Tectonic Tensions: Violence as an Interaction Across Space in Northern Ireland Hannes Muellery, Dominic Rohnerz, David Schoenholzerx October 31, 2013

Abstract This article presents a model of con‡ict that is robust to the unit of aggregation. The framework models violence as an interaction across space. Depending on the level of aggregation, interactions across the units of observation can cause most violence. We call this tectonic tensions and show how they can be captured empirically. The exclusion of tectonic tensions can lead to severe omitted variable bias in empirical studies that use disaggregated data. The extent of this bias is illustrated with data on religious composition and violence from Northern Ireland. In addition, we model and estimate the impact of spatial separation on levels of violence and show that our measure of tectonic tensions can predict the placement of barriers (i.e. so-called "peace lines") in Northern Ireland. The article discusses the importance of military technologies for the spatial distribution of violence. Finally, we argue that tectonic tensions add a level of complexity to cross-country studies of ethnic con‡ict. Keywords: Con‡ict, Ethnic Violence, Spatial Data, Transport Costs, Polarization, Segregation, Northern Ireland, Insurgency. JEL Classi…cation: D74, K42, N44, Z10.

Acknowledgements: We thank Joan Maria Esteban and Debraj Ray for very useful comments. We also thank seminar participants at the UC Berkely, NYU Abhu Dhabi and CEMFI Madrid. All errors are all ours. y Institut d’Analisi Economica (CSIC), Barcelona GSE. Email: [email protected]. z Department of Economics, University of Lausanne. Email: [email protected]. x Department of Economics, UC Berkeley. Email: [email protected].

1

1

Introduction

Phenomena like con‡ict are characterized by the fact that their spatial spread does not necessarily correspond to the political borders of a country and that the intensity of violence varies locally. Hence, a challenge faced by research is to determine the right spatial level of analysis. Up until recently, the default option was to use as units of observation a country. More recently, to address worries of unobserved heterogeneity, many studies have performed analysis at a much more disaggregate level such as local districts or cells in GIS-based raster data. However, the more disaggregated the data the more important become interactions between the di¤erent units. Casualties in a community are, for example, often not caused by perpetrators from within the same community. This is both an empirical problem and a theoretical one. Figure 1 illustrates the shortcoming in the current state-of-the-art. The …gure shows a spatial distribution of characteristics A and B and two violence distributions. Violence, indicated by a grey shading, in the example on the left is distributed randomly across areas with characteristic A. Violence in the example on the right is concentrated at the boundary between characteristic A and B. The standard method of regressing violence on characteristics within units will not indicate any di¤erence between these two patterns. Both examples will register as a correlation of violence with characteristic A. Yet, violence in the right graph is likely generated by an interaction between characteristics A and B. This article develops a methodology that distinguishes between the two cases in Figure 1 and makes the analysis immune to di¤erent levels of aggregation. We start with the assumption that con‡ict is motivated by factors at the aggregate level. We posit that a) the likelihood of achieving a group’s military and political goals is an increasing function of the fatalities it causes, b) recruitment and support for paramilitary organizations active in con‡ict are local, c) speci…c individual characteristics (religion, ethnicity, socioeconomic status) determine the likelihood of supporting certain armed groups. To these standard ingredients we add the element of "transport cost of violence", i.e. chances of success are smaller when attacking targets that are further away. We show that this element is crucial if one wants to predict spatial violence patterns. If one assumes positive transport costs this gives rise to what we call tectonic tensions, the occurrence 2

of violence at interfaces between homogenous areas of opposed groups. Tectonic tensions do not arise if one assumes negligible transport costs. We use this simple framework to analyze the religious dimension of the Northern Irish con‡ict. Northern Ireland –being a rare example of a developed country experiencing an intense con‡ict– provides a unique setting that allows us to match detailed con‡ict events and location data with …ne-grained census data on the exact number of members from di¤erent religious groups in 582 local administrative wards (in a robustness check we also use district and cell-level data). We can observe patterns of violence by state forces, republicans and loyalists separately. We show that the con‡ict between loyalists and republicans, i.e. sectarian violence, can be predicted extremely well with tectonic tensions. In order to show that tectonic tensions were also a crucial factor for government policy we study the placement of barriers between religious communities in Northern Ireland. Many of these walls were erected on ward boundaries. We show that the placement of a wall on a boundary can be predicted by tectonic tensions. The baseline risk of construction more than doubles with a standard-deviation increase in tectonic tensions. This result holds up to controlling for ward …xed e¤ects and within-ward religious composition on both sides of the ward boundary. While neglecting local interaction is a general methodological problem for all phenomena with spatial heterogeneity, it is particularly pressing for the study of "complex warfare". Civil con‡icts often blur the traditional distinction between insurgency and sectarian violence. The current con‡icts in Syria, Mali, Iraq and Afghanistan, for example, share elements with traditional guerrilla warfare, but also feature a large amount of violence between di¤erent religious or ethnic groups. Our data allows us to distinguish between these sectarian and insurgent elements of con‡ict. Violence with state involvement follows a di¤erent pattern than violence between the religious communities. Insurgent and counter-insurgent violence takes place to a large extent in Catholic strongholds irrespective of their neighborhood composition. We argue that this is due to the military balance between state forces and republican paramilitaries. Overwhelming military power meant that state forces patrolled directly in Catholic areas regardless of neighborhood

3

composition. This low level of transport costs for the state forces led to a very di¤erent geographic distribution of casualties with state involvement. The spatial distribution of ethnic groups is an important determinant of levels of violence. This has consequences for empirical studies of violence. Empirical work with disaggregated data faces the problem of omitted variable bias and is in danger of missing tectonic tensions. Our framework provides a simple test for this bias. Country-level analysis of ethnic polarization and factionalization could also bene…t from taking spatial clustering into account. If opposed groups live far apart, violence can be low even if polarization is high. If groups are mixed then violence is likely higher. The paper is organized as follows: Section 2 surveys the related literature, while Section 4.1 provides a discussion of the context of the "Troubles" in Northern Ireland. In Section 3 we set up a simple formal model, in Section 4 we present the data and empirical strategy, and in Section 5 we carry out the econometric analysis and present the results. In section 7 we use the model to shed light on several aspects of the con‡ict in Northern Ireland. Section 6 discusses what lessons our …ndings bear for other studies using cross-country or cell-level data analyzing con‡ict or other phenomena. Section 7 concludes.

2

Related Literature

Our paper is related to the literature on ethnic and religious con‡ict between domestic groups. Most theoretical papers in economics model ethnic con‡ict as a strategic interaction between a small number of aggregate players on the nation level (Esteban and Ray, 2008, 2011; Caselli and Coleman, 2013). These frameworks are able to show that ethnic con‡ict is more salient than class con‡ict and that the risk of turmoil increases in ethnically polarized societies, i.e. in societies with a small number of sizeable groups. Rohner (2011) builds one of the rare models of ethnic con‡ict where interaction and social tensions happen at a disaggregate individual level, and …nds that ethnic fractionalization, polarization and segregation fuel con‡ict. Recently, the nexus between ethnic con‡ict and trust has also received attention by the theoretical literature (Rohner, Thoenig and Zilibotti, 2013; Acemoglu and Wolitzky, 2012). In other strands of the

4

theoretical con‡ict literature, Besley and Persson (2010, 2011) focus on the role of state capacity in civil wars, while Morelli and Rohner (2013) emphasize the impact of natural resources on civil con‡ict. However, while all of these contributions study the overall, aggregate likelihood of con‡ict, none of them contains predictions of spatial violence patterns on the sub-national level. This paper o¤ers a way to translate theories on the national level into predictions of spatial violence patterns. The simple idea is that, for a given level of motivation, violence is increasing in the number of targets and perpetrators of violence. To the best of our knowledge this, almost trivial, point has been disregarded up until now. Most empirical studies of ethnicity and civil war focus at the country year level. While the impact of ethnic fractionalization on ethnic …ghting has been found to be ambiguous (cf. Fearon and Laitin, 2003; Collier and Hoe- er, 2004; Collier and Rohner, 2008; Collier, Hoe- er, and Rohner, 2009), it has been found that ethnic polarization fuels the risk of civil war (Montalvo and Reynal-Querol, 2005; Esteban, Mayoral and Ray, 2012). Cederman and Girardin (2007) …nd that ethno-nationalist exclusion of minority groups increases the risk of ethnic con‡ict. Michalopoulos and Papaioannou (2011) …nd that the division of ethnic groups by arbitrary national boundaries leads to con‡ict. There is an increasing number of papers that study violence at a disaggregate, local level (e.g. Rohner, Thoenig, and Zilibotti, 2013; Berman and Couttenier, 2013; Dube and Vargas, 2013), but most of these contributions do not contain a formal model of war and/or focus on geographical features related to characteristics of the terrain and the availability of natural resources. They typically do not take into account the local ethnic composition, usually due to data limitations, as it is extremely hard to …nd …ne-grained and reliable data on ethnic group location and size for politically unstable countries. There are a few papers selecting an intermediate level of disaggregation and building a panel dataset on the ethnic group level covering a large number of countries (cf. for example, Buhaug, Cederman, and Rod, 2008; Esteban, Morelli and Rohner, 2012). However, their level of disaggregation is still much less …ne-grained than in the current paper, and they do not focus on local

5

ethnic cleavages and the interaction of ethnic groups across regions. Research that studies insurgency and violence for one country at a very …ne-grained level is still rather scarce. Kalyvas (2006) argues that …ghting groups use a combination of persuasion and coercion to win support of the local population and to extract important intelligence information about their opponents. The insurgents and governing forces use –especially in regions of incomplete control–discriminate or indiscriminate violence in the goal of establishing control over an area. As indiscriminate violence is ine¢ cient, armed groups prefer to apply discriminate violence whenever intelligence information permits. Kalyvas stresses the role of military control for recruitment into armed groups. Our work can be regarded as translating Kalyvas’ideas of military control into a context in which population characteristics a¤ect the ease at which paramilitaries recruit. More importantly, perhaps, our model allows for an empirical study of the interaction of geographic units across space. We see this as a natural way of extending the idea of territorial control and bringing it to the data.1 There is a small literature in political science studying –inspired by the epidemiological literature on the spread of diseases–di¤usion and clustering patterns of violence over space and time (Townsley, Johnson, and Ratcli¤e, 2008, Schutte and Weidmann, 2011). Further, also VectorAutoregression (VAR) models have been used to study cyclical …ghting: for example, Jaeger and Paserman (2008) study whether there is "tit-for-tat" over time in the Palestinian-Israeli con‡ict. They …nd using VAR and Granger causality tests that while Israeli troops strike back promptly and strongly after Israeli fatalities, a similar retaliation is not observed for Palestinian …ghters. Our study draws attention to the fact that Palestinian killings of Israelis trigger a response not locally but many kilometers away. Without a clear classi…cation of victim and perpetrator of violence in the data this study would therefore not be possible. There are also papers studying the e¢ ciency of particular counter-insurgency strategies: Lyall (2010) …nds that village "sweep"operations carried out by pro-Russian Chechen forces are more

1

Bhavnani et al (2011) provide a model of control in geographic space with three groups but do not test this

aspect empirically.

6

e¢ cient in reducing posterior retaliation attacks than similar "sweep" operations performed by Russian troops, which would be consistent with the view that a "coethnicity advantage" in counter-insurgency exists due to better information. Kocher, Pepinsky and Kalyvas (2011) …nd that U.S. aerial bombing –a form of indiscriminate violence–in Vietnam was counter-productive and resulted in an increased likelihood that the Viet Cong ultimately gained control of an a¤ected area. Berman, Shapiro and Felter (2011) build a model with government, rebels and civilians, where the government has the choice between repression and public good provision. They then show empirically for Iraqian microdata that public good provision has reduced insurgency e¤orts. We show in one of our applications that the troop surge in Northern Ireland lowered republican violence beyond areas most directly a¤ected. This externality is an important factor for studies using disaggregated data. Novta (2013) builds a simulation-based model of how con‡ict spreads. Contrary to our setting that models a situation of insurgency and terrorism, her framework is designed to study traditional military warfare between two standing armies. The features of her model are found to be consistent with the spread of violence in the 109 municipalities of Bosnia. Novta models the armed groups in each municipality as separate, myopic players who can only attack in their home village while the focus of our, much simpler, framework lies precisely on the across-ward attacks. Predicting spatial violence patterns is important because violence a¤ects political and economic outcomes locally. Besley and Mueller (2012) show, for example, that there were distinct di¤erences in the economic e¤ect of the Northern Ireland con‡ict driven by di¤erences in local violence levels. Compared to peaceful areas, housing in the most violent areas sold for between 2 and 17 percent less - depending on the level of violence. If ethnic tensions have an impact on countries then we would expect these to be biggest in the violent areas. And, depending on how predictable violence is, we might …nd areas with no violence to be completely una¤ected.

7

3

Model

3.1

The Set Up

We model a violence in a country with W regions indexed by w. In each region live two groups g 2 fc; pg, of size Cw and Pw . Regions have the population size Nw = Cw + Pw . Denote the total number of members of the two groups in all W regions as C and P . Violence in these regions is conducted by two paramilitary groups which recruit themselves from the groups g. The e¤ort in con‡ict (support for paramilitaries, deployment of troops...) in these two groups is given by

p

and

c.

Assume, for example, that

g

is the likelihood that a

given member of group g joins the paramilitaries. In order to follow the standard cross-country model we assume that this e¤ort is derived from a contest function at the aggregate level. The equilibrium likelihood of joining the paramilitaries in this sort of model follows the form

where

p

= Fp (C; P; )

c

= Fc (C; P; )

is the deeper motivation behind violence (natural resource rents, taste di¤erences in

public goods, etc.). The standard contest function in this example would make the likelihood

p

decreasing in P and increasing in C and . In this model the likelihood

g

is constant across the population. The expected number of

perpetrators from group p in region w is therefore given by number of perpetrators from group c in region w is

3.2

p Pw .

Analogously, the expected

c Cw .

Model Without Transport Costs

Assume no transport costs of violence. Each member of group C can then attack each member of group P - regardless of the distance between them. The perpetrators ( p P and

c C)

in this

model are able to attack every member of the opposed group with equal probability. To translate this into violence patterns at the local level we assume that violence takes place where the targets

8

live.2 The expected number of casualties in group c in region w is then proportional to EA (P; Cw ) /

pP

Cw

(1)

and analogously, the expected number of casualties in group p in region w is proportional to EA (C; Pw ) /

cC

Pw :

An empirical study with micro data will not be able identify con‡ict as an interaction. This is because the total number of perpetrators

cC

and

pP

does not vary at the local level. This

makes violence in w proportional to the number of targets Cw and Pw . In other words, the interaction between the groups becomes indistinguishable from violence being directly generated by characteristic Cw (and Pw ). We will return to this …nding below.

3.3

Model With Transport Costs

What happens if we add the notion of "transport costs" to this model? Transport costs imply that the distance between the "base" of the perpetrator and the target matters. Call the distance between region w and region v, d(w; v). The average distance d(w; w) between individuals within the same geographic units is d0 and across regions we have a …nite number of possible distances D such that d0 < d1 < ::: < dD . It will be useful to de…ne neighborhoods n1 (w) :::nD (w) as those regions that are at distance d1 :::dD to region w. We expect a positive transport cost of violence, which implies that the probability of a successful attack decreases with distance. We denote the likelihood that a given individual in group c successfully attacks a given individual in group p at distance di as Pr (c; di )

c k(di );

where k(d) is the decay function of violence. We assume that k(dj ) < k(dj

2

3 1 ).

An alternative is to model the allocation of violence as a Colonel Blotto game. The empirical implications of

this alternative would be less clear-cut. 3 This assumption is very important for the conclusions we draw from our framework. We therefore test it empirically.

9

Analogously the likelihood that a given individual in group p successfully attacks a given individual in group c at distance di is Pr (p; di )

p k(di ):

The probabilities Pr (c; di ) and Pr (p; di ) are best thought of as the likelihood that an interaction between members of the two groups lead to a casualty given that this interaction takes place across distance di . In the empirical section we will estimate these probabilities directly. For simplicity, we maintain the assumption of constant

p

and

c.

This assumption makes

sense if, for example, individuals …rst choose to join the paramilitaries and then realize their local e¤ectiveness, k(d), at a later stage.4 The number of perpetrators from group p in region w, for example, is then still given by

p Pw .

Total attacks by group p on group c in w are then

EA (P; Cw ) = Pw +

Pr (p; d0 )

D X

Pni (w)

Cw Pr (p; di )

(2) Cw

i=1

where Pni (w) denotes the total number of group p members that live in the neighborhood ni (w) P Pr (p; di ) Cw therefore captures all the casualties caused of ward w. The term D i=1 Pni (w) by attacks on group c from outside of w. Total attacks by group c on group p in w are EA (C; Pw ) = Cw +

Pr (c; d0 )

D X

Cni (w)

Pw Pr (c; di )

Pw :

i=1

This model maintains the structure of the model without transport costs. The number P still appears in equation (2) but is now split into its elements Pni (w) and Pw .5 The di¤erent elements are then weighted by the probabilities Pr (p; d0 ) ; Pr (p; d1 ) ::: Pr (p; dD ). A lower weight is given to those members of group p that live furthest away from w.

4

While this assumption might appear very restrictive it is not. A model of endogenous e¤ort would make

decreasing function of transport costs. Con‡ict e¤ort

g

g

a

would then be smaller in regions that face higher transport

costs. This would not, however, change our results qualitatively as our model already predicts less violence in areas with higher transport costs. Subject to a functional form change in k(d) the two models would be indistinguishable. P 5 To see this note that P = D i=1 Pni (w) + Pw .

10

The change in Pr (c; di ) with di will depend on the violence technologies of groups c and p and the amount of military control the groups have. They can be estimated directly if data on Cw , Pw and d(w; v) is available. Note, that we made the implicit assumption that perpetrators that live close to many targets can in‡ict more casualties. The alternative would be to assume that every perpetrator can only hit a limited number of targets. Violence would then not increase in the number of targets.

3.4

Predictions

In what follows we will focus on outcomes in region w. We do this to be as close as possible to the empirical speci…cation used in most micro studies. These typically try to explain region-level outcomes (attacks in our case) with characteristics of these regions. Total expected attacks in region w are EAw = Pw

[Pr (p; d0 ) + Pr (c; d0 )] Cw ! D X + Cw Pni (w) Pr (p; di ) i=1 D X

+

Cni (w)

Pr (c; di )

i=1

!

(3)

Pw

where the …rst term describes fatalities that arise from interactions within the region and can be explained by its characteristics Cw and Pw . The term shows both the attacks from members of group p and c. The second and third term describe attacks that arise from interactions across the boundaries of the region. The second term describes victims of group c in region w who were attacked by group p in the regions in the neighborhood of w. The third term describes victims of group p in region w who were attacked by group c in the neighborhood of w. An obvious simpli…cation of the model is to assume that Pr (p; di ) = Pr (c; di ) = Pr (di ) so that we can rewrite equation (3) EAw = (2 +

Pw D X

Cw )

Pni (w)

Pr (d0 ) Cw + Cni (w)

(4) Pw

Pr (di ) :

i=1

This illustrates the main point of this paper. Micro studies which focus on interactions within 11

geographic units will miss violence that is generated by interactions across regional boundaries. We call our measure of this part of violence tectonic tensions. Proposition 1 (TECTONIC TENSIONS) An explanation of violence at the local level that disregards interactions with regions in the neighborhood omits the terms TT =

D X

Pni (w)

Cw + Cni (w)

Pw

Pr (di ) :

(5)

i=1

Tectonic tensions are generated by the interaction of the groups across ward boundaries. The terms Pni (w)

Cw

Pr (di ), for example, capture attacks from the neighborhood on Cw : If these

are likely to succeed (high Pr (di )) then tectonic tensions become an important driver of local violence. This is particularly important in areas that are pure, i.e. in regions that have Pw = 0 or Cw = 0. The only way for violence to arise are then attacks from outside. In our empirical implementation we will focus mostly on the most immediate interaction of regions with their immediate surroundings. These are captured by the term T T (1) = Pn1 (w)

Cw + Cn1 (w)

Pw

Pr (d1 ) :

(6)

The presence of tectonic tensions can lead to biased estimates of the e¤ect of within-region characteristics. This is a particular problem for the study of within-region characteristics if these characteristics are correlated with T T . For a given population size Nw measures of within-region polarization and fractionalization, for example, are proportional to our measure of within-region tension Pw

Cw .6 Equations (3) and (4) can therefore be used to analyze the e¤ect of within-

region polarization. Assume, without loss of generality, that Pw

Cw . An increase in the

number of Cw then increases polarization and violence within the region w. We obtain the following proposition on the e¤ect of polarization: Proposition 2 (LOCAL POLARIZATION) Assume that group p is a majority in region w (Pw

6 P 2N

Cw ). For …xed population size in region w a marginal increase in polarization in region

For two groups the standard formulas for fractionalization and for polarization become, F RAC =

C N

P 2C C 2P P , and P OL = 4(( N ) N + (N ) N ) = 4N

C N

, which are both monotonically increasing in (P

12

P C N N

C).

C +N

P N

=

w (a marginal increase in Cw and simultaneous decrease in Pw ) changes the level of violence in that region by @EAw @P OL

= 2 +

(Pw D X

Cw )

Pr (d0 )

Pni (w)

Cni (w)

(7) Pr (di ) :

i=1

Proof. See the appendix. Intuitively, attacks within the region always increase in polarization. This is the idea behind testing con‡ict theories with micro data. Two problems complicate this endeavour. Changes in composition within a region a¤ect attacks from across the region’s boundaries. If, for example, PD Cni (w) Pr (di ) < 0 then an increase in Cw in region w leads to a reduction of i=1 Pni (w) attacks from other regions. This e¤ect can dominate if Pw and Cw are close to each other.

A second problem is scale. Our model suggests that changes in scale a¤ect the level of violence. Units that are more densely population will experience a higher absolute number of casualties. The term Pw

Cw captures these di¤erences in size. It increases with a higher number of persons

living in a geographic unit, even if the population share is constant. Our model also allows us to analyze the e¤ect of spatial separation of groups in con‡ict. Assume that group c is in the minority in region w and that the neighborhood is otherwise balanced. Then imagine that a member of group c in region w is swapped with a member of group p in a region v in the neighborhood. This "sorting" of the groups always leads to a reduction in violence overall. Proposition 3 (SEGREGATION) Assume Pw > Cw and Pu = Cu in all other regions u 6= w: An exchange of a marginal member of group c in region w with a member of group p in a region v at distance d (w; v) leads to a fall of expected attacks in region w by @EAw @SEG

=

2

(Pw

+ (Pw

Cw )

Cw )

Pr (d0 )

(8)

Pr (d (w; v))

and to a rise of attacks in region v by @EAv = (Pw @SEG

Cw ) 13

Pr (d (w; v))

(9)

so that the overall amount of violence falls i¤ Pr (d (w; v)) < Pr (d0 ). Proof. See the appendix. The …rst term in equation (8) captures the fact that interactions within region w are reduced. However, the exchange triggers more interactions across regional boundaries. The exchange of within-region interactions with across-region interactions decreases violence if Pr (d (w; v)) < Pr (d0 ). The core element is therefore the decay function. If transport costs are strictly increasing with distance violence will fall with segregation. It should be obvious from equations (3) and (4) that an increase in transport costs always decreases violence. But how does violence evolve with decreasing costs? It is useful to return to the model without transport costs to understand this. As we have shown before the link between the proximity of perpetrators and violence vanishes. Proposition 4 (TRANSPORT COSTS) Violence at distance di is decreasing with an increase in transport costs (decrease in k(di )). If transport costs of violence are negligible, we have Pr (di ) = Pr for all i = 0:::D: We can then express expected violence in equation (4) as EAw =

W X v=1

Pv

!

Pr Cw +

W X v=1

Cv

!

Pr Pw

(10)

The number of casualties a military organization can in‡ict is decreasing in the transport cost this group faces. This is the reason behind measures like road-blocks and other restrictions of movement commonly deployed in the context of con‡ict. The second part in proposition 4 takes up the results derived in the model without transport costs. Violence in this model is linear in Cw and Pw . Only the geographical distribution of targets then matters for the distribution of violence. Proposition 4 is particularly relevant in the context of asymmetric warfare where state forces are militarily more powerful than insurgent groups. In cases like Northern Ireland, where state forces stayed in control of most of the territory for most of the con‡ict, we expect transport costs to be extremely low. The most obvious signal of this asymmetry in the Northern Irish context was open patrolling of state forces in Catholic strongholds.

14

We now turn towards the e¤ect of urbanization, i.e. of increasing population density in some areas for a given total population (note that population growth would always increase the scope for violence). In order to capture the e¤ect of urbanization without altering local ethnic composition, we can focus on the (somewhat extreme) thought experiment of a country with the same ethnic composition in all regions. Now consider the case where the population of a whole area of the country decides to move to other parts. Proposition 5 (URBANIZATION) Assume a given total population, N , and a given constant ethnic population composition. A larger population density (i.e. more people per square kilometer of populated area) will reduce the average distances between group p and c and therefore increasing the total expected attacks. As an example think of climate change resulting in rougher climate conditions with frequent ‡oods and storms that make a whole part of the county unsuitable for settlement. While the total population and ethnic composition is unchanged, in such a case population density increases in the populated regions, as the same number of people now live closer together.

4

Background on the Con‡ict and Data

4.1

Context of Con‡ict in Northern Ireland

The Northern part of Ireland, Ulster, has been religiously divided since its conquest by England and the Reformation, taking both place in the 16th century.7 Since then the Catholic population from Gaelic Irish origin and the Protestant population of English and Scottish settlers have lived "separate lives" characterized by very stable patterns of land holdings and relatively few religiously mixed marriages (Mulholland, 2002). When the Republic of Ireland achieved independence from Britain in 1919, the six Northern counties of Ireland remained part of the UK.

7

This subsection draws heavily on Mulholland (2002).

15

In the early 1920s "Troubles" broke out with the Irish Republican Army (IRA) challenging British authority over Ulster and engaging in violent combats against the British troops and Protestant paramilitary organizations such as the Ulster Volunteer Force (UVF). The following decades were characterized by "home rule" and the new Parliament of Northern Ireland at Stormont near Belfast. The political divide persisted between the Catholic Nationalists (also called Republicans) who wanted to join the Republic of Ireland and the Protestant Unionists (also called Loyalists) who wanted to remain united with the UK. While in the 1950s and early 1960s there were relatively low levels of political violence, in 1968 the situation became again more confrontational when the Civil Rights Movements asked for more rights for Catholic citizens. Some of the initially peaceful demonstrations and marches were met with repression and resulted in fatalities. From August 1969 onwards sectarian violence exploded. In September 1969 radical militants took control of the previously dormant IRA and created its radical wing, the Provisional IRA. The "Provos" achieved an ever tighter grip of traditional Catholic working class strongholds like the Falls Road in Belfast or the Bogside in Derry. Further, alarmed by the rise of the IRA and the seeming willingness of the UK government to make political concessions, loyalist paramilitary organizations stepped up in the 1970s, intimidating Catholic families from mixed and Protestant areas and starting a violent campaign against civilian Catholics. After 1976 the UK built up a stronger Royal Ulster Constabulary (RUC) that together with the British Army and the SAS troops stepped up e¤orts to militarily weaken the IRA. This e¤ort led the loyalist paramilitaries to lower their violence and the IRA to retrieve from large-scale open confrontations and to adopt a cellular structure common in terrorist organizations. The con‡ict moved from an insurgent war to terrorist campaign. Even carefully planned attacks by the paramilitary groups had to rely on operational centres based on religion. Dillon (1999), for example, describes an IRA operation in October 1972 as follows: "The intelligence o¢ cer of the 1st Battalion said Twinbrook was the best for an assault on

16

the laundry van [...]. He reckoned that if the van was attacked in Twinbrook an IRA unit could make an escape with ease and be in the safety of the Andersontown district within a matter of …ve to ten minutes."(Dillon 1999, page 42). In Andersontown the 1971 census counted 5588 Catholics and 51 Protestants. The quote shows that the IRA was operating from and around this Catholic ward. This made attacks on Protestants and state forces close to Andersontown more likely.

4.2

Data

We use two main data sources. Data on religious composition is from the UK 1971 census and is provided by NISRA. Most data on violence comes from Sutton (1994) and has been updated by the Con‡ict Archive on the Internet (CAIN) website. We use address data in the description of killings to derive geo-references data. We then use these references to match killings to wards and grid-cells. It should be stressed that the violence data is unique as it reports the religion of each victim (unless for members of the state forces) and the group that attacked him or her. We have data on 582 wards and, alternatively, 3571 grid cells to which we aggregate the data on killings. Table 1 shows the summary statistics of all relevant variables. The number of Catholics (Cw in the model) and Protestants (Pw in the model) are in thousands. We also show all interaction variables used in the regression analysis. The …rst measure is the number of within-ward interactions. On the ward level the mean indicates that the average ward featured 720,000 interactions between Catholics and Protestants. The variable "tectonic tensions in n(1)" is our main variable and is calculated according to equation (6). The ward mean indicates almost 11 million interactions with population of the opposed religion in the direct neighborhood. The variable "tectonic tensions in n(2)" is calculated analogously only that now the neighborhood are those wards (cells) that are not direct neighbors but the neighbor’s neighbors. Table 1 also summarizes our data on con‡ict-related casualties. The special feature of this data is that it reports both perpetrators and victims of violence. This allows us to test our ideas about the data in detail. The main distinction we make is between violence involving state forces and violence that does not involve state forces. Table 1 shows that this splits the data roughly

17

in half. We also report summary statistics on the dyad level. Each dyad consists of two wards which are neighbors by the contiguity criterion. We discuss the structure of this data more in detail in section (5.2).

5

Empirical Results

In this section we illustrate our theoretical …ndings with the data from Northern Ireland. Section (5.1) explores the role of tectonic tensions. We contrast the standard model that uses only within-ward variation with our model and show that tectonic tensions are an important factor in explaining violence levels. This view is supported in section (5.2) where we show that tectonic tensions can predict the placement of barriers between wards. This highlights the crucial importance of interactions across space in the Northern Irish context. Section (5.3) investigates the role of transport costs empirically. We …rst show that the amount of violence per interaction indeed becomes smaller with distance. This is con…rmed in a speci…cation with electoral districts as unit of observation. We then use information on violence perpetrators and victims to distinguish violence with state involvement. Our results are as suggested by proposition 4. Violence with state involvement shows a much weaker relationship to tectonic tensions and is instead linear in the number of Catholics. Positive transport costs are crucial ingredients for tectonic tensions to arise. Finally, we run a robustness check using cell-level data in section (5.4). The importance of tectonic tensions in explaining violence is con…rmed with this data.

5.1

Tectonic Tensions

Figure 2 shows the relationship between Catholic-Protestant interactions and the absolute number of casualties at the ward level. Our measure of within-ward interactions (Pw

Cw ) is reported on

the X-axis. It takes higher values where a higher number of Catholics and Protestants interact within the ward boundaries. The positive slope in …gure 2 implies that a growing interaction between Protestants and Catholics leads to an increase in casualties. As we show below this

18

pattern holds up if we control for population. It is important to note that this pattern does not arise with the classic measure of polarization which norms the number of Catholics and Protestants by the population in the ward. Figure 3 shows the relationship between the standard measure of polarization (factionalization) and casualties.8 The positive relationship from …gure 2 disappears. However, …gure 3 reveals that violence is relatively high both in very polarized areas (on the right) and very homogenous areas (on the left). The results in …gure 3 speak against a model which endogenizes e¤ort with local characteristics. Such a model would yield the known result that more polarized wards are more violent. This is an important di¤erence to the cross-country literature which has focused on the motivation to engage in violence. In our model the number of possible interactions is key for explaining variation in violence. This gives scale an important role to play. The empirical evidence in …gures 2 and 3 supports this assumption. Figure 4 shows the impact of our measure of tectonic tensions in the immediate neighborhood T T (1) = Pn1 (w)

Cw + Cn1 (w)

Pw :

The variable on the X-axis now counts interactions of Catholics and Protestants across ward boundaries. Again we observe an increasing slope. This provides …rst evidence for the validity of proposition 1. Interactions across ward boundaries lead to more violence. Shaftsbury, for example, faces relatively high tectonic tensions due to the fact that over 17,000 Catholics and over 41,000 Protestants live in its direct neighborhood so that the ward experienced over 240 million interactions across ward boundaries.9 We now turn towards an explicit test of our theoretical …ndings. Table 2 shows our estimates of equations (3) and (4). Column (1) ignores interactions across ward boundaries. It con…rms the pattern from …gure 2. The coe¢ cient can be directly interpreted as Pr (d0 ) =2. In other words,

8 9

The same pattern arises if one correlates casualties per capita with polarization. T T (1) = 41 2:6 + 17 8:2 = 246

19

per one million interactions between Catholics and Protestants within a ward there were two additional casualties in the Troubles.10 Put di¤erently, an increase of one standard deviation in within ward interactions (SD = 1:72) leads to an increase in casualties by more than one standard deviation (SD = 2:71). Column (1) also controls for the number of Catholics and Protestants in the ward so that this pattern is not driven by variation in population. The number of Catholics correlates positively with violence. We will return to this …nding in the next section. Table 2, column (2) reports the equivalent regression for across-ward interactions. We …nd a strongly signi…cant, positive coe¢ cient on tectonic tensions which con…rms …gure 4. The coe¢ cient is smaller which already hints to a probability Pr (d0 ) > Pr (d1 ). However, with a standard deviation of almost 30 the explanatory power of tectonic tensions is very high. A one standard deviation increase in tectonic tensions leads to an increase of two standard deviations in casualties. Our estimate of equation (4) is in Table 2, column (3). Both within-ward and across-ward interactions still imply an increase in casualties. Tectonic tensions, in particular, remain an important driver of casualties. This con…rms proposition 1. The point estimate of the coe¢ cient on within-ward interaction more than halves when we add the tectonic tension variable. This is in line with the idea put forward in proposition 2 - interactions across geographic units can introduce biases in the within-ward analysis. Omitting interactions with the neighborhood could have lead to a considerable upward bias on the coe¢ cient in column (1). Column (4) runs the same regression with district …xed e¤ects as additional controls. Results are robust. Column (5) provides our estimates of equation (3). We split the tectonic variable into two interaction terms to capture the two directions of violence. The interaction between Protestants within the ward and Catholics in the neighborhood (Cn1 (w)

Pw ) is positive and highly sig-

ni…cant. This indicates that members of the Catholic community attacked Protestants across ward boundaries. The reverse is also true. The interaction of Catholics with Protestants in the

10

Note that attacks at this level are in both directions.

20

immediate neighborhood Pn1 (w)

Cw also led to more casualties.

What do these results tell us about the role of segregation that is stipulated in proposition 3? Table 2, column (3) shows that for an increase of Cw

Pw by 1, violence in each direction of

this interaction increases by Pr (d0 ) =

0:771 = 0:386: 2

Since both Cw and Pw are measured in thousands this implies that within a ward casualties increase by 0:771 for 1 million more interactions between Catholics and Protestants. From column (3) we can also read that an increase of the interaction Cn1 (w) Pw or Pn1 (w) Cw by 1 leads to an increase of casualties by Pr (d1 ) = 0:155 which implies a ratio of Pr (d0 ) = Pr (d1 )

2:5: This suggests positive transport costs of violence

so that proposition 3 applies. Sorting population into wards would have reduced violence if the neighborhood was (relatively) balanced. Take the example of Falls - a ward in the centre of Belfast with 9,400 Catholics and about 1,000 Protestants. In the neighborhood of Falls lived 32,500 Protestants and 20,900 Catholics. Our model from equation 2 predicts a reduction in violence from replacing Protestants in Falls with Catholics from the neighborhood if @EA1 @SEGG

=

2

(9:4

+ (9:4 (20:9

1)

1)

Pr (d0 )

(11)

Pr (d1 )

32:5)

Pr (d1 ) < 0

or Pr (d0 ) = Pr (d1 ) > 1:25 which is clearly satis…ed. A sorting of Protestants and Catholics in the Falls neighborhood would have lowered violence. We can use our estimates to simulate a policy in which the 1000 Protestants in Falls are replaced by 1000 Catholics from the neighborhood. According to our

21

estimates violence in the ward would fall by about 7 percent.11 This change in violence occurs because within-ward interactions are replaced by across-ward interactions. Violence would also have been reduced in the ward that exchanged population with Falls because more Protestants in this ward imply less attacks from the neighborhood.

5.2

Tectonic Tensions and Barriers

The policy of building walls and fences was one of the ways in which the UK government tried to contain sectarian violence. If the predictions of our model are correct we would expect that the barriers separating di¤erent streets and areas in Northern Ireland (i.e. the so-called peace lines or peace walls) should be built on ward boundaries with high tectonic tensions. In order to test this idea we obtained geo-referenced information on peaceline constructions. We describe our data collection in the appendix. Our dataset contains 118 peace lines, out of which 72 were located between wards and 46 were within wards. Building on the peace line data, we construct a cross-sectional dataset on the level of ward-pairs (i.e. dyads). Each ward pair only appears once in the dataset (to avoid double-counting). Given that all variables at the dyad level are "symmetric" our results do not depend on which dyad direction we omit. We use as dependent variable a dummy of whether a given ward pair ever had at least one peace line erected between these two wards. In 36 out of our 1603 ward-pairs (i.e. in about 2.2% of cases) this variable takes a value of one. As before, our main independent variable is our measure of tectonic tensions. Only here we calculate it at the dyad level. In other words, the tectonic tensions between two wards w, v is T T (w; v) = Pv

Cw + Cv

Pw :

Note that for the construction of all variables the numbers of Catholics and Protestants are expressed in thousands. If, for example, 1000 Catholics live in ward v and 1000 Protestants live in ward w then tectonic tensions take a value of 1.

11

The two levels of violence are 9:4 2 0:386 + 326:4 0:155 = 58 and 348:4 0:155 = 54:

22

We include ward …xed e¤ects (which we are able to do, as a given ward has several neighbors and hence appears in several dyads), and in some of the speci…cations we control for the total number of Catholics in the dyad, for the total number of Protestants in the dyad and for withinward interactions Pw

Cw ; Pv

Cv .

The results are displayed in Table 4. Column 1 runs the benchmark regression with only tectonic tension as independent variables. Tectonic tension is a strong, positive predictor of the construction of a peace line between two wards. It is signi…cant at the 1% level and quantitatively important. Adding a standard deviation of tectonic tensions more than doubles the baseline risk of peace line construction. In column 2 we control for the total number of Catholics and of Protestants in both wards of the dyads. The coe¢ cient of our variable of interest is unchanged. In column 3 we show that the result is also unchanged when controlling for religious tension within the ward boundaries. Columns 4-6 are the mirror image of columns 1-3, but restricting the analysis to the city of Belfast, where most of the peace lines are located. The results are very similar, with the coe¢ cient on our measure of tectonic tensions being even a bit larger and statistically signi…cant. This highlights the fact that understanding the con‡ict in Northern Ireland at this level of disaggregation would have been impossible without taking into account interactions across space. Highly costly measures were taken to prevent movement and violence across wards which di¤ered in religious composition. This suggests that the government believed violence would decrease with an increase of transport costs. We now turn towards the crucial role that transport costs play for our empirical strategy.

5.3

The Role of Transport Costs

Transport costs play a crucial role in our model. Interactions across space can only be identi…ed if transport cost of violence are not negligible. In section (5.1) we estimated the simplest possible function of transport costs to show that interactions across space matter. We now relax the assumption that attacks are only possible in the direct neighborhood. In Table 4, column (1) we reproduce our main speci…cation from Table 2, column (3). Table

23

4, column (2) adds the tectonic tension with wards that are not a direct neighbor to ward w but neighbors to its neighbors, n1 (w). Formally, this tectonic tension variable is given by T T (2) = Pn2 (w)

Cw + Cn2 (w)

Pw :

(12)

The coe¢ cient on this measure is positive but not signi…cant. The coe¢ cient on T T (1) falls slightly. In column (3) we add tectonic tensions for an additional neighborhood. The coe¢ cient on T T (2) now changes sign but remains insigni…cant. The coe¢ cient on T T (3) is not signi…cant either. Tectonic tensions in the immediate neighborhood, T T (1), still predict violence. The inclusion of just one layer of neighborhoods, n(1), in the previous sections seems justi…ed. In column (4) we provide an additional view on transport costs by aggregating wards to electoral districts (with about 5 wards forming one district). If we run the speci…cation from column (1) on these larger units the image is as before - only that now the coe¢ cients on within ward interactions and T T (1) fall dramatically. In particular, the coe¢ cient within district interactions has a similar magnitude as the interaction with the neighborhood n(2) in column (2). Given the larger size of the units this is not unrealistic. Intriguingly, the coe¢ cient on the linear term of Catholics does not fall. This suggests that the fall in the other coe¢ cients is not simply a scale e¤ect. Throughout all speci…cations in Table 2 we found that the coe¢ cient on the number of Catholics is positive and signi…cant. Proposition 4 states that this linear terms should be driven by a group of perpetrators that target Catholics and do not face transports costs. In the case of Northern Ireland this condition was satis…ed to some extent for state forces like the British Army who fought elements within the Catholic community but did not rely on local Protestant support to the same extent as the Loyalist paramilitaries. The hypothesis from proposition 4 is then that the number of Catholics in a ward can predict violence involving state forces. In order to test this hypothesis we split casualties according to whether they involved state actors either on the side of victims or on the side of perpetrators. We then use the speci…cation from Table 2, column (2) for both sub-samples. The results are

24

reported in columns (5) and (6) of Table 3.12 Violence involving the state (in column (6)) correlates closely with the number of Catholics. Interaction terms as predicted by equation (4), in contrast, play only a minor role in the prediction of the violence involving the state. This pattern can be explained by the fact that state forces were deployed directly in Catholic wards. This makes the number of Catholics a direct predictor both of attacks by state forces and attacks on state forces. This changes considerably in column (5) which con…rms the earlier pattern of interactions across space. These empirical …ndings illustrate the potential of our model in identifying the character of con‡ict. If military organizations require local support then our model predicts very di¤erent violence patterns than if these groups can deploy without local support. In cases of weak states, for example, violence with state involvement will behave much more like sectarian violence in the Northern Irish case.

5.4

Analysis on the Cell Level

Up until now we have focused on endogenously determined units - wards and districts. This is convenient as size in these units is chosen so as to make them more homogenous in population. Wards in urban areas are signi…cantly smaller than in rural areas. The main advantage of this bending of the geography is that we can match census data and violence data to an astonishing degree of precision in the cities. However, the ward level view induces an endogenous structure to the model which implies di¤erent distances in our neighborhood de…nition. This is only justi…able if the transport costs of violence are higher in densely populated areas. To alleviate the concerns from this way of looking at the data we now test our ideas of tectonic tensions with grid-level census data provided by NISRA. Here the units have the size of 1 km2 regardless of the underlying population density. This gives less densely populated areas a much larger weight and generally leads to highly heterogenous units in terms of population. We de…ne

12

More results with a …ner deconstruction of violence is discussed in the appendix and shown in Table A1.

25

a neighborhood n(1) as the 8 cells surrounding the cell.13 Table 5 reproduces the speci…cations from Table 2 in columns (1) to (3). Columns (1) and (2) con…rm the role of tectonic tensions and intra-ward interactions in explaining violence. The coe¢ cient on tectonic tensions indicates that a change of one standard deviation in this variable implies about 0.3 casualties (about 1/3 of a standard deviation) more in the Troubles. This does not change when we add both tectonic tensions and intra-ward interactions in the same regression. The coe¢ cient on tectonic tensions barely changes. However, we now …nd no evidence that intra-cell interactions lead to more violence. One way to explain this surprising result is that interactions within and across wards are highly correlated now and so it is di¢ cult to tell them apart. Another interpretation is that there is a non-monotonic relationship between transport costs in terms of distance and the intensity of interactions. In the models used to describe violence by insurgents this is actually not atypical.14 In this view there could be targets that are "too close to home". Columns (4) and (5) recreate the …ndings from Table 4. They show that, again, the number of Catholics is a direct predictor of violence involving state forces. Tectonic tensions are relatively more important in the prediction of sectarian violence in column (4).

6

Relevance for Other Empirical Work

There is an increasing body of cross-country studies of civil wars that focus on nationwide indicators of ethnic polarization or fractionalization (cf. Fearon and Laitin, 2003; Collier and Hoe- er, 2004; Montalvo and Reynal-Querol, 2005; Collier and Rohner, 2008; Collier, Hoe- er, and Rohner, 2009; Esteban, Mayoral and Ray, 2012). The emphasis of this literature is solely on nationwide ethnic diversity, hence neglecting all information on local ethnic diversity. As shown in our theory the latter is important: For similar nationwide ethnic polarization scores, a country

13 14

We experimented with this de…nition and results are robust to this. See, for example, the algorithms described in Shakarian and Subrahmanian (2011).

26

with two or three large ethnic groups that are geographically separated such as Switzerland, Belgium or Canada has lower local ethnic tensions than places like Bosnia, Rwanda or Guatemala where ethnic groups inhabit the same geographical areas. According to our theory, a given level of motivation at the aggregate level will lead to higher levels of violence with lower segregation. This suggests a complementarity between measures of polarization and segregation at the country level. In recent years there has also be a boom of articles studying civil war with the help of georeferenced, disaggregated data, such as for example Buhaug, Cederman, and Rod (2008), Berman, Shapiro, and Felter (2011), La Ferrara and Harari (2012), Rohner, Thoenig and Zilibotti (2013), Novta (2013), and Berman and Couttenier (2013). As shown by our theory, running regressions that explain local violence with only the characteristics of a given cell or district, will be misspeci…ed when there is signi…cant violence between these units. Note that using simply existing spatial econometrics tools will not solve this problem, as the extent of such between-cell or between-district killings will depend on the interaction between relative population characteristics of the cells or districts involved. To capture the full e¤ect of tectonic tensions the regression speci…cations need to rely on an underlying structural theory of con‡ict between groups. The importance of interactions between characteristics across spacial units also seems of importance for other work inside and outside the con‡ict the literature. There is a large number of economic decisions that is a¤ected by the interaction of geographic features (plains next to mountains, for example). In the move towards more and more disaggregated data these interactions should receive close attention.

7

Conclusion

Most theoretical work on con‡ict models the motivation for con‡ict in detail but does not make predictions on the spatial distribution of violence on the local level. Our theory takes the motivation for violence as given but models the link to local characteristics in more detail. In this way our framework aims to provide a spatial prediction for current theories of con‡ict. Our framework allows for the identi…cation of "transport costs" of violence. If these are not

27

negligible our theory predicts the rise of tectonic tensions - violence appears where larger areas of opposed groups meet. We have also shown that local cell- or district-level analysis can su¤er from high levels of omitted variable bias if interactions across space are not taken into account. We derive two main predictions regarding the spatial distribution of violence in Northern Ireland. First, areas next to tectonic tensions are relevant because the political support in the civilian population creates the base from which attacks are launched. Second, areas in which only the Protestant majority live remain relatively peaceful because violence of state agents is targeted at areas who support insurgents. Our …ndings here relate to Besley and Persson (2011) who show that the control of the state by a group leads to asymmetries in the extent of violence. Our ideas also complement existing work that links violence to the degree of military control within geographic units.15 We show that what de…nes military in‡uence depends on the type of actor. Insurgent and sectarian violence originates in ethnically de…ned areas of control. The ability to project this violence capacity decays with distance. This is what generates the need for interaction terms across space. In other words, the ability to transport violence is a crucial determinant for how violence correlates with local characteristics. Several avenues seem promising for future research: First, a model of endogenous local motivation and transport costs seems a logical next step. Such a model would allow us to trace both local and aggregate con‡icts into violence patterns. Second, we aim to produce measures of segregation and tectonic boundaries for a variety of countries and to perform a cross-country analysis of the e¤ect of nationwide and local ethnic diversity on con‡ict. Third, we warmly encourage studies that apply the current framework to the analysis of other cases than Northern Ireland and to the study of other phenomena than con‡ict where spatial heterogeneity of intensity and local interactions play an important role. Migration, for example, is most attractive where rich areas are close to poor areas. Other examples are research questions in regional science such as the study of urbanization patterns and local economic activity, topics

15

See, for example, Kalyvas (2006) and Berman et al (2011).

28

in electoral politics, such as the study of local campaigning in national elections, or public health policies such as anti-AIDS campaigns.

References [1] Acemoglu, Daron, and Alexander Wolitzky, 2012, "Cycles of Distrust: An Economic Model", mimeo, MIT. [2] BBC,

2009,

"Forty

years

of

peace

lines",

1

July

2009,

http://news.bbc.co.uk/2/hi/uk_news/northern_ireland/8121228.stm [3] Belfast Interface Project, 2012, "Belfast Interfaces: Security Barriers and Defensive Use of Space", report, Belfast Interface Project. [4] Berman, Eli, Jacob Shapiro, and Joseph Felter, 2011, "Can Hearts and Minds Be Bought? The Economics of Counterinsurgency in Iraq", Journal of Political Economy 119: 766-819. [5] Berman, Nicolas, and Mathieu Couttenier, 2013, "External shocks, internal shots: the geography of civil con‡icts", mimeo, Graduate Institute of International and Development Studies, Geneva, and University of Lausanne. [6] Besley, Timothy, and Hannes Mueller, 2012, “Estimating the Peace Dividend: The Impact of Violence on House Prices in Northern Ireland”, American Economic Review 102: 810-833. [7] Besley, Timothy, and Torsten Persson, 2010, "State Capacity, Con‡ict and Development", Econometrica 78: 1-34. [8] Besley, Timothy, and Torsten Persson, 2011, "The Logic of Political Violence", Quarterly Journal of Economics 126: 1411-1445. [9] Bhavnani, Ravi, Dan Miodownik, and Hyun Jin Choi, 2011, "Three Two Tango: Territorial Control and Selective Violence in Israel, the West Bank, and Gaza", Journal of Con‡ict Resolution 55(1): 133-158.

29

[10] Buhaug, Halvard, Lars-Erik Cederman, and Jan Ketil Rod, 2008, "Disaggregating EthnoNationalist Civil Wars: A Dyadic Test of Exclusion Theory", International Organization 62: 531-551. [11] Bruckner, Markus and Antonio Ciccone, 2010, "International Commodity Prices, Growth and the Outbreak of Civil War in Sub-Saharan Africa", The Economic Journal, 120: 519– 534. [12] Caselli, Francesco, and Wilbur John Coleman II, 2013, "On the Theory of Ethnic Con‡ict”, Journal of the European Economic Association 11: 161-192. [13] Cederman, Lars Erik, and Luc Girardin, 2007, "Beyond Fractionalization: Mapping Ethnicity onto Nationalist Insurgencies", American Political Science Review 101: 173-85. [14] Collier, Paul and Anke Hoe- er, 2004, "Greed and grievance in civil war", Oxford Economic Papers 56: 563-95. [15] Collier, Paul, Anke Hoe- er, and Dominic Rohner, 2009, "Beyond Greed and Grievance: Feasibility and Civil War", Oxford Economic Papers 61: 1-27. [16] Collier, Paul, and Dominic Rohner, 2008, "Democracy, Development, and Con‡ict", Journal of the European Economic Association 6: 531-40. [17] Conley, T., 1999, "GMM Estimation with Cross Sectional Dependence," Journal of Econometrics 92, 1-45. [18] Dillon, Martin, 1999, "The Dirty War: Convert Strategies and Tactics Used in Political Con‡icts", New York: Routledge Chapman & Hall. [19] Dube, Oendrila, and Juan Vargas, 2013, "Commodity Price Shocks and Civil Con‡ict: Evidence from Colombia", forthcoming in Review of Economics Studies. [20] Esteban, Joan, Laura Mayoral, and Debraj Ray, 2012, "Ethnicity and Con‡ict: An Empirical Investigation", American Economic Review 102: 1310-1342.

30

[21] Esteban, Joan, Massimo Morelli, and Dominic Rohner, 2012, "Strategic Mass Killings", mimeo, IAE, Columbia University and University of Lausanne. [22] Esteban, Joan, and Debraj Ray, 2008, "On the Salience of Ethnic Con‡ict", American Economic Review 98: 2185-202. [23] Esteban, Joan, and Debraj Ray, 2011, "Linking Con‡ict to Inequality and Polarization", American Economic Review 101: 1345-374. [24] Fearon, James, and David Laitin, 2003, "Ethnicity, Insurgency, and Civil War", American Political Science Review 97: 75–90. [25] Jaeger, David, and Daniele Paserman, 2008, "The Cycle of Violence? An Empirical Analysis of Fatalities in the Palestinian-Israeli Con‡ict", American Economic Review 98: 1591-1604. [26] Jarman, Neil, 2005, "Mapping Interface Barriers", report, Institute for Con‡ict Research, Belfast. [27] Kalyvas, Stathis, 2006, The Logic of Violence in Civil War, Cambridge UK: Cambridge University Press. [28] Kocher, Matthew Adam, Thomas Pepinsky, and Stathis Kalyvas, 2011, "Aerial Bombing and Counterinsurgency in the Vietnam War", American Journal of Political Science 55: 201-218. [29] La Ferrara, Elia, and Maria‡avia Harari, 2012, "Con‡ict, Climate and Cells: A Disaggregated Analysis", IGIER Working Paper n. 461. [30] Lyall, Jason, 2010, "Are Coethnics More E¤ective Counterinsurgents? Evidence from the Second Chechen War", American Political Science Review 104: 1-20 [31] Michalopoulos, Stelios, and Elias Papaioannou, 2011, "The Long-Run E¤ects of the Scramble for Africa", NBER Working Papers 17620. [32] Montalvo, José, and Marta Reynal-Querol, 2005, "Ethnic Polarization, Potential Con‡ict, and Civil Wars", American Economic Review 95: 796-816. 31

[33] Morelli, Massimo, and Dominic Rohner, 2013, "Resource Concentration and Civil Wars", mimeo, Columbia University and University of Lausanne. [34] NISRA,

2006,

Shape…le

with

location

of

peace

lines,

http://www.nisra.gov.uk/geography/Peacelines.htm [35] Novta, Natalija, 2013, "Ethnic Diversity and the Spread of Civil War", mimeo, NYU. [36] Plümper, Thomas, and Eric Neumayer. 2010. Model Speci…cation in the Analysis of Spatial Dependence, European Journal of Political Research, 49 (3), pp. 418-442. [37] Rohner, Dominic, 2011, "Reputation, Group Structure and Social Tensions", Journal of Development Economics 96: 188–199. [38] Rohner, Dominic, Mathias Thoenig, and Fabrizio Zilibotti, 2013, "War Signals: A Theory of Trade, Trust and Con‡ict", forthcoming Review of Economic Studies. [39] Rohner, Dominic, Mathias Thoenig, and Fabrizio Zilibotti, 2012, "Seeds of Distrust: Con‡ict in Uganda", Journal of Economic Growth, 2013, 18: 217-252. [40] Schutte, Sebastian, and Nils Weidmann, 2011, "Di¤usion patterns of violence in civil wars", Political Geography 30: 143-152. [41] Shakarian, Paulo, and V.S. Subrahmanian, 2011, "Geospatial Abduction: Principles and Practice", Springer. [42] Sutton, Malcolm, 1994, "Bear in Mind these Dead ... An Index of Deaths from the Con‡ict in Ireland 1969-1993", Belfast: Beyond the Pale Publications. [43] Townsley, Michael, Shane Johnson and Jerry Ratcli¤e, 2008, "Space Time Dynamics of Insurgent Activity in Iraq", Security Journal 21: 139-146.

A

Proof of Propositions

The result in equation (7) can be derived from equation (4). An increase in polarization is equivalent to an simultaneous increase in Cw and decrease in Pw . Formally the e¤ect of this can 32

be derived by setting Pw = Nw EAw = (2 +

Cw so that (Nw

D X

Cw )

Pni (w)

Cw )

Pr (d0 )

Cw + Cni (w)

(Nw

Cw )

Pr (di )

i=1

and deriving @EAw =@Cw . The e¤ect of segregation can be studied analogously with equation (4) by regarding segregation as four simultaneous changes: increasing Cw , lowering Pw , increasing Pv and lowering Cv .

B

Data on Peacelines

First, we collected data on the location of the peace lines. For this we drew on various lists of peace lines containing geographical information (Jarman, 2005; BBC, 2009; Belfast Interface Project, 2012), on the geo-referenced map of peace lines from NISRA (2006) and on correspondence with the Department of Justice of Northern Ireland, which provided us with additional information in response to our freedom of information request DOJ FOI 12/136. Combining all this sources and using a geo-referenced map of all wards of Northern Ireland, we have been able to put together a novel dataset on the location of peace lines. Peace lines running parallel to borders between two wards and lying either directly on the ward border or in-between the ward border and the nearest street are counted as peace lines separating two wards. Peace lines located in only one ward and not meeting the above criterion are counted as within-ward peace lines. We have not encountered problematic cases that could not be associated to neither of the two categories above (i.e. there have not been peace lines running perpendicular to ward borders and crossing them etc). Our dataset contains 118 peace lines, out of which 72 are located between wards and only 46 are within wards. This suggests that indeed wards are a politically salient unit in the Northern Ireland con‡ict.

33

C

Attacks by Perpetrators and Targets

Table A1) analyses the four violence interactions empirically. The table shows casualties of Catholics (civilians and paramilitaries) caused by Loyalist paramilitaries and state forces, casualties of Protestants caused by Republican paramilitaries, and casualties of state forces caused by Republican paramilitaries. Tectonic tensions are an important explanatory variable in all of these. However, violence against and from state forces is also linear in the number of Catholics. This shows that state forces did, at least to some degree, act without facing transport costs of violence.

34

Table 1: Summary Statistics

WARD LEVEL catholics protestants catholics * protestants tectonic tension n(1) tectonic tension n(2) catholics * protestants in n(1) catholics in n(1) * protestants casulties casulties with state involvement casulties without state involvement

number of units 582 582 582 582 582 582 582 582 582 582

mean 0.67 1.14 0.72 10.97 26.15 5.48 5.48 2.71 1.03 1.61

standard deviation 0.97 1.26 1.72 29.85 63.40 19.34 18.48 7.59 3.63 4.88

min 0 0 0 0 0 0 0 0 0 0

max 9.40 9.76 21.93 325.34 808.41 305.46 232.96 72.00 37.00 43.00

CELL LEVEL catholics protestants catholics * protestants tectonic tension n(1) tectonic tension n(1+2) catholics * protestants in n(1) catholics in n(1) * protestants casulties

number of units 3571 3571 3571 3571 3571 3571 3571 3571

mean 0.11 0.22 0.17 2.24 5.69 1.12 1.12 0.77

standard deviation 0.45 0.78 1.62 20.87 50.45 11.18 12.11 4.94

min 0 0 0 0 0 0 0 0

max 7.97 14.68 55.92 594.62 1215.66 296.84 425.09 134.00

DYAD LEVEL peaceline tectonic in dyad catholics protestants

number of units 3206 3206 3206 3206

mean 0.02 2.12 1.41 2.56

standard deviation 0.15 6.08 1.70 2.32

min 0 0 0 0

max 1.00 8.01 15.91 17.60

Table 2: Main Results (1)

(2) (3) (4) Dependant variable: total number of casualties

(5)

2.080***

0.771*

0.784*

0.805*

(0.373)

(0.416)

(0.441)

(0.419)

0.182***

0.155***

0.154***

(0.0248)

(0.0274)

(0.0327)

VARIABLES protestants * catholics

tectonic tension in n(1)

protestants * catholics in n(1)

0.174*** (0.0267)

catholics * protestants in n(1)

0.134*** (0.0450)

catholics

protestants

2.738***

2.027***

1.542**

1.399*

1.938**

(0.528)

(0.646)

(0.685)

(0.726)

(0.819)

0.0856

0.0156

-0.376

-0.509

-0.625

(0.431)

(0.513)

(0.608)

(0.688)

(0.678)

-0.0545

0.0539

0.0425

-0.0207

(0.148)

(0.147)

(0.179)

(0.155)

-0.0574

-0.0437

-0.0686

-0.00406

(0.0851)

(0.0929)

(0.0999)

(0.105)

no

no

yes

no

catholics in n(1)

protestants in n(1)

discrict fixed effects

no

Observations 582 582 582 582 582 R-squared 0.622 0.735 0.744 0.750 0.745 Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. "Total number of casualties" is the sum of all casualties in the troubles between 1971 and 1979. Catholics and Protestants are measured in 1000s. The coefficients on interaction terms are therefore casualties per million. "Tectonic tension in n(1)" is caclulated as in equation 6. "Catholics in n(1)" is the total number of Catholics in wards adjacent to the ward.

Table 3: Tectonic Boundaries and the Construction of Barriers

(1)

tectonic tension in dyad

catholics in dyad

protestants in dyad

within-ward interactions

(2) (3) (4) (5) Dependent variable: creation of peace lines between the two wards of a dyad

(6)

0.0144***

0.0144***

0.0143***

0.0152***

0.0152***

0.0151***

(0.0031)

(0.0031)

(0.0030)

(0.0034)

(0.0034)

(0.0034)

-0.0538***

-0.0692***

-0.1556***

-0.1646***

(0.0174)

(0.0187)

(0.0198)

(0.0198)

-0.0039

-0.0148

-0.1063***

-0.1114***

(0.0203)

(0.0217)

(0.0250)

(0.0252)

0.0175

0.0153

(0.0114)

(0.0123)

Ward Fixed Effects Yes Yes Yes Yes Yes Yes Sample All All All Belfast Belfast Belfast Observations 1603 1603 1603 176 176 176 R-squared 0.729 0.729 0.733 0.745 0.745 0.749 Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Catholics and Protestants are measured in 1000s. The coefficients on interaction terms are therefore casualties per million. "Tectonic tension in dyad" is the number of Catholics in ward v multiplied by Protestants in w plus the number of Protestants in v multiplied by the number of Catholics in ward w. "Within-ward interactions" is the number of Catholics in A multiplied by the number of Protestants in A plus the number of Catholics in B multiplied by the number of Protestants in B.

Table 4: The Role of Transport Costs

(1)

tectonic tension in n(1)

0.771* (0.416) 0.155*** (0.0274)

0.801* (0.434) 0.113*** (0.0432) 0.0248 (0.0251)

1.542** (0.685) -0.376 (0.608) yes yes

1.118* (0.584) -0.433 (0.587) yes yes

tectonic tension in n(2) tectonic tension in n(3) catholics protestants catholics in n protestants in n

(3)

(4)

(5)

(6)

casualties without state involvement

casualties with state involvement

0.0301*** (0.00118) 0.00683** (0.00273)

0.729** (0.306) 0.100*** (0.0224)

0.00672 (0.277) 0.0503*** (0.0175)

2.363*** (0.236) -0.150 (0.0960) yes yes

-0.330 (0.453) 0.143 (0.428) yes yes

1.800*** (0.426) -0.655*** (0.236) yes yes

total number of casualties

VARIABLES protestants * catholics

(2)

0.588 (0.448) 0.123*** (0.0397) -0.00214 (0.0326) 0.0242 (0.0211) 0.889 (0.551) -0.329 (0.569) yes yes

Observations 582 582 582 118 582 582 R-squared 0.744 0.751 0.756 0.994 0.731 0.563 Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. "Total number of casualties" is the sum of all casualties in the troubles between 1971 and 1979. "Without state involvement" are casualties in which state forces were involved neither as victims or perpetrators. Catholics and Protestants are measured in 1000s. The coefficients on interaction terms are therefore casualties per million. "Tectonic tension in n(1)" is caclulated as in equation 6. "n(2)" are wards that are adjacent to wards in the direct neighbourhood but which are not direct neighbours themselves. "n(3)" are, analogously, the neighbour's neighbour's neighbours. "Catholics in n" and "patholics in n" are a set of controls for each of the neighborhoods, analogous to table 2. Column (4) is on the district level.

Table 5: Regressions at the Cell Level (1) VARIABLES protestants * catholics

protestants catholics in n(1) protestants in n(1)

(3)

Dependant variable: total number of casualties

(5) casualties without state involvement

(6) casualties with state involvement

1.167***

-0.0361

0.221

-0.367

(0.186)

(0.675)

(0.412)

(0.290)

0.115***

0.118*

0.0683*

0.0509*

(0.0232)

(0.0620)

(0.0368)

(0.0271)

4.987***

3.364***

3.380***

0.786*

2.653***

(0.905)

(0.804)

(0.845)

(0.468)

(0.484)

-0.0239

-1.452**

-1.432**

-0.706*

-0.706***

(0.257)

(0.739)

(0.606)

(0.372)

(0.255)

0.153

0.149

0.108

0.0415

(0.138)

(0.146)

(0.0851)

(0.0859)

0.254**

0.252**

0.182**

0.0634

(0.129)

(0.117)

(0.0732)

(0.0505)

tectonic tension in n(1) catholics

(2)

Observations 3,571 3,571 3,571 3,571 3,571 R-squared 0.562 0.654 0.654 0.675 0.494 Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. "Total number of casualties" is the sum of all casualties in the troubles between 1971 and 1979. Catholics and Protestants are measured in 1000s. The coefficients on interaction terms are therefore casualties per million. "Tectonic tension in n(1)" is caclulated as in equation 6. "Catholics in n(1)" is the total number of Catholics in the 8 cells adjacent to the cell. "N(1+2)" includes indirect neighbors.

Table A1 - Violence Split by Perpetrators and Targets

VARIABLES protestants * catholics tectonic tension in n(1) catholics protestants catholics in n(1) protestants in n(1)

Observations

(1)

(2)

(3)

(4)

protestants killed by republican paramilitaries

catholics killed by loyalist paramilitaries

0.272** (0.133) 0.0176** (0.00704) -0.172 (0.152) 0.0867 (0.247) 0.0458 (0.0312) -0.0240 (0.0365)

0.0878 (0.211) 0.0570*** (0.0182) -0.268 (0.337) -0.00369 (0.174) 0.0442 (0.0672) 0.00639 (0.0316)

0.00759 (0.132) 0.0280*** (0.00792) 1.095*** (0.214) -0.378*** (0.126) -0.0246 (0.0518) 0.0218 (0.0190)

-0.000875 (0.168) 0.0223* (0.0125) 0.705** (0.275) -0.277** (0.128) -0.0259 (0.0731) 0.00842 (0.0221)

582

582

582

582

state forces killed by catholics killed by state republican paramilitaries forces

R-squared 0.453 0.605 0.501 0.471 Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. "Total number of casualties" is the sum of all casualties in the troubles between 1971 and 1979. Catholics and Protestants are measured in 1000s. The coefficients on interaction terms are therefore casualties per million. "Tectonic tension in n(1)" is caclulated as in equation 6. "Catholics in n(1)" is the total number of Catholics in wards adjacent to the ward.

Figure 1: Two Violence Patterns

A A A A A A A A A B B B peaceful violent

A A A A A A A A A B B B

0

20

casu ualties 40 4

60

8 80

Figure 2: Casualties and Within-Ward Interactions

0

5

10 15 catholics x protestants

20

0

20

alties casua 40

60

8 80

Figure 3: Casualties and Polarization/Factionalization

0

.2

.4

.6 polarization

.8

1

0

20

casu ualties 40 4

60

8 80

Figure 4: Casualties and Tectonic Tensions

0

50

100

150 200 tectonic tensions

250

300

350