect past conflict outcomes. Within our panel dataset, we construct indicators for Placebo Refugee Sites, which is coded as a 1 for province-years for which the province will eventually host refugee sites, 0 for provinceyears of provinces that will never host refugee sites (even though they may belong to countries 7 Conducted with the Field Information and Coordination Support Section (FICSS) of UNHCR on December 30, 2013 by teleconference and January 12-18, 2015 in person at the UNHCR headquarters in Geneva. 18

20 that host refugee sites), and NA for the year before a province begins to host refugee sites (to be conservative in case the exact date of the refugee site creation is an estimate, especially for informal refugee settlements) and then all the subsequent province-years once a province starts to host a refugee site, thus dropping true treated observations. 8 Since these placebo observations do not yet host refugees, there should be no effect of these sites on the outcomes, which is what we find. This suggests that we control for the correct confounders. 4 Results 4.1 Effect of Refugee Sites on Conflict Onset Figure 3 shows the main results for conflict onset. We first show the baseline predicted probabilities with 95% confidence intervals (black estimates on white background) for each of the four cases: 1. province-years with no refugee sites and with are no other refugee sites elsewhere in the same country-year, 2. province-years with refugee sites but with no other sites elsewhere in the same country-year, 3. province-years with no refugee sites but with other sites elsewhere in the same country-year, and 4. province-years with refugee sites with other sites elsewhere in the same country-year. The effect sizes for hosting refugee sites are thus the difference-inmeans between the previous two, hosting versus non-hosting estimates (black estimates on gray background). The placebo estimates are shown alongside (blue estimates). The first plot uses the Full Data of all province-years, the second plot the Subset Data of only province-years in countries that have ever hosted refugees. With respect to H1a on the direct effect of refugees on conflict onset, province-years with refugee sites are no more likely to experience increased likelihood of onset than non-hosting 8 We do the same for Placebo Refugee Camps and Placebo Refugee Settlements for heterogeneous effects analysis. 19

21 Predicted Probability Actual site Placebo (future) site No Site and No Other Site but No Other No Other Province Site No Site but Other Effects of Refugee Site on Civil War Onset Site and Other Other Province Site Predicted Probability No Site and No Other Site but No Other No Other Province Site No Site but Other Site and Other Effects of Refugee Site on Civil War Onset (Subset to only countries who have ever hosted Refugee Sites) Other Province Site Figure 3: This figure shows the predicted probabilities baseline estimates with 95% CIs (black) of conflict onset for provinces hosting a refugee site or not, conditional on whether there are other provinces in the same country-year hosting a refugee site or not (white background) compared to placebo estimates (blue). It also shows the difference-in-means effect size (gray background) of having a refugee site in a province-year versus not. The first plot uses the Full Data of all province-years, the second uses the Subset Data of only province-years in countries who have ever hosted refugees. Both plots show that conditional on no other provinces with refugee sites in the same country-year, a province with a refugee site has a decreased likelihood of conflict onset than a province without a refugee site, suggesting a conditional risk reduction effect of refugees. In the other cases, provinces with refugee sites do not have a greater likelihood of conflict onset. province-years. In fact, counter to conventional wisdom and much of the existing research, both plots show that conditional on no other provinces with refugee sites in the same country-year, a province-year with refugee sites has a decreased likelihood of conflict onset (i.e. experience a new conflict) than a province without refugee sites. These effect sizes are about 1.43 percentage points with the Full Data and 4.82 percentage points with the Subset Data. Given the baseline propensity for conflict onset is 4.69% and 11.41% respectively, hosting refugee sites almost halves the likelihood of conflict onset, a substantively large effect. The null effects shown by the placebo 20

22 test suggests that this conditional risk reduction effect of refugee sites is not due to unobservable confounders. For H2a on the indirect effect of refugees on conflict onset, we see that province-years without refugee sites are no more likely to experience increased likelihood of onset even if there are other provinces in the same country-year hosting refugees. Again, the null effects from the placebo test suggest that we have included the correct confounders. Please find the full regression table of these models 5 in the Appendix. 4.2 Effect of Refugee Sites on Conflict Incidence Figure 4 shows the main results on conflict incidence. Similar to the previous section, we first show the baseline predicted probabilities with 95% confidence intervals (black estimates on white background) for each of the four cases. The effect sizes for hosting refugee sites are again the difference-in-means between the hosting versus non-hosting estimates (black estimates on gray background). The placebo estimates are shown alongside (blue estimates). The first plot uses the Full Data of all province-years, the second plot the Subset Data of only province-years in countries who have ever hosted refugees. With respect to H1b and H2b, both plots show that province-years with refugee sites are no more likely to experience conflict incidence, compared to the placebo estimates (blue). In the second plot which uses the subset data again confirms the conditional risk reduction effect; it shows that conditional on no other provinces with refugee sites in the same country-year, a province-year with refugee sites has a decreased likelihood of conflict incidence (i.e. experiencing conflict) than a province without refugee sites. This effect sizes is about 4.33 percentage points, a substantively significant effect given the baseline propensity for conflict incidence for these countries is 25.97%. Again, the null effects shown by the placebo test again suggests that this 21

23 Predicted Probability Actual site Placebo (future) site No Site and No Other Site but No Other No Other Province Site No Site but Other Effects of Refugee Site on Civil War Incidence Site and Other Other Province Site Predicted Probability No Site and No Other Site but No Other No Other Province Site No Site but Other Site and Other Effects of Refugee Site on Civil War Incidence (Subset to only countries who have ever hosted Refugee Sites) Other Province Site Figure 4: This figure shows the baseline predicted probabilities estimates with 95% CIs (black) of conflict incidence for provinces hosting a refugee site or not conditional on whether there are other provinces in the same country-year hosting a refugee site or not (white background) compared to placebo estimates (blue). It also shows the difference-in-means effect size (gray background) of having a refugee site in a province-year versus not. The first plot uses the Full Data of all province-years, the second uses the Subset Data of only province-years in countries who have ever hosted refugees. The second plot shows that conditional on no other provinces with refugee sites in the same country-year, a province with a refugee site has a decreased likelihood of conflict incidence than a province without a refugee site, again suggesting a conditional risk reduction effect of refugees. In the other cases, provinces with refugee sites do not have a greater likelihood of conflict incidence. conditional risk reduction effect is not due to unobservable confounders. Please find the full regression table 6 in the Appendix. 4.3 Robustness Checks To confirm these findings, we perform the following robustness checks described in greater detail in the Appendix. First, we rerun the analysis, dropping one country at a time (with replacement). This exercise confirms that results are not driven by any outlier countries (section 22

24 7.2.1). Similarly, we repeat the analysis, dropping one region at a time (again, with replacement) and again, show that results persist, although some coefficients lose statistical significance (section 7.2.2). The conditional risk reduction effect is largest when East Africa is dropped from the analysis, which is consistent with the qualitative case study research pointing to Burundi, Kenya, and the Democratic Republic of Congo as countries where refugees have been involved in spreading conflict. Next, we rerun the analysis using a modified Subset Data. Unlike the Subset Data described in the previous sections, which only includes countries who ever have and ever will host refugee sites within the panel period, this modified version is subsetted dynamically to address concerns over possible collider bias, which can occur in analysis at a certain time that conditions on variables in the future (Morgan and Winship, 2014). In our case, at time t, we are including the subset of all countries that have not only hosted refugees by time t but also will host refugees after time t. Thus, we also create a Dynamic Subset Data in which at any year t, we only include the subset of countries who have already hosted refugees in the past. Results in section using this data generally confirm that refugee sites are not positively associated with conflict onset or incidence, yet the conditional risk reduction effect loses statistical significance because we are dropping many observations. Additionally, because terrain is an important consideration for where refugees settle, we rerun our Conflict Onset and Conflict Incidence models including all terrain (or land cover type) variables: evergreen, deciduous, wetlands, croplands, barren, urban, shrub, herbaceous, and water-snow-ice. Specifically, these variables represent the sum total of 300 meter squared grid-squares within a given province consisting of each particular terrain type. 9 Results do not 9 We do not include these variables in the main analysis because they lead to issues of perfect separation in estimating the logistic regressions. 23

25 change under this specification (section 7.2.4). Finally, to examine whether refugee sites might have unforeseen future effects on conflict, we replace the outcome variables, Conflict Onset and Conflict Incidence at time t with conflict onset and incidence at time t + 5. Regardless whether a hosting province is the only one in the same country-year, there are no effects of refugee sites on future conflict (section 7.2.5). 4.4 Discussion of Counterintuitive Findings Our results show a null and, in some cases, negative effect of the presence of refugees on conflict. Thus, we seem to substantively contradict conventional wisdom and the existing seminal quantitative research of Salehyan and Gleditsch (2006). These latter scholars find that increased refugee populations from neighboring countries are associated with increased likelihood of conflict onset. While there are many reasons why our subnational analysis differs from their country-level analysis, one substantial difference in the two studies is the respective time periods of coverage. Our subnational data covers the current post-cold War period from 1989 to 2008 while theirs runs from 1946 to 2001 and consists primarily of years during the Cold War. Conflict dynamics between the two periods differ, particularly in terms of the proxy wars fought across developing countries during the Cold War, which might explain, at least in part, the difference in results that we and Salehyan and Gleditsch (2006) uncover. It is possible, for instance, that refugees had a general positive effect on conflict during the Cold War but not afterward. We, therefore, attempt to reconcile the two findings by replicating and extending Salehyan and Gleditsch (2006) s analysis, specifically separating the Cold War and post-cold War periods. We begin by extending Salehyan and Gleditsch (2006) s original country-year analysis. Specifically, we replicate their study for the post-cold War period from 1990 through 2013 (the most recent 24