Understanding Violent Attacks against Humanitarian Aid Workers

Understanding Violent Attacks against Humanitarian Aid Workers1 Kristian Hoelscher Jason Miklian Håvard Mokleiv Nygård PRIO Abstract What factors explain attacks on humanitarian aid workers? Most research has either tended to describe trends rather than analyse the reasons underlying attacks, or lacks the empirical evidence to support causal assertions. In moving this agenda forward, we present to our knowledge the first cross-national time-series study that identifies factors related to violent attacks on humanitarian aid workers. Drawing on security, civil conflict, and criminal violence literatures, our theoretical framework explores three groups of potential explanatory factors: dynamics of conflict; the political economic context; and aspects of humanitarian sector operations. Using a global sample at the country-level from 1997-2014, we identify factors related to lethal and non-lethal attacks on humanitarian workers. Our results indicate that the presence and severity of armed conflicts are related to increased attacks on aid workers, but neither conflicts that actively target civilians nor levels of criminal violence increase risks to humanitarian workers. We also find more economically developed and politically stable countries are safer for aid workers, and that the presence of an international military force does not add to aid worker risk. Keywords: Humanitarianism; Humanitarian Security; Civil Conflict; Humanitarian Aid; United Nations; Civilian Violence; Development Aid; International Non-Governmental Organizations; Risk and insecurity; Aid Worker Attacks. 1 Authors listed alphabetically. Contributions are equal. Our thanks go to Kristin Bergtora Sandvik and Abby Stoddard for useful discussions and feedback; and to Ida Rudolfsen for excellent research assistance. 1 1 Introduction Humanitarian aid is experiencing a golden age. Global aid spending by governments and private actors has increased 400% since 2000, amounting to $25 billion USD in 2014 (Alnap 2010; GHA 2015). The number of international non-governmental (INGO) aid workers2 has nearly quadrupled in the past 25 years, with 115,000 INGO aid workers in 1997, 210,000 in 2008, and 450,000 in 2014 (Alnap 2010; Fast 2014). These increases are also a function of global instability, as some 58 million people were displaced in 2014 alone – the highest total ever recorded (GHA 2015). Yet this international goodwill has also created hazardous side-effects. Aid delivery areas tend to be in conflict or crisis zones, increasing operational insecurity and at times blurring motivations for aid disbursement. Recognizing that aid and the political and military goals of foreign powers are increasingly conflated (Carmichael and Karamouzian 2014), many have questioned these motivations in light of the increasing numbers of aid workers attacked and killed while in the field. In 2013, 461 aid workers were attacked, representing the most violent year ever in terms of absolute numbers (HO 2013). Figure 1 shows the locations of global aid worker attacks between 1997 and 2014, with bubble sizes corresponding to the number of aid worker attacks in each country during this period. As is clear from the map, aid worker attacks occur disproportionally in a set of countries such as Afghanistan, Pakistan, and Syria, and drive a narrative that aid work is becoming more dangerous as conflict actors increasingly disrespect maxims of humanitarian neutrality (Stoddard et al. 2006; Humanitarian Outcomes 2014; Brooks 2015).3 While a persuasive narrative, there is a lack of empirical evidence explaining why aid workers are attacked to support it. This knowledge gap has important implications for security and risk 2 “Aid Workers” are defined by Humanitarian Outcomes (2015) as ‘the employees and associated personnel of not- for-profit aid agencies (both national and international) that provide material and technical assistance in humanitarian relief contexts. This includes both emergency relief and multi-mandated (relief and development) organizations: NGOs, the International Movement of the Red Cross/Red Crescent, donor agencies and the UN agencies belonging to the Inter-Agency Standing Committee on Humanitarian Affairs (FAO, OCHA, UNDP, UNFPA, UNHCR, UNICEF, UN-Habitat, WFP and WHO) plus IOM and UNRWA. The aid worker definition includes various locally contracted staff (e.g., drivers, security guards, etc.), and does not include UN peacekeeping personnel, human rights workers, election monitors or purely political, religious, or advocacy organizations’. 3 To put the figures into perspective, international aid workers are killed at a rate of approximately 50 deaths per 100,000 workers, below that of loggers (108/100k) or pilots (64/100k) in the United States (BLS 2014), and comparable to the homicide rate in violent urban centers such as Caracas, Venezuela (100/100k) or Kingston, Jamaica (50/100k). (Guardian 2014) 2 protocols in humanitarian organizations; for donors and researchers engaging with humanitarian insecurity. We aim to fill this gap by combining existing datasets on aid worker deaths, violent conflict, security and development to more accurately assess which national-level conditions place aid workers at risk. The study is motivated by the research question: what factors explain the spatial and temporal distribution of attacks on humanitarian workers? Figure 1: Number of aid worker attacks, 1997 – 20144 The following section reviews the state-of-the-art on the trends and suggested causes of aid worker attacks to identify potential explanatory factors. Section three outlines our theoretical framework and hypotheses, and section four outlines our data and empirical strategy. Section five presents our results and discussion, with section six discussing limitations and caveats. Section seven considers the implications of our research and concludes. 2 Background and Literature Measuring and clarifying humanitarian insecurity An emerging literature has examined aid worker attacks over the past decade. Stoddard et al (2006; 2009), Fast (2010; 2014) and Wille & Fast (2013) have all made important contributions in highlighting the threats that humanitarian workers face. A considerable resource has been the Aid Worker Security Database (AWSD) a ‘global compilation of reports on major security 4 Size of bubbles is proportional to number of attacks. Countries in grey have seen at least one year of internal armed conflict (as defined by Gleditsch et al. 2002) over the same period. 3 incidents involving deliberate acts of violence affecting aid workers’ (Humanitarian Outcomes, 2015: online). The AWSD has provided annual data since 1997 on lethal and non-lethal attacks on aid workers. This data has been instrumental in highlighting modes of aid worker deployment and patterns of aid worker attacks, including modality of attack, and institutional affiliation of the target. An important series of studies has examined the rising incidence of aid worker attacks in the past decade. (Stoddard et al. 2006; 2009; Humanitarian Outcomes, 2013; 2014) Stoddard et al. (2006) suggest that security protections have deteriorated, particularly for local UN staff. Wille & Fast (2013) also compare targeting of international and national staff, suggesting that fatalities of national staff have increased proportionately over time, and that this is more pronounced in national divisions of UN and Red Cross agencies. The key message is that it has become more dangerous to be an aid worker in the field (Fast 2014), particularly in conflict affected regions. Figure 2: Yearly aid worker attacks and total number of battle deaths, 1997 – 2014 Figure 2, shows the total number of global aid worker attacks from 1997 to 2014, and the number of battle related deaths globally in the same period (UCDP 2015). The dashed vertical line marks the start of the current conflict in Syria that has coincided with a substantial increase in both the number of aid worker attacks and battle deaths. Broadly, the figure seems to support 4 the notion that attacks on aid workers are increasing over time. Yet despite the recent increase in aid worker attacks and battle related deaths, these measures do not appear to coincide over the past two decades, calling into question the implicit or assumed relationship between armed conflict and attacks on aid workers. Yet while Figure 2 suggests a general upward trend in attacks on aid workers in recent decades, this does not necessarily imply that risk to individual aid workers has increased. If we break down Figure 2 by country, we find that the mean number of annual attacks for countries that experience at least one attack between 1997 and 2014 has not increased. Rather, the mean number of attacks in any given country has been relatively constant in the past two decades. Two factors are likely responsible for the increase in attacks reported in Figure 2, particularly in recent years. First is the increasing number aid workers in the field, with the 1997-2014 time period covered here having seen a steady increase in the number of aid workers deployed (Humanitarian Outcomes, 2015). Considering this, per capita rates of attacks based on the number of humanitarian workers in the field might suggest that the overall risk has stayed more or less constant since the late 1990s – and potentially even back to the mid-1980s (Shiek et al. 2000). Second, the increased number of total attacks appears driven by a small number of countries registering an above average number of incidents. Afghanistan and Syria, for example, in the past few years have seen many more aid worker attacks than previously seen in other countries.5 Yet even if we acknowledge that the field is becoming more dangerous, we still do not fully understand why patterns are changing, and different approaches and viewpoints of researchers, in-the-field practitioners, and policy actors emphasise different causal pathways. Fast (2010) offers a clean baseline of the state of the field, arguing that two streams constitute the majority of scholarship: an epidemiological approach that emphasizes “proximate causes” of violence against aid workers using empirical data; and “deep causes” literatures that attempts to explain violence through the lens of increasing politicization of humanitarian aid. These may include (among others): shrinking humanitarian space endangering relief workers; embedded operations with military actors increasing the exposure of aid workers to frontline fighting or targeted 5 The area to the right of the vertical line in figure 2 and the shaded areas in figure 3 suggest that this increase in the total number of attacks is being driven by a small number of countries. 5 attacks; or the threat of terrorism or anti-Western sentiment incentivizing attacks on aid workers. The evidence base that would support or reject these explanations, however, is relatively poor – both at a country-specific and a global level. Moreover, comprehensive studies of those who perpetrate attacks against aid workers are rare, leaving most assumptions about the motivations behind attacks to lean towards conjecture.6 It is our contention that a key drawback of the current state of the art on humanitarian security is that it often focuses on description of trends rather than explanation of underlying causal or proximate explanatory factors; and relies primarily on anecdotal or ad hoc evidence to interpret or explain trends in available data. While this analysis has inarguably been important, such a focus may inadvertently obscure the explanatory factors that motivate humanitarian attacks; feed misleading narratives about why aid workers are targeted; sideline issues related to dealing with ‘legal protection gaps and disparities in staff vulnerability’7; or bias certain policy actions or security considerations in response. Given the changing role of aid in the dynamics of war (Wood & Sullivan 2015) – and the changing role of conflict and humanitarian engagement in the 21st century (Dononi et al, 2008) – a closer examination is needed to better understand (and ultimately work to reduce) violence against humanitarian workers. To this end, we build on Fast’s classification to develop a simple theoretical framework to understand aid worker attacks, presented in section three. Before doing so, we explore two major conceptual trends in INGO field operations: reducing risk by going local; and forsaking neutrality for integration. INGO risk reduction strategies Aid organizations are working much more extensively in remote field settings, doing more sophisticated work, and undertaking a wider variety of tasks than just a decade ago8. At the same time, gaps between needs and practices illustrate the increasing securitization of field workers. Many aid organisations also consider risks themselves to be more unpredictable, adding to perceptions of insecurity at both individual and organizational (reputational) levels (Egeland 6 Notable exceptions include MSF (2014), which explores how those in crisis view aid actors, and Sparrow (2013), who provides a vivid account of the justification of attacks on medical aid workers by Syrian fighting forces. 7 On this point see Brooks (2015). 8 See for example, Miklian 2014; and Sandvik & Hoelscher, forthcoming. 6 2011). Traditionally, pan-institutional risk profiling was done by developing case-specific guidance for knowing ‘when to leave’ before the situation became too volatile. Yet in the interest of engaging with those most in need, organizations are increasingly looking to instead figure out ‘how to stay’ in even the most dangerous situations (Egeland 2011). This new philosophy has required INGOs to employ new strategies to attempt to reduce operational risk, first and foremost by compartmentalizing security strategies for domestic and international expatriate staff.9 This employs a multi-pronged risk reduction approach: leaning more heavily on trusted national partners; presenting a less-visible local profile;10 increasing staff security mechanisms (Kurtenbach and Wulf 2012); and introducing "remote management” to subcontract more dangerous or inaccessible tasks to local NGO staff instead of their international equivalents (Miklian et al. 2011; Brooks 2015). The reasons for this are many: national staff tend to greatly outnumber their international counterparts (by a factor of 10 to 1) and constitute a lower risk profile in both political, economic and reputational terms for the INGO; yet they are also less likely to receive security training due to the expense (MA 2015). Internationals can also employ more personal strategies to reduce risk like leaving the region, that local staff cannot (Rubenstein 2014). Further, there is also growing similarities between corporate, military and INGO risk and security strategies in fragile and conflict areas (Avant and Haufler 2012). ‘Embedded aid’ has re-ignited the debate over the effectiveness and local perception of using armed escorts for aid workers and their cargos, and is a matter of fierce debate in heavily-militarized conflict zones such as Afghanistan (Olson 2006).11 However, the standardization of risk strategies within the INGO community tends to be reactive, constituting good-faith efforts to keep employees safe in challenging work environments. Security strategies are also influenced by independent assessments of risk in given areas – such as the US State Department Travel Alert system, UK Travel Advisory system, or UN Duty Station classification system – and once assessments are made for a given area, most INGOs 9 Rowley et al. (2013) provide a systematic review of security practices of major aid INGOs operating today. 10 This is contrasted with the heavily criticized efforts of INGOs in the 1990s and 2000s to increase brand exposure (especially through international news channels) by planting logos and insignias across relief materials. 11 Perhaps paradoxically, this has led to an increase in both supply of and demand for new NGOs operating in conflict zones, which then exposes even more national staff to insecure environments (MA 2015). 7 implement standard protocols for employee protection. But given that aid workers face special political and professional concerns that differ from tourists or government officials (and differ between local and international staff), it is possible that these assessments over- or under-play personal risks for INGO staff. INGO integration strategies INGOs have a long history of seeing themselves operating as removed from – but working in parallel to – the conflicts they operate in. Neutrality and impartiality have served as an INGO gold standard for several reasons: to gain access to provide humanitarian assistance in contested spaces; to protect workers by framing them as ‘illegitimate targets’; and to be seen as above the fray of violence and worthy of protection by all fighting sides. But the end of the Cold War ushered in a dramatic expansion of the aid sector, with many INGOs moving from purely humanitarian work to also taking on development portfolios, which tend to be more political in nature. In response, there are concerns about links between aid actors and donor governments in the Global South (Abiew 2012); the business-like actions of humanitarian aid agencies (Weiss 2013); and how these links widen the gap between INGOs and their national implementing agents (Miklian et al. 2012). The loudest critiques have come in the field of security studies. Chandler (2006), Mills (2005), Duffield et al. (2001) and others argue that the 'new humanitarianism' linking INGOs to western governments to operationalize peace and security has eroded local perceptions of the value and neutrality of 'humanitarian' space. One contributing factor is mission creep, as today ‘humanitarian’ work can include everything from state-building to democracy promotion (Barnett 2005). Some believe this places INGO staff at additional risk. Abiew (2012:208) argues that “(t)he integration of politics and humanitarian action has been a major reason behind the attack on humanitarian aid workers and their inability to deliver aid to the neediest”” Within this critical strand, increasing politicization and deeper ties with governments and/or militaries is considered a key factor motivating violence against INGO workers. Some have tied humanitarian insecurity to the militarization of aid (Lischer, 2007), and military-embedded humanitarian operations (Barry & Jeffries 2002). Others counter that this assumes that aid 8 INGOs were truly neutral and apolitical before, a dubious proposition given governments have securitized aid programs for over 120 years (Barakat et al. 2010). The harshest critiques view contemporary INGO work as little more than attempts to socially engineer societies based on ‘western’ ideals (Duffield 2012; Richmond 2007), in the process threatening the continued validity of the humanitarian enterprise (Donini et al. 2008). While these critiques may hold for certain contemporary conflict situations (ISIS comes to mind), studies of integrated UN missions that merge military, humanitarian and political action are less conclusive (e.g. Combaz 2013; Ferreiro 2012; Donini 2011). A common thread in both INGO risk reduction and integration strategies is that conclusions are contextual, few ascribe global lessons to data trends, and even fewer account for the socio- political context of violence or type of conflict fought. Simply, there is much we still do not know. In the absence of a counterfactual, it is uncertain if violence against aid workers is a function of increasing risk or changes in INGO operations. It is possible that aid workers are indeed increasingly targeted – even in per capita terms – but the improvement of security by INGOs for their field staff may have reduced the number of attacks to below what they might otherwise have been without such improvements. It is also possible that aid worker attacks are less of a ‘special’ phenomenon than assumed, and merely reflect levels of overall violence in country.12 As such, aid workers operating on more dangerous regions may simply be attacked with greater frequency due to the underlying regional insecurity – either in the form of criminal or political violence. Such unanswered questions illustrate the need to further scrutinize available data sources and collect new information to better understand where improved security for the most vulnerable workers is needed, and how to ensure its delivery. 3 Theoretical Framework and Hypotheses Our theoretical framework identifies three broad themes that may explain of attacks on humanitarian workers. These are: 1) the nature of conflicts that humanitarian agencies operate within; 2) the political factors economic conditions that may motivate attacks; and 3) the structure of humanitarian operations. 12 ACF International (2015) touches on this difference between aid worker and civilian attacks in a recent discussion paper, asking both whether it is useful to differentiate between these humanitarian and civilian victims; and whether more focus should be placed on global efforts to enhance protection of civilians who experience violence to a far greater extent than humanitarian workers. 9 Conflict Dynamics One explanation assumes that attacks are a function of the location and intensity of fighting. Humanitarian agencies engage in myriad conflict and post-conflict environments, and travel and operational security for aid workers is more difficult in these situations (Humanitarian Outcomes, 2013; 2014). Qualitative evidence suggests conflict dynamics play a significant role in how and where aid workers are targeted (Fast 2014). Quantitatively, Stoddard et al. (2006) find that for a small sample of countries with a high presence of aid workers, attacks were greater in countries with interstate wars and under conditions of civil violence; but less likely where civil wars occurred. We anticipate the presence and severity of conflict in a country influences how humanitarian workers are targeted. Hypothesis 1a: Attacks on aid workers will be greater where conflicts are present in a country. Hypothesis 1b: Attacks on aid workers will be greater where conflicts are more violent. Further, the type of conflict may have an effect on how aid workers are targeted. In particular, in conflicts where insurgents are seeking secession, rebel groups might be more suspicious of the presence of aid workers in their territory. We suggest: Hypothesis 2: Attacks on aid workers will be greater where conflict actors control territory or aim to do so. The strategic use of violence during a conflict may also influence how aid workers are targeted. We consider the extent to which combatants target civilian populations as potentially increasing aid workers risk. If civilians make up a large number of casualties in a conflict, combatants use violence less discriminately against all non-combatants (Eck and Hultman 2007), and may have fewer reservations against targeting neutral parties: Hypothesis 3: Attacks on aid workers will be greater where conflict actors are actively targeting civilians. Political and Economic Context The political institutional context or governance capacity of state may be factors which explain targeted violence against humanitarian workers. Generally, strong democratic and autocratic 10 states are more able than weak or transitional states to ensure a modicum of territorial security and legitimacy over the use of violence within their borders. Transitional states often have higher rates of civil conflict (Hegre et al. 2001) and social violence (Fox & Hoelscher 2012), and may present greater risks to humanitarian actors as INGO governance and capacity-building programming may be perceived as challenging state authority. We hypothesize that: Hypothesis 4: Attacks on aid workers will be greater in weak, inconsistent, or transitional states. We also suggest economic incentives play a role in how humanitarian workers are targeted. As representatives of wealthy, visible foreign entities, INGO workers may be valuable kidnapping targets, or individuals owning lootable material goods. Therefore the economic drivers of criminal and political violence,– namely low economic development and high inequality (e.g. Østby, 2013; Fajnzylber, Lederman & Loayza, 2002; Moser, 2004) – may increase the likelihood of opportunistic attacks against high-value targets such as humanitarian workers. We hypothesize: Hypothesis 5: Attacks on aid workers will be greater in states with lower levels of economic development. Hypothesis 6: Attacks on aid workers will be greater in states with higher levels of inequality. Relatedly, then, humanitarian attacks may also be a function of general levels of insecurity in a society. Indeed, Buchanan and Muggah suggest that criminal motivations are a major factor in explaining attacks on humanitarian and development workers, and that ‘by far their biggest risk is tied to criminal violence’ (2005:24). Hypothesis 7: Attacks on aid workers will be greater in states with higher levels of criminal violence. Humanitarian Operations A third set of explanations study the nature by which humanitarian INGOs operate. Here we consider if the ways humanitarian agencies deploy field staff affect how staff are targeted. For example, if agencies operate in more complex or dangerous environments, or place humanitarian staff in precarious frontline situations, it follows that aid workers may be increasingly targeted. 11 First, increasing attacks may simply be due to organisations deploying staff in more risky, dangerous environments in general. To assess this, we examine how attacks are related to country risk profiles used by international organisations. Hypothesis 8: Attacks on aid workers will be greater where operations occur in countries with a higher risk profile. Another important consideration is the closer operational ties between militaries and INGOs. Some argue that conducting humanitarian operations alongside the military places aid workers in greater danger, either by blurring lines between military actor and humanitarian agent (Duffield et al. 2001); or due to humanitarian workers being perceived of being biased toward a particular party in a conflict (Abiew 2012). Moreover, certain conflict actors may be opposed to the real or perceived political, developmental and humanitarian agendas of aid organizations. In such cases the presence of actors such as the US military or NATO may encourage extremist groups to attack humanitarian actors perceived to be agents of great powers (e.g. Fast, 2010). Hypothesis 9: Attacks on aid workers will be greater where NATO ground operations are present. Similarly, humanitarian security in post-conflict environments may be affected by integration between aid INGOs and United Nations peacekeeping operations (PKOs). UN peacekeeping missions are assumed to create space for secure humanitarian entry into post-conflict areas, though some argue this also leaves humanitarian workers vulnerable to attack given their international and/or ‘Western’ associations (InterAction 2011). Our general assumption is that UNPKOS make aid workers safer, and that both the presence and size of UNPKOs will have an effect. Hypothesis 10a: Attacks on aid workers will be lower where UN PKOs are present. Hypothesis 10b: Attacks on aid workers will be lower the larger the UN peacekeeping force. Further, the type of mandate employed may also be important, as PKOs are most effective at curtailing violence where transformational mandates – those designed to address the roots of the conflict, such as economic reconstruction and institutional transformation (i.e. reform of police, army, judicial system, elections) – are employed (Hegre et al. 2011). We suggest: 12 Hypothesis 10c: Attacks on aid workers will be greater where UN PKOs have traditional mandates rather than transformational mandates. 4 Data and Empirical Strategy To examine our hypotheses we use time-series data across a global sample of countries between 1997 and 2014, with the country-year as unit of analysis. Our baseline model is a negative binomial regression. This estimates the count of events, here aid worker attacks, when events are over-dispersed compared to what the Poisson distribution would predict.13 We fit a negative binomial regression where this over-dispersion is modelled by assuming that each individual observation follows the Poisson distribution, but in addition a variable 𝑣𝑖 is added to the individual effects where 𝑒 𝑣𝑖 is gamma distributed with mean 0 and variance 𝛼 (Hilbe 2011). Let 𝔦 index country-years, then our model is given by: 𝑦𝑖 ~ 𝑃𝑜𝑖𝑠𝑠𝑜𝑛(𝜇𝑖 ) 𝜇𝑖 = exp(Χ𝑖 𝛽 + 𝑂𝑓𝑓𝑠𝑒𝑡𝑖 + 𝑣𝑖 ) 1 𝑒 𝑣𝑖 ~ 𝐺𝑎𝑚𝑚𝑎 ( , 𝛼) 𝛼 As the data contains a large number of zero values, we also fit a zero-inflated negative binomial, as well as a Poission regression and a simple OLS with country fixed and random effects to test robustness. Reported in the appendix, these specifications yield substantively the same results. Dependent Variables Our main dependent variables measure lethal and non-lethal attacks on humanitarian workers between 1997 and 2014 using incident data drawn from the AWSD (Humanitarian Outcomes, 2015). The AWSD uses systematic media filtering and information directly provided by aid agencies to compile counts and descriptions of global attacks on humanitarian aid workers, regularly cross-checking figures with regional and field-level consortiums.14 Our analyses primarily use combined lethal and non-lethal attacks on aid workers as the main dependent variable, yet also disaggregate analyses for aid workers killed, wounded, and kidnapped.. 13 Formally this means the variance of the counts does not equal the mean of the counts. 14 For full description of the data and methodology see Humanitarian Outcomes (2015) 13 Independent Variables Our key independent variables cover three theoretical constructs: conflict dynamics; political and economic context; and humanitarian operations. Conflict Dynamics To measure conflict presence (Hypothesis 1a) we use a dummy based on whether the UCDP/PRIO Armed Conflict Database (ACD) (Gleditsch et al. 2002, Pettersson and Wallensteen 2015) registers an internal armed conflict in the country. To measure conflict intensity (Hypothesis 1b) we use dummies for high intensity conflict and low intensity conflict, and a measure of the total count of battle deaths in a given year, all from UCDP. For conflict type (Hypothesis 2) we use the UCDP/PRIO ACD measure of whether the conflict is territorial, meaning challengers want succession or regional autonomy; or governmental where challengers want to change the political system, or change the composition or replace the government. To test whether aid workers are attacked more frequently where civilians are targeted (Hypothesis 3), we include a dummy variable indicating whether the country experiences one-sided violence (Eck and Hultman 2007). Political and Economic Context For Hypothesis 4, we use the absolute and squared polity score based on the Marshall & Jaggers (2009) widely used Polity IV index15 to measure political institutional consolidation. We measure the effect of economic development (Hypothesis 5) using GDP per capita and the effect of income inequality (Hypothesis 6) using the GINI index (Solt 2014). To test the effect of levels of criminal violence on attacks on humanitarian workers (Hypothesis 7) we use the national homicide rate, regarded as the most robust and accurately reported measure proxying generalised insecurity. We draw this data from the World Development Indicators 2015 update.16 Humanitarian Operations We test the effect of humanitarian operations in risky environments (Hypothesis 8) using data taken from the ICRG classification of political risk (PRS Group 2015). This is a scale ranging 15 This measures a country’s placement on a 21-point scale between full autocracy (-10) and full democracy (10). 16 See http://data.worldbank.org/data-catalog/world-development-indicators 14 from 0 to 12 determining risk of government instability, with 12 indicating very low risk. The ICRG uses this variable as a proxy for the overall political risk of a country, and we argue it is a reasonable proxy for general levels of country risk. For Hypothesis 9 assessing foreign military intervention, we use a dummy indicating the presence of a NATO ground mission in a given year. In Hypothesis 10 we also measure the effect of presence (10a), size (10b) and mandate (10c) of UN PKOs. Here we use a dummy measuring the presence of a PKO and a measure of the total PKO budget, both derived from (Hegre et al. 2011); and dummy variables indicating whether the PKO mandate is traditional or transformational (Doyle and Sambanis 2006). Control Variables We control for several additional factors. We include the log total national population since countries with larger populations generally see more conflict (Hegre and Sambanis 2006) and/or might attract more aid workers. To further control for other aspects of state capacity not captured by our political or economic independent variables, we include a variable measuring the number of years since last regime change; and the number of years the country has been in peace. Finally, to account for inertia effects of aid worker attacks and potential autocorrelation in the data, we include we include a lagged dependent variable to control for and autocorrelation in the data in all estimations reported below. 5 Results and Discussion Table 1 reports the first set of results, focusing on conflict dynamics. Here we estimate the effect of the explanatory variables on the total count of both lethal and non-lethal aid worker attacks. Results strongly support hypotheses 1b and 1b. The estimates in Column 1 for the effect of internal armed conflict are large and clearly different from zero. This holds for both for minor conflicts (those incurring between 25 and 999 battle-related deaths per year) and major conflicts (those incurring over 1000 battle deaths per year). Both in this estimation and all subsequent estimations we find, unsurprisingly, that aid workers are much more likely to be attacked in countries experiencing conflict than in peaceful countries. Somewhat surprisingly though, the difference between minor and major conflicts is not that large regarding the effects on the number of expected attacks. An average country with a minor conflict is likely to see 7 aid 15 worker attacks annually, while a similar country with a major armed conflict is likely to see 14 attacks.17 Table 1: Negative binomial regression, Conflict Dynamics, 1997 -- 2014. (1) (2) (3) (4) Conflict BRD Territory One sided Minor conflict 1.282*** 1.627*** (0.346) (0.424) Major conflict 1.898*** 2.066*** (0.359) (0.377) Battle deaths 0.000394* (0.000) Territorial 0.229 (0.448) One sided violence 0.00148 (0.001) ln(population) 0.208 0.288** 0.136 0.0814 (0.114) (0.111) (0.126) (0.215) ln(GDP capita) -0.646*** -0.488** -0.699*** -0.220 (0.136) (0.166) (0.180) (0.243) ln(Time in peace) -0.328** -0.515*** -0.364** -0.339** (0.115) (0.136) (0.124) (0.112) Time since regime change -0.00188 0.0109 0.00919 0.0216 (0.007) (0.006) (0.005) (0.015) Polity 2 0.011 0.0267 0.117*** -0.0496 (0.031) (0.034) (0.031) (0.047) Polity^2 -0.013* -0.0180** -0.0190* -0.00525 (0.006) (0.007) (0.008) (0.008) Aid worker attacks (t-1) 0.0934* 0.361*** 0.148** 0.238*** (0.037) (0.080) (0.049) (0.035) _cons 2.253 1.266 3.568* 1.656 (1.158) (1.192) (1.809) (2.152) lnalpha _cons 1.830*** 1.951*** 1.692*** 1.471*** (0.187) (0.177) (0.184) (0.178) AIC 3313.4 2834.1 1462.7 1503.4 ll -1647.7 -1407.1 -719.4 -741.7 N 2692 2480 1779 383 Standard errors in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001 The relationship also holds when measuring conflict intensity using battle-related deaths (Column 2). Figure 3 illustrates the effect of battle deaths on the expected number of aid worker attacks. It simulates the expected count of aid worker attacks for an average country in conflict 17 Proportionally this increase is, of course, substantial. 16 as log battle deaths increases (King et al. 2000),18 showing a strong and significant effect of conflict intensity. An increase in battle deaths from around 400 deaths per year, a medium intensity conflict, to 2000 deaths a year, a high intensity conflict, roughly doubles the expected amount of aid worker attacks. For aid organizations the lesson here is clear, the more violent the situation they deploy to, the more care they need to take. Figure 3: Expected aid worker attacks as conflict intensity increases (from Table 1). Column 3 analyses Hypothesis 2 to ascertain whether aid worker attacks are conditioned by aims of the rebels. Whether a conflict is fought over government or territory is only defined for countries in conflict, and consequently we only include these states in this analysis. We find no evidence indicating that rebel motivation significantly influences aid worker attacks. Conflicts where rebels are seeking secession or regional autonomy (and where rebels often have a strong regional presence) do not show more attacks on aid workers than in conflicts motivated to change the government or the policies of the state. We also find no evidence indicating that countries experiencing one-sided violence, i.e. where either the government or insurgents are actively targeting civilians, have higher rates of attacks on aid workers (Hypothesis 3). Column 3 reports that the effect of one-sided violence on aid worker attacks is essentially zero. This is potentially encouraging news. Periods of one-sided 18 An average country in conflict has a (log) population of 10, 11 years since last regime change, and saw 15 attacks at t-1. We use this country profile for all simulations reported below. 17 violence are situations where civilian populations are especially vulnerable and in need of humanitarian assistance. That these situations do not appear to be more dangerous to aid workers may further encourage the international community to provide more extensive support for vulnerable populations. Table 2. Negative binomial regression, Conflict Dynamics, Aid worker attacks disaggregated, 1997 -- 2014. (1) (2) (3) Killed Wounded Kidnapped Minor conflict 1.371*** 1.214** 0.803 (0.342) (0.400) (0.435) Major conflict 2.546*** 1.712*** 2.485*** (0.433) (0.415) (0.658) Polity 2 0.0470 0.0207 -0.0435 (0.031) (0.034) (0.039) Polity^2 -0.00812 -0.0116 -0.0200* (0.008) (0.008) (0.008) ln(population) 0.123 0.275* 0.148 (0.134) (0.108) (0.123) ln(GDP capita) -0.570*** -0.479** -0.379 (0.158) (0.162) (0.249) ln(Time in peace) -0.370* -0.176 -0.314* (0.173) (0.140) (0.154) Time since regime change 0.00799 -0.0000134 0.00716 (0.007) (0.007) (0.010) Aid worker killed (t-1) 0.280*** (0.060) Aid worker wounded (t-1) 0.320** (0.103) Aid worker kidnapped (t-1) 0.385* (0.152) _cons 1.223 -0.581 -0.0440 (1.159) (1.273) (1.771) lnalpha _cons 1.474*** 1.909*** 2.697*** (0.208) (0.221) (0.394) AIC 1640.9 1784.6 1207.9 ll -809.4 -881.3 -592.9 N 2480 2480 2480 Standard errors in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001 Table 1 grouped all attacks on aid workers, lethal as well as non-lethal, together, yet certain conflict dynamics may encourage different types of attacks on aid workers. Delving deeper, Table 2 distinguishes between counts of aid workers killed (Column 1), wounded (Column 2), and kidnapped (Column 3). Regarding the effect of conflict intensity, we find little or no difference between whether aid workers are killed or wounded, but that kidnappings occur at a 18 much lower rate in minor armed conflicts than major armed conflicts. As discussed more below, we also find that economic development affects lethal and non-lethal attacks differently. While less developed countries see more killed and wounded aid workers, this is not the case for kidnappings where the effect of log GDP per capita is indistinguishable from zero. Next, we examine the political and economic context. Results are shown in Table 3. Overall, we find strong support for Hypothesis 5 (level of development) in two of our economic context variables; but no evidence for Hypothesis 4 (inconsistent states), 6 (level of inequality) and 7 (homicide rate). Table 3: Negative binomial regression, Political and Economic context, 1997 -- 2014. (1) (2) (3) Inconsistent Gini Homicides Minor conflict 1.282*** 1.627*** (0.334) (0.424) Major conflict 1.899*** 2.066*** (0.406) (0.377) Gini -0.0171 (0.018) Homicides 0.00664 (0.010) Polity 2 0.0109 0.117*** 0.129** (0.031) (0.031) (0.047) Polity^2 -0.0129* -0.0190* -0.0246* (0.007) (0.008) (0.011) ln(population) 0.187 0.136 0.199 (0.104) (0.126) (0.178) ln(GDP capita) -0.463** -0.699*** -0.954*** (0.167) (0.180) (0.246) ln(Time in peace) -0.258* -0.364** -0.809*** (0.122) (0.124) (0.180) Time since regime change 0.00520 0.00919 0.0157* (0.005) (0.005) (0.006) Aid worker attacks (t-1) 0.156*** 0.148** 0.452* (0.042) (0.049) (0.212) _cons 0.995 3.568* 6.355** (1.162) (1.809) (2.396) lnalpha _cons 1.772*** 1.692*** 2.243*** (0.183) (0.184) (0.198) AIC 2777.2 1462.7 1045.9 ll -1377.6 -719.4 -512.9 N 2480 1779 1290 Standard errors in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001 19 We report the effect of regime type in Column 1 using both standard and squared Polity measure. We find no support for Hypothesis 4. The assumption that autocracies and democracies would have lower rates of attacks on aid workers while inconsistent states, also called semi- democracies or hybrid regimes, would see a greater number does not hold. As this effect is non- linear and therefore hard to interpret directly from the coefficients, we illustrate the effect through a simulation in the left panel of Figure 4. Here we find a tendency towards more aid worker attacks in the most authoritarian countries, those below -5 on the Polity scale, yet surprisingly find evidence supporting our hypothesis that inconsistent states see more aid worker attacks. This is especially interesting considering abundant evidence suggesting these countries are more conflict prone and instable (Knutsen and Nygård 2015) and experience greater rates of social violence (Fox & Hoelscher 2012). Figure 4: Expected number of aid worker attacks across Polity regime scale (left) and log GDP per capita (right) (from Table 3) 250 200 200 Expected aid worker attacks Expected aid worker attacks 150 150 100 100 50 50 0 0 -10 -5 0 5 10 5 6 7 8 9 10 11 12 Polity regime scale log(GDPcap) 10/90 pctile Attacks 10/90 pctile Attacks Generally, more developed countries measured in terms of GDP per capita see fewer aid worker attacks. Supporting Hypothesis 5, this pattern holds for other measures of state capacity and state consolidation, namely ‘time since regime change’ and ‘time in peace’. We find higher capacity states see fewer attacks, presumably as they are better able to protect aid workers present in the country. The right panel of Figure 4 illustrates the effect in a simulation with the same parameters as above. It shows that the higher the levels of economic development, the fewer aid worker attacks expected. 20 Table 4: Negative binomial regression, Humanitarian operations, 1997 – 2014 (1) (2) (3) (4) (5) Risk NATO Presence Budget Mandate Conflict 1.052*** 1.063*** 1.053*** 1.081*** 1.025*** (0.224) (0.228) (0.207) (0.213) (0.201) Political risk -0.261** (0.081) NATO -0.273 (0.520) PKO 0.990** (0.365) ln(PKO budget) 0.144* (0.058) PKO transformational 0.389 (0.435) PKO traditional 1.406** (0.542) Polity 2 -0.0251 0.0149 0.0288 0.0247 0.0366 (0.043) (0.032) (0.031) (0.031) (0.033) Polity^2 -0.00710 -0.0136* -0.0175** -0.0149* -0.0200*** (0.008) (0.006) (0.006) (0.006) (0.005) ln(population) 0.138 0.197 0.230* 0.213* 0.240* (0.122) (0.103) (0.104) (0.104) (0.106) ln(GDP capita) -0.660*** -0.454** -0.439* -0.423* -0.476* (0.188) (0.165) (0.177) (0.174) (0.188) ln(Time in peace) -0.372** -0.278* -0.197 -0.244* -0.188 (0.119) (0.128) (0.113) (0.124) (0.112) Time since regime change 0.00908 0.00584 0.00508 0.00607 0.00523 (0.005) (0.006) (0.005) (0.005) (0.005) Aid worker attacks (t-1) 0.0984** 0.150*** 0.142*** 0.141** 0.147*** (0.036) (0.042) (0.043) (0.043) (0.042) _cons 5.460** 0.916 0.298 0.385 0.548 (1.755) (1.133) (1.124) (1.126) (1.169) lnalpha _cons 1.662*** 1.774*** 1.740*** 1.749*** 1.736*** (0.223) (0.183) (0.180) (0.183) (0.177) aic 2086.8 2779.3 2758.2 2768.4 2754.8 ll -1032.4 -1378.7 -1368.1 -1373.2 -1365.4 N 1909 2480 2480 2480 2480 Standard errors in parentheses * p < 0.05, ** p < 0.01, *** p < 0.001 Contrary to expectations, Column 2 shows no effect of economic inequality as measured by the GINI index on aid worker attacks (Hypothesis 6). If anything the effect seems to be negative, suggesting states that are more unequal have fewer attacks, but this effect is not different from 0. Simply economic inequality seems to have no effect of aid worker attacks. Moreover, the effect of general level of insecurity as measured by the national homicide rate (Column 3, Hypothesis 7) appears unrelated to the number of aid worker attacks. Countries with 21 high homicide rates see just as many attacks on humanitarian workers as countries with low homicide rates. This could potentially be good news for aid organizations, perhaps indicating the effectiveness of security protocols when deploying to insecure countries.19 Table 4 looks at how the risk profile of the country (Hypothesis 8), the presence of international military forces (Hypothesis 9) and the dynamics of UN PKOs (Hypothesis 10) affect violence against aid workers. Column 1 reports the result for the country political risk. Not surprisingly, and in line with Hypothesis 8, we find that countries that are ranked as having a lower risk score (i.e. they get a high score on the ICRG political risk variable) see much fewer attacks on aid workers.20 This indicates that countries that have functioning state institutions, and therefore have a lower likelihood of seeing (irregular) government instability, are also associated with lower levels of risk of aid workers. Hypothesis 9 posits that we will see more attacks on aid workers in countries where NATO forces are present. To test this we code a dummy variable indicating all country-years in which NATO forces where deployed. Results are reported in Column 2. We find no evidence indicating that countries where NATO forces are present have more attacks on aid workers. The estimated effect cannot be reliably distinguished from zero, but if anything, it is negative – countries with NATO forces see fewer attacks on aid workers. To assess Hypothesis 10 we estimate the effects of peacekeeping presence, budgets and mandates. Results for the presence of UNPKOs (10a) are shown in Column 3. We find that countries with Peacekeepers present see more attacks on aid workers. This is not surprising. The coarse dummy measure essentially only distinguishes between countries that currently are in or have recently been in conflict and those in peace, and given this these results align with our findings in Hypothesis 1. We examine this further in Column 4, which reports results for the budget of the PKO force (10b). For this, we find a positive result that is clearly distinguishable from 0. In other words, the larger the budget of the peacekeeping force – which we assume is highly correlated with the size of the force – the more attacks against aid workers we expect to see. 19 Note, however, that data for homicides is quite poor; consequently we have far fewer observations available to analyze this hypothesis than for most others. 20 For simplicity, we report the result using a dichotomized conflict measure in this table. 22 This effect is substantively interesting, and is simulated in Figure 5. It shows that an increase in PKO budgets from USD 7 million, the budget of a small observer force like the 1988– 1991 United Nations Iran-Iraq Military Observer Group, to USD 400 million, the budget of a force the size of United Nations Interim Force in Lebanon (UNIFIL), doubles the expected count of aid worker attacks. At first glance this might appear disheartening – deploying PKO forces is associated with more attacks on aid workers.21 This finding, however, is mediated by the result for the type of PKO mandate (Hypothesis 10c) reported in Column 5. Here PKOs with traditional mandates are indeed associated with more aid worker attacks; whereas there is no effect for PKOs with transformational mandates. Figure 5: Expected number of aid worker attacks as (log) PKO budget increases (from Table 4). This result then is fully in line with recent research showing that only PKOs with transformational mandates are effective at reducing conflict (Hegre et al. 2011). In contrast, traditional peacekeeping forces – mostly tasked with observing the terms of truce or peace agreements, or policing a buffer zone and assisting in negotiating a peace agreement – are not equipped to use lethal force to protect themselves and civilians, and seem less capable of creating 21 Conversely, the UN may selectively deploy larger PKOs to situations where aid workers are more likely to be attacked. We cannot rule out such a relationship using our approach here. If anything, however, this contradicts the more general argument that PKOs are only deployed to the “easy cases”, yet recent studies that try to tackle this endogeneity find no support for this type of argument (Vivalt forthcoming). 23 environments safe for aid workers. Nevertheless, we find little clear support for Hypothesis 10 in general, as the presence of PKOs, either traditional or transformational, and regardless of budget, does not create a situation on the ground where we see fewer aid worker attacks.22 6 Limitations to the analysis This study is among the first to analyse the determinants of attacks against aid workers. As such, we recognize the present analysis has a number of limitations that future research should address. First among these is the quality of the data. Despite the AWSD providing an excellent source of data, refinements and improvements may include researchers addressing: the quality of coverage and extent of reporting of attacks; unpacking what constitutes an ‘incident’; and – important in determining risk rates – accurately defining and determining the number of workers in the field. Fast (2010: 368) notes that: “A major problem with the existing research and documentation is the lack of consistency in definitions and statistics of the various data sources, which makes it very difficult to determine whether aid workers are targeted in increasing numbers and with increasing severity. One problem…is defining exactly what constitutes a ‘security incident’ and who is an ‘aid worker.’” While the data landscape can be improved, we note that if a definition is consistent over time, and a database populated with a consistent definition is used, then it does strengthen our ability to draw inferences. However, we are still left with limitations connected especially to the highly aggregated nature of available data, and are unable to perform potentially important sub-national analysis. Further, data is ill-equipped to examine certain questions including: how agency deployment methods influence the number of attacks per aid worker; which types of aid workers 22 In addition to analyses reported here, in Table 5 in the appendix we report robustness tests that check first (1) whether the results are sensitive to modeling excess zeros in the data or (2) sensitive to assuming a Poisson instead of a Negative binomial distribution. We find our results are not sensitive to these modeling choices. Second, it could be that aid worker attacks are associated with a particular set of countries or that we have omitted time-constant country level variables. We therefore estimate OLS regressions with country (1) fixed and (2) random effects. Again our results are robust to estimating such models. 24 under which conditions are most at risk; and, to an extent, are we even counting all attacks that occur.23 Regarding the spatial unit of analysis, a large literature has documented the local-level dynamics of conflict (Buhaug and Gates 2002). In conducting a country-year level analysis, we aggregate what are likely often local phenomena up to the country level. This potentially masks large within-country variation in the location and causes of aid worker attacks, which may be important in countries such as the Democratic Republic of Congo and Syria. Therefore, since we lack geo-located data on where aid worker attacks occur, it is possible that attacks we associate with on-going conflicts may occur far from actual conflict zones. Alternatively, in certain countries aid workers may stay clear of active conflict zones, reducing their exposure and risk and thus depressing the number of reported attacks in the country overall. Moreover, data may also include attacks possibly unrelated to field activities done by aid workers. Humanitarian security may also be affected by decisions regarding the characteristics of aid workers deployed. As noted, there are well-documented concerns that national staff are placed in more precarious situations than international counterparts. Yet given their local knowledge and profile, local workers may be a safer alternative to visible expatriate staff in dangerous settings. Such dynamics could influence the results reported above, but unfortunately, we lack adequate staffing data to either test or control for this using the empirical strategy here. Getting closer to these dynamics might necessitate in-depth case studies of large international humanitarian organizations. This would enable an examination of precisely the factors that influence how organizations make decisions about how, why, who and where to deploy. Combining this with case studies of conflict dynamics in relevant countries could allow a mapping of how conflict dynamics interact with various aspects of humanitarian organisations’ operations. Further, it is entirely possible that different humanitarian organizations have fundamentally different ways of dealing with the risk of attacks; and ideally comparative research would examine different organizations security protocols and operational responses to on-ground risk. At this point, however, the data does not allow for this. 23 In addition to those discussed here, certain limitations may relate to specific aspects of the data. For example, kidnappings are under-reported for various reasons, including reputational issues for INGOs and not ‘setting a precedent’. Killings, conversely, are nearly always reported. Economic attacks (such as kidnapping or extortion) might therefore be under-represented in the data. 25 7 Conclusion In this article we have considered the changing dynamics of humanitarian risk and integration to inform an examination of the factors associated with violence against aid workers. We find that the dynamics of conflict matter in understanding attacks on aid workers. In particular, more intense conflicts see more attacks, which is somewhat disheartening considering this are precisely where aid workers are most needed. Interestingly, we find no effect of general levels of violence, or of NATO presence on the estimated number of aid worker attacks. Situations in which NATO deploys forces see just as many, or as few, attacks on aid workers as those where NATO is not present. Due to ground-breaking data collection and research efforts (Humanitarian Outcomes 2015), the knowledge base on humanitarian insecurity is increasing. The crucial next step is to unpack the micro-level dynamics of these attacks. We already, to a large extent, have the data on conflict dynamics to do such an analysis, yet still lacking is systematic, fine-grained information about who, where, and what aid workers are doing. Addressing this might involve research collaboration with humanitarian agencies to access information about staff deployment patterns; undertaking fieldwork or working with disaggregated data within particular countries or conflicts to understand deep case-specific processes; or interviewing perpetrators of violence to understand motivations for attacking aid workers in different contexts.24 Finally, given the re-conceptualization of humanitarianism in the 21st century,25 it is important to consider impacts of new operational paradigms and evolving threats to aid workers for future humanitarian engagement. Our hope is that this article contributes to refining knowledge about the conditions underlying violence against humanitarian workers, and can assist aid organizations to better prepare for and prevent humanitarian attacks. 24 See also Spang (2009) 25 For instance, Rubenstein (2014) argues for moving beyond the humanitarian paradigm to restore humanitarian space; Briscoe (2013) argues for creating new humanitarian spaces in authoritarian, criminal or otherwise inherently violent areas not officially considered conflict zones but nevertheless of heightened interest to INGOs. 26 8 Bibliography Abiew, Francis Kofi (2012) Humanitarian action under fire: Reflections on the role of NGOs in conflict and post-conflict situations. International Peacekeeping 19(2): 203-216. Action contre la Faim (ACF International) (2015) Ensuring the Protection Aid Workers: Why a Special Mandate Holder is Necessary, Discussion paper. Available at: http://www.actioncontrelafaim.org/sites/default/files/publications/fichiers/discussion-paper- smh-131015.pdf. Accessed 16 November, 2015. Active Learning Network for Accountability and Performance in Humanitarian Action (Alnap) (2014). The State of the Humanitarian System: Assessing Performance and Progress. London: Alnap. Avant, Deborah & Virginia Haufler (2012) Transnational organizations and security. Global Crime 13(4): 254-275. Barakat, Sultan; Sean Deely, & Steven Zyck (2010) ‘A tradition of forgetting’: Stabilisation and humanitarian action in historical perspective. Disasters 34(3): 297-319. Barnett, Katy (2004) Security report for humanitarian organizations. ECHO Security Review. Brussels: Directorate-General for Humanitarian Aid. Barnett, Michael (2005) Humanitarianism Transformed. Perspectives on Politics 3(4): 723-735. Barry, Jane & Anna Jefferys (2002) A bridge too far: Aid agencies and the military in humanitarian response. Network paper 31, London: Humanitarian Practice Network. Brooks, Julia (2015) Humanitarians under attack: Tensions, disparities, and legal gaps in protection. ATHA White Paper Series. Cambridge: Harvard Humanitarian Initiative. Buchanan, Cate & Robert Muggah (2005) No Relief: Surveying the Effects of Gun Violence on Humanitarian and Development Personnel. Geneva: Small Arms Survey. Buhaug, Halvard & Scott Gates (2002) The geography of civil war. Journal of Peace Research 39(4): 417-433. Bureau of Labor Statistics (BLS) (2014) National Census of Fatal Occupational Injuries in 2014. Available at: http://www.bls.gov/news.release/archives/cfoi_09172015.pdf. Accessed 10 November, 2015. Carmichael, Jason-Louis & Mohammad Karamouzian (2014) Deadly professions: Violent attacks against aid-workers and the health implications for local populations. International Journal of Health Policy and Management 2014(2): 65-67. 27 Cederman, Lars-Erik; Kristian Skrede Gleditsch, & Halvard Buhaug (2013) Inequality, Grievances, and Civil War. Cambridge: Cambridge University Press. Chandler, David (2001) The road to military humanitarianism: How the human rights NGOs shaped a new humanitarian agenda. Human Rights Quarterly 23(3): 678-696. Choi, Seung-Whan & Idean Salehyan (2013) No good deed goes unpunished: Refugees, humanitarian aid, and terrorism. Conflict Management and Peace Science, 30(1): 53-75. Combaz, Emile (2013) The impact of integrated missions on humanitarian operations. GSDRC Working Paper. Birmingham: GSDRC, University of Birmingham. Dandoy, Arnaud & Marc-Antoine Pérouse de Montclos (2013) Humanitarian workers in peril? Deconstructing the myth of the new and growing threat to humanitarian workers. Global Crime 14(4): 341-358. Donini, Antonio (2011) Between a rock and a hard place: Integration or independence of humanitarian action? International Journal of the Red Cross 93(881): 141-157. Donini, Antonio; Larissa Fast, Greg Hansen, Simon Harris, Larry Minear, Tasneem Mowjee, & Andrew Wilder (2008) The State of the Humanitarian Enterprise. Boston: Feinstein International Center, Tufts University. Duffield, Mark (2012) Challenging environments: Danger, resilience and the aid industry. Security Dialogue 43(5): 475-492. Duffield, Mark; Joanna Macrae & Devon Curtis (2001) Editorial: Politics and humanitarian Aid. Disasters 25(4): 269-275. Eck, Kristine & Lisa Hultman (2007) One-sided violence against civilians in war: Insights from new fatality data. Journal of Peace Research 44(2): 233-246. Egeland, Jan (2011) To Stay and Deliver: Good Practice for Humanitarians in Complex Security Environments. Geneva: OCHA. Fajnzylber, Pablo; Daniel Lederman & Norman Loayza (2002) What causes violent crime? European Economic Review 46(7): 1323-1357. Fast, Larissa (2010) Mind the gap: Documenting and explaining violence against aid workers. European Journal of International Relations 16(3): 365-389. Fast, Larissa (2014) Aid in Danger: The Perils and Promise of Humanitarianism. State College: University of Pennsylvania Press. Ferreiro, Marcos (2012) Blurring of Lines in Complex Emergencies: Consequences for the Humanitarian Community. The Journal of Humanitarian Assistance, online. 28 Fox, Sean & Kristian Hoelscher (2012) Political order, development and social violence. Journal of Peace Research 49(3): 431-444. Guardian, The (2014). The 10 World Cities with the Highest Murder Rates. Available at: http://www.theguardian.com/cities/gallery/2014/jun/24/10-world-cities-highest-murder-rates- homicides-in-pictures. Accessed 10 November, 2014. Hegre, Håvard; Lisa Hultman & Håvard Mokleiv Nygård (2011) Simulating the effect of peacekeeping operations, 2010–2035. In John Salerno et al, eds. Social Computing, Behavioral- Cultural Modeling and Prediction. Berlin Heidelberg: Springer: 325-332. Hegre, Håvard & Nicholas Sambanis (2006) Sensitivity analysis of empirical results on civil war onset. Journal of Conflict Resolution 50(4): 508-535. Hilbe, Joseph (2011) Negative Binomial Regression. Cambridge: Cambridge University Press. Humanitarian Outcomes (2013) The New Normal: Coping with the Kidnapping Threat. Aid Worker Security Report 2013. New York: Humanitarian Outcomes. Humanitarian Outcomes (2014) Unsafe Passage: Road Attacks and their Impact on Humanitarian Operation. Aid Worker Security Report 2014. New York: Humanitarian Outcomes. Humanitarian Outcomes (2015) Aid Worker Security Database. Available at: www. https://aidworkersecurity.org. Accessed 19 November 2015. InterAction (2011) A Humanitarian Exception to the Integration Rule. Available at: http://www.interaction.org/document/interaction-statemetn-un-integration. Accessed 25 August 2015. Kalyvas, Stathis (2006) The Logic of Violence in Civil War. Cambridge: Cambridge University Press. King, Gary; Michael Tomz & Jason Wittenberg (2000) Making the most of statistical analyses: Improving interpretation and presentation. American Journal of Political Science 44(2): 347- 361. Knutsen, Carl Henrik & Håvard Mokleiv Nygård (2015) Institutional characteristics and regime survival: Why are semi-democracies less durable than autocracies and democracies? American Journal of Political Science 59(3): 656-670. Lischer, Sarah (2007) Military intervention and the humanitarian “force multiplier”. Global Governance: A Review of Multilateralism and International Organizations 13(1): 99-118. 29 Medicines Sans Frontieries (MSF) (2014) In the Eyes of Others: How People in Crises Perceive Humanitarian Aid. Geneva: MSF. Miklian, Jason (2014) The Past, present and future of the liberal peace. Strategic Analysis 38(4): 493-507. Miklian, Jason; Åshild Kolås & Kristoffer Liden (2011) The perils of ‘going local’: Liberal peace-building agendas in Nepal. Conflict, Security and Development 11(3): 285-308. Mills, Kurt (2005) Neo-humanitarianism: The role of international humanitarian norms and organizations in contemporary conflict. Global Governance 11(2):162-168. More Altitude (MA) (2015) World Humanitarian Day 5 Years On: 5 Security Trends. Available at: https://morealtitude.wordpress.com/tag/aid-worker-statistics/. Accessed 27 August 2015. Moser, Caroline (2004) Urban violence and insecurity: An introductory roadmap. Environment and Urbanization 16(2): 3-16. Olson, Lara (2006) Fighting for humanitarian space: NGOs in Afghanistan. Journal of Military and Strategic Studies 9(1): 1-28. Østby, Gudrun (2013) Inequality and political violence: A review of the literature. International Area Studies Review 16(2): 206-231. Pringle, Catherine & Derica Lambrechts (2011) The risk of humanitarianism: Towards an inclusive model. Strategic Review for Southern Africa 33(2): 51-80. PRS Group (2015) International Country Risk Guide Methodology. Political Risk Group. Available at: http://www.prsgroup.com/PDFS/icrgmethodology.pdf. Accessed November 16, 2015. Richmond, Oliver (2007) Emancipatory forms of human security and liberal peacebuilding. International Journal 62(3): 458-477. Rowley, Elizabeth; Lauren Burns and Gilbert Burnham (2013) Research review of nongovernmental organizations’ security policies for humanitarian programs in war, conflict, and postconflict environments. Disaster Medicine and Public Health Preparedness. 7(3): 241- 250. Rubenstein, Leonard (2014) A way forward: in protecting health services in conflict: Moving beyond the humanitarian paradigm. International Review of the Red Cross 95(890):331-340. Sandvik, Kristin Bergtora & Kristian Hoelscher (2016) The Reframing of the War on Drugs as a “Humanitarian Crisis”: Costs, Benefits and Attendant Consequences. Latin American Perspectives, forthcoming. 30 Sheik, Mani; Maria Gutierrez, Paul Bolton, Paul Spiegel, Michel Thieren & Gilbert Burnham (2000) Deaths among humanitarian workers. BMJ 321(7254): 166-168. Sims, Christopher & Tao Zha (1999) Error bands for impulse responses. Econometrica 67(5): 1113-1155. Solt, Frederick (2014) The Standardized World Income Inequality Database. Working paper. SWIID Version 5.0, October 2014. Spang, Lyra (2009) The humanitarian faction: The politicization and targeting of aid organizations in war zones. International Affairs Review 18(1): online. Sparrow, Annie (2013) ‘Syria’s Assault on Doctors’. New York Review of Books. 3 November. Stoddard, Abby; Adele Harmer & Katherine Haver (2006) Providing Aid in Insecure Environments: Trends in Policy and Operations. London: Overseas Development Institute. Stoddard, Abby; Adele Harmer, & Victoria DiDomenico (2009) Providing aid in insecure environments: 2009 update. HPG Policy Brief 34. UCDP (2015) UCDP Battle-Related Deaths Dataset v.5-2015. Uppsala Conflict Data Program, www.ucdp.uu.se, Uppsala University. Vivalt, Eva (2016) Peacekeepers help, governments hinder. Journal of Conflict Resolution, forthcoming. Weidmann, Nils; Doreen Kuse & Kristian Skrede Gleditsch (2010) The geography of the international system: The CShapes dataset. International Interactions 36(1): 86-106. Weiss, Thomas (2013) Humanitarian Business. London: Polity. Wille, Christina & Larissa Fast (2013) Security Facts for Humanitarian Agencies. Shifting Patterns in Security Incidents Affecting Humanitarian Aid Workers and Agencies: An Analysis Of Fifteen Years Of Data (1996-2010). Vevey, Switzerland: Insecurity Insight. Wood, Reed & Christopher Sullivan (2015) Doing harm by doing good? The negative externalities of humanitarian aid provision during civil conflict. Journal of Politics 77(3): 736- 748. 31 9 Appendix Table 5: Robustness tests for core conflict dynamics model, attacks on aid workers, 1997 -- 2014 (1) (2) (3) (4) ZINB Poisson Fixed Random Conflict 0.521*** 0.919*** 2.042*** 1.790*** (0.150) (0.038) (0.212) (0.183) Polity 2 -0.00371 -0.000628 0.138*** -0.00982 (0.015) (0.005) (0.041) (0.012) Polity^2 -0.0104*** -0.00835*** 0.00296 0.000557 (0.003) (0.001) (0.007) (0.003) ln(population) 0.158* 0.189*** 2.585*** -0.00494 (0.065) (0.019) (0.697) (0.047) ln(GDP capita) -0.417*** -0.304*** 0.439 -0.103 (0.086) (0.027) (0.403) (0.071) ln(Time in peace) -0.265*** -0.383*** -0.426*** 0.0930 (0.062) (0.030) (0.120) (0.060) Time since regime change 0.00648 -0.00810*** -0.00458 -0.000301 (0.004) (0.002) (0.011) (0.002) Aid worker attacks (t-1) 0.0959*** 0.0349*** 0.412*** 0.664*** (0.016) (0.001) (0.020) (0.016) _cons 2.033* 0.529* -27.10*** 0.689 (0.892) (0.255) (6.134) (0.679) inflate Conflict -1.022** (0.385) Battle deaths -0.00357 (0.002) _cons 0.791** (0.252) lnalpha _cons 1.035*** (0.165) AIC 2740.1 5952.2 12449.7 . ll -1357.1 -2967.1 -6215.9 N 2480 2480 2480 2480 Column 1 reports a zero inflated model. For this, we use the intensity of the conflict as covariates in the first step (the inflation step) of the model. This, not surprisingly, reduces the effect of conflict on the estimated count of aid worker attacks but the effect is still clearly significant. Column 2 estimates a Poisson model instead of a negative binomial model, again the effect of conflict, and indeed all the other covariates, is robust to this. Columns 3 and 4 and estimate OLS models with country fixed (Column 2) and random (Column 4) effects. If anything the fixed effects regression that only focuses on within country variation, and thus represents our most 32 conservative estimate, show a stronger effect of conflict on aid worker attacks than the above results. The results also hold for the random effects estimation. The map in Figure 6 shows, for reference, the sum of the number of aid worker attacks for each country in the world in 2014. Figure 6: Aid worker attacks by country, 201426 Attacks 0 1 2 3 4 5 9 10 13 18 22 23 25 42 127 26 Map produced using CShapes (Weidmann et al 2010). 33