The geo-temporal evolution of violence in civil conflicts: A micro analysis of conflict diffusion on a new event data set

Abstract

Existing works on diffusion fail to account for the incapacitating effects conflict events may have on the operational capability of the combatant sides and how these effects may determine the evolution of a conflict. I hypothesize that it is those events with losses on the state side that are likely to be associated with geo-temporal spillovers, whereas events with insurgency losses are less likely to be associated with future mayhem in their vicinity. To test my arguments, I first introduce a new, comprehensive and detailed event dataset on the long-running civil conflict in Turkey. The Turkish State–PKK Conflict Event Database (TPCONED) includes the exact date and county-level location for the fatal events of the armed conflict between the Turkish state and the rebel organization PKK since its very beginning in 1984 with detailed information on combatant casualties. I then employ a split population bi-probit model which allows me to comprehensively depict the geotemporal evolution of the conflict by acknowledging, estimating and accounting for the variation in the underlying conflict proneness across locations as a latent variable that shapes the diffusion of events. The results of the statistical analyses offer support for my hypotheses and reveal that how events evolve over space and time is conditioned by the damages suffered by the combatant sides. I demonstrate the robustness of these results on a matched sample I obtain by employing the Coarsened Exact Matching (CEM) on the data.

Keywords

civil conflict conflict event dataset spatial analysis split population model

Introduction

How do civil conflicts evolve over time and space? Are they chaotic episodes of armed violence in fragile states or is there a predictable pattern to the diffusion of events that we can decipher? Given the prevalence and destructiveness of civil conflicts it is imperative that we answer these questions, because if we can figure out the spatial and temporal dynamics, we can then hope to devise pre-emptive actions and policies to prevent or at least contain these humanitarian disasters.

The literature on conflict diffusion is dominated by works on international contagion and focuses on the factors that facilitate the spread of civil conflicts between countries (Buhaug & Gleditsch, 2008; Ward & Gleditsch, 2010; Cederman, Girardin & Gleditsch, 2009; Cederman et al., 2013; Lake & Rothchild, 1998; Lane, 2016; Salehyan & Gleditsch, 2007; Weidmann, 2015). Compared to this rich literature on cross-country diffusion, the literature on diffusion of events within conflicts is still in its early stages. Nevertheless, existing studies have already firmly established that conflict events exhibit spatial and temporal interdependencies (Townsley, Johnson & Ratcliffe, 2008; Hegre, Ostby, & Raleigh, 2009; Lyall, 2009; Raleigh et al., 2010; Weidmann & Ward, 2010; Schutte & Weidmann, 2011; O’Loughlin & Witmer, 2012). This strong result has then led scholars to explore the mechanisms that explain this spatio-temporal clustering of violence. Recent works have theorized about and provided empirical evidence for the role played by the relative capabilities of combatants (Holtermann, 2016; Beardsley, Gleditsch & Lo, 2015), accessibility of locations via road networks (Zhukov, 2012), environmental conditions (Carter & Veale, 2015), and retaliatory motives (Braithwaite & Johnson, 2012; Linke, Witmer & O’Loughlin, 2012) in determining the time and location of conflict events. Surprisingly however, none of these works pays any attention to the incapacitating effects these events may have on the warring factions and how these effects may determine the evolution of the conflict. Even those studies that emphasize the tit-for-tat nature of conflicts seem to have forgotten what Machiavelli reminded Lorenzo Di Pierro De Medici 400 years ago: ‘(men) can avenge themselves of lighter injuries, of more serious ones they cannot’ (Machiavelli, n.d.: 7).

In this study, I argue that how an insurgency spreads within a country is associated with how conflict events affect the operational capabilities of the warring sides, or in Machiavelli’s words, with the amount of injury each side suffers as a result of each event. Employing a split population bi-probit model I test my arguments on a new event dataset I introduce on the long running civil conflict in Turkey. The results offer support for my arguments. Confirming previous diffusion studies, I find that once a civil conflict starts, geotemporal interdependencies play an important role in determining how events evolve over time and space. Significantly contributing to our understanding of these interdependencies, and in line with my theoretical arguments, my findings indicate that the geotemporal evolution of the conflict is conditioned by the damages suffered by the combatant sides, and that it is events with losses on the state side that are likely to be associated with geotemporal spillovers, whereas events with insurgency losses are less likely to be followed by further violence in the neighborhood.

In the next section, I discuss diffusion patterns in civil conflicts and introduce my theoretical arguments. Then I introduce my case, my data, my statistical model, the results and robustness checks in the following sections.

Theory and literature

The conflict diffusion literature has two main branches. The first branch focuses on international diffusion and analyzes the mechanisms behind transnational spillovers of violence. The works under this branch are in fact cross-country studies of conflict onset that assess a country’s risk of experiencing a civil conflict based on its local characteristics and its interactions with the outside world. Among the interactions identified in this literature as transborder carriers of conflict risk are refugee flows (Salehyan & Gleditsch, 2007; Braithwaite, 2010; Rüegger, 2018), communication networks (Weidmann, 2015; Beiser, 2016), circulation of arms and combatants (Lane, 2016; Bara, 2018; Braithwaite & Chu, 2018), learning and strategic emulation (Buhaug & Gleditsch, 2008; Maves & Braithwaite, 2013; Hill, Rothchild & Cameron, 1998; Forsberg, 2008; Ayres & Saideman, 2000), and external sponsorship of insurgencies (Gleditsch, Salehyan, & Schultz, 2008; Cederman, Girardin & Gleditsch, 2009; Schultz, 2010).

The findings of this impressive literature provide theoretical and methodological insights for the second branch which takes a micro-level approach to analyze the within-conflict diffusion of violence. This literature is still in its early stages; however, existing works have already firmly established that conflict events exhibit a strong spatial and temporal interdependency (Townsley, Johnson & Ratcliffe, 2008; Hegre, Ostby & Raleigh, 2009; Lyall, 2009; Raleigh et al., 2010; Weidmann & Ward, 2010; Schutte & Weidmann, 2011; O’Loughlin & Witmer, 2012). Recent works have investigated the mechanisms that explain this spatio-temporal clustering and provided empirical evidence for the role played by factors such as the accessibility of locations via road networks (Zhukov, 2012), environmental conditions (Carter & Veale, 2015), internal displacements (Bohnet, Cottier & Hug, 2018), and state coercion (Duffy Toft & Zhukov, 2012) in determining the time and location of conflict events. Interestingly however, attention is yet to be paid to the incapacitating effects conflict events may have on the warring sides, and how these effects may influence the likelihood of geotemporal spillovers. This, I argue, is a serious shortcoming. Civil conflicts involve strategic actors making strategic decisions on whether, when and how to act. It is, thus, not possible to construct an accurate understanding of how conflict events evolve over time and space without taking into account those factors that shape these decisions. Operational capability is one of the most important of such factors. Operational capability refers to the capacity of conflict actors to carry out successful combat operations against the adversary on the battlefield, and is mainly determined by resources available to them and their ability to successfully use those resources (Tellis et al., 2000). Note that conflict events can inflict damages on both of those determinants of operational capability and as such can have incapacitating effects on conflict actors.

The argument that the evolution of a conflict is associated with how conflict actors are affected by events has in fact already been raised by those works that emphasize the tit-for-tat nature of conflicts and the retaliatory motives of conflict actors (Jaeger & Paserman, 2008; Lyall, 2009; Haushofer, Biletzki & Kanwisher, 2010; Linke, Witmer & O’Loughlin, 2012; Kocher, Pepinsky & Kalyvas, 2011; O’Loughlin & Witmer, 2012; Braithwaite & Johnson, 2012; Schutte & Donnay, 2014). However, even those works implicitly assume that combatants will always have the capacity to react against all instances of violence. In this study, I relax this assumption by controlling for combatant casualties in each violent event as a measure of the damage to the operational capability of warring sides.

The size of the military force is an important resource in all armed conflicts, but especially so for insurgents fighting against powerful state adversaries. In fact, the opportunity-cost theories of conflict onset posit the availability of labor as the binding constraint on the production of violence by insurgencies (Grossman, 1991; Mikulascheck & Shapiro, 2018). Most civil conflicts are fought between organized, well-armed, and sizeable state military forces and relatively much smaller and ill-equipped insurgent groups.¹ Balcells & Kalyvas (2014: 1391) use the term ‘irregular conflicts’ to refer to civil conflicts with such power asymmetry. The asymmetry in resources in irregular civil conflicts means that, compared to state forces, casualties are expected to be more costly and debilitating for insurgents since each combatant corresponds to a higher share of their operational capability. The relative difficulty of recruiting and training replacements inflates this cost further. Heavy losses against state forces may also have a deterrent effect on both the insurgents and their civilian support base which may then render recruitment even more difficult and may even lead to defections.

The asymmetry in resources not only renders them more valuable for insurgents but also shapes the way they utilize them. Due to a state’s material advantages, insurgent groups stand a high risk of defeat if they try to fight the state with conventional tactics in consistent theaters of combat. They thus favor mobility and guerrilla warfare. Staying mobile helps them evade attacks. It also gives them an opportunity to compete with the state’s armed forces by varying targets and using the element of surprise to their advantage (Beardsley, Gleditsch & Lo, 2015). McColl’s (1969) influential account of how revolutionary insurgencies evolve also emphasizes the need for mobility.

Note that this type of irregular warfare creates a pattern in which insurgents proactively stage hit-and-run attacks which then drive state forces into reactive counterinsurgency operations.² But, the insurgents’ capacity to sustain such a pattern depends on whether they can hit their targets and run afterwards without getting hit themselves. Casualty counts give us a grim account of their success in doing so. While security force casualties provide a measure of the hit the state side takes, insurgency casualties provide a measure of their (in)ability to run. It follows that with each insurgent casualty this pattern becomes less sustainable, and the likelihood of future hit-and-run attacks in the vicinity goes down. On the other hand, state casualties can be expected to bolster this pattern. It is reasonable to expect damages on state forces to make it easier for insurgents to escape after an attack. Successful attacks may also boost morale among rebels, help them in their efforts to gain public support and find new recruits (Kalyvas, 2006) thereby allowing them to increase the geotemporal scope of their activities.

Casualties are important in determining the actions of state forces as well. The size advantage and the relative ease of recruitment shield the operational capability of state forces against losing soldiers, thus state casualties are less likely to have a dampening effect on conflict activity. On the contrary, in many cases, because losses on the state side carry heavy political costs for state leaders (Kibris, 2011), this can lead them to resort to retaliatory actions that perpetuate violence (Jaeger & Paserman, 2008). Holtermann (2016) argues that state casualties can also be seen as a measure of the relative capacity of rebels and as such they can indicate higher likelihood of future conflict events in the neighborhood. Relatedly, losses in one area may lead to a transfer of state military resources from nearby locations and may leave those areas vulnerable.

If these are valid mechanisms then we should expect events with insurgency casualties to curb the potential for future events in nearby locations and events with state security force casualties to be harbingers of geotemporal spillovers. The following two hypotheses are derived from these expectations:

Hypothesis 1: In irregular civil conflicts, casualties on the state side are positively associated with future conflict events in other locations in the neighborhood.

Hypothesis 2: In irregular civil conflicts, casualties on the insurgency side are negatively associated with future events in other locations in the neighborhood.

I test these hypotheses on a new and detailed event dataset I introduce on the long-running civil conflict in Turkey. The Turkish State–PKK Conflict Event Database (TPCONED) includes the exact date and county-level location for the fatal events of the armed conflict between the Turkish state and the rebel organization PKK in the 1984–2018 period with detailed information on combatant casualties.

Statistical analysis of conflict diffusion offers an empirical challenge because it comes with very specific and demanding data requirements. In order to be able to track the geotemporal path of violence, one needs a comprehensive and complete event dataset with detailed information on the time, location and characteristics of conflict events. Comprehensiveness, which is full geotemporal coverage of the conflict, is important to make sure that any observed association is not specific to a period or a location. Completeness, which is not having any missing observations, is even more important. Missing observations introduce a serious selection bias in any statistical analysis, but they are even more problematic in diffusion studies since with each missing observation another step in the geotemporal progression of events gets lost and the accuracy of the dataset in reflecting the diffusion of violence weakens. TPCONED is a comprehensive and complete event dataset on the armed conflict between the rebel organization PKK and the Turkish state, and as such, it avails the armed conflict in Turkey as a rich case study for understanding the geotemporal dynamics of civil conflicts.

The conflict

Since late 1984, Turkey has been suffering from an insurgency campaign led by the Kurdish separatist guerrilla organization the Kurdistan Workers’ Party (PKK). The organization was first founded with the goal of establishing an independent Kurdish state in southeastern Turkey, though later on in the 1990s, it appeared to have rolled back on its goal to a federational structure that would grant more autonomy to the large Kurdish minority in the country. The armed conflict between the PKK and the Turkish security forces (TSF) has been geographically concentrated in southeastern and eastern regions which have traditionally been inhabited by

Figure 1.

Combatant casualties, 1984–2018

ethnic Kurds. First dismissed by Turkish governments as the acts of a handful of outlaws, this irregular conflict has been going on for more than 35 years.³ Financially, it has cost the country billions of dollars. But more importantly, it claimed the lives of tens of thousands.

Figure 1 depicts the total number of TSF and PKK casualties over time. As can be seen, the 1990s has been the bloodiest period of the conflict. The insurgency announced a unilateral ceasefire after its leader Abdullah Öcalan was captured in 1999 and ceased its attacks in the early 2000s. Unfortunately, peace in the area did not last long and violence flared up again after 2004. The latest ceasefire was announced in March 2013 as part of a peace process which unfortunately broke down in July 2015 only to bring more bloodshed. Figure 2 depicts the geographical distribution of combatant casualties at the province level and reveals their geographical concentration.

The Turkish State–PKK Conflict Event Database – TPCONED

I test my hypotheses with a new, high-resolution event dataset that I introduce in this study on the armed conflict between the rebel organization PKK and the Turkish state. This is one of the longest running civil conflicts in the world, and it plays an important role in the political turmoil of the Middle East, and as such it has a lot to tell us about civil conflicts. However, due to scarcity of reliable data, it has been analyzed by only a handful of empirical studies so far. I release the Turkish State–PKK Conflict Event Database (TPCONED) with the hope of

Figure 2.

Total combatant casualties, 1984–2018

motivating those much-needed studies that will bring this deadly conflict into the spotlight.

TPCONED (https://wrap.warwick.ac.uk/138227) is a dataset on the fatal events of the armed conflict between the rebel organization PKK and the Turkish state that took place on Turkish soil and in which there was at least one combatant casualty. It contains detailed information on 7,063 conflict events with 17,308 PKK casualties and 7,514 state casualties over the course of the conflict in the 1984–2018 period.⁴ For each event, the exact date, location at the county (town) level, location at the province level, number of TSF and PKK casualties, and the data sources are listed. The dataset has been in the making since 2009, and relies on a wide range of local sources which include reports and publications by the Turkish Ministry of Defence, Turkish General Staff, the Directorate General of Press and Information of Turkish Prime Ministry and by various other state offices and local administrations in Turkey; the archives of the Gendarmerie Museum in Ankara; the digital and hard copy archives of five major Turkish daily newspapers (Cumhuriyet, Milliyet, Hürriyet, Sabah, and Zaman); PKK publications; daily and yearly reports by the Turkish Human Rights Association; scholarly works (Tezcür, 2016); and personal contacts and interviews with the families of the casualties. I provide detailed information on these sources in section A1 in the Online appendix.

Note that my sources belong to different political affiliations and views. I cross-checked contentious events across these sources to make sure that the resulting dataset is as comprehensive and accurate as possible. A very important thing to note here is that combatant casualties do not only serve as a binary decision rule for an event to be included in TPCONED but more importantly they provide a solid confirmation of the validity of each event in the dataset.

Turkey has a draft army and a mandatory military service system that requires each Turkish man, when he comes of age, to serve in the army for about a year. This mandatory system is supported by a military and religious culture that glorifies the army and service to the country. When a soldier or a police officer dies in service, he or she is considered a martyr, both legally and culturally. A state funeral is held for each of them in their hometowns. These funerals have always been important public events attended by high-level military and state officials and thousands of locals. The national and the local media cover the story of the fallen soldier or officer, his grief-stricken family, and the funeral ceremony, and by so extensively publicizing the casualties on the state side, they create the most credible way of tracking down the events and the geotemporal evolution of the conflict.

A similar glorification of combatant casualties is also present on the PKK side. They also refer to their casualties as martyrs and commemorate them by special publications. Although very limited compared to the information on the Turkish side, such resources present valuable information on conflict events.

Hence, all the events recorded in TPCONED have been confirmed by information on the identity of the each of the 7,514 TSF casualties involved and of a good portion of the PKK casualties.⁵

To demonstrate the accuracy, precision and comprehensiveness gains that TPCONED accomplishes, I present a systematic and detailed comparison (á la Donnay et al., 2018) of TPCONED and the UCDP-GED database (Sundberg & Melander, 2013), which is one of the most commonly used conflict event databases by researchers, in section A2 in the Online appendix. Note that the GED does not have any data on the conflict for its first five years, and then has only a total of 4,682 conflict events recorded for the 1989–2018 period. Moreover, of the recorded events, only 54%, have location information at the county level, that is, location at the county level is missing in 46% of the observations. With nearly half the observations missing, it becomes quite impossible to track the evolution of the conflict at the county level with the GED data. Unfortunately, the problem with missing observations persists at larger geographical aggregations as well. For 7.5% of the observations, reported location covers almost the entire conflict zone.⁶

Another major problem that diminishes the quality of existing conflict event databases is that most of them rely on international news agencies. Going back to the GED example, nearly all the observations under theTurkish state–PKK conflict are coded from sources like the BBC, Reuters, Agence France Press, etc. Admittedly, these international news agencies themselves rely on local sources for information, but they tend to rely more on state sources or agencies which, in many cases, feed them with biased information. In the Turkish case for example, they mostly rely on the Anadolu (News) Agency which is in fact owned and controlled by the Turkish state. Moreover, they tend to be biased towards more substantial events in their reporting while ignoring smaller ones. They also tend to summarize the information they receive from local sources and leave out certain details, like precise location of the event, which they deem irrelevant for their international readers. Moreover, in many cases, a lot gets lost in translation.

Statistical model and analyses

Simply stated, diffusion refers to the temporal and geographical spread of conflict events. In this study, I define a conflict event as any conflict-related lethal violence that results in the death of at least one combatant. I use casualties as a measure of the size of losses in terms of operational capability of each warring faction in each conflict event.

In most civil conflicts, violence remains geographically clustered, and a great deal of places never experience any events while a small subset of locations accounts for nearly all the violence in a recurring fashion. Civil conflicts are violent struggles about state formation (Weinstein, 2007; Tull, 2004; Pegg, 1998), in other words, insurgencies are would-be states. Note that no such would-be-state-maker takes up arms without a targeted territory. Insurgents may target to take over a country as a whole or some parts of it, either because it is the ethnic homeland, the promised territory, or because it is rich in economic resources, or home to ideologically sympathetic populations, etc. The fact that insurgents target certain territories implies that there is an underlying, base risk distribution over space, and that the geotemporal evolution of the conflict is correlated with this distribution. This implication is consistent with the observed geographical clustering of conflict events. It is also consistent with McColl’s (1969) account of how insurgencies evolve. The targeted base areas become the hubs of conflict events, the areas in their periphery experience relatively fewer incidents as insurgents try to expand, and the rest of the territory remains outside the conflict zone. But if particular territorial units are not likely to experience any conflict events, then their inclusion in the sample may lead the diffusion analysis astray. Braumoeller & Goertz (2002) and Mahoney & Goertz (2004) argue that including irrelevant cases where the outcome of interest is impossible induces erroneous inference. Having said that, trying to exclude them can also be very problematic since it requires distinguishing between the relevant and irrelevant areas when relevancy is unobservable. Any ad hoc identification risks excluding relevant areas or including irrelevant ones. Note that a similar problem had been discussed in the study of international conflicts. There the argument is that some country dyads, because they lack any opportunity to fight, are not relevant for an empirical study on the determinants of international war. To deal with this problem, Clark & Regan (2003) and Xiang (2010) propose split population models in which they estimate relevance as a latent variable. In this study, I follow their proposition and employ a split population bi-probit model to analyze the geotemporal evolution of conflict events.

In a split population model an additional binary choice regression – used as a selection step – is added to a standard statistical model to capture the idea that there are two data-generating processes behind the observed data. In the context of modelling civil conflict diffusion, the selection step is meant to estimate the underlying risk distribution over territorial units and identify the potential base areas. A split population model, in that sense, can be used to bring together the civil conflict onset literature which explains the occurrence of violence in relation to specific conditions existing prior to the conflict, and the diffusion literature which emphasizes the importance of geotemporal interdependencies.⁷

My dependent variable Y_{i, t} is a binary incidence variable that takes on the value 1 if a conflict event took place in location i in time t, and 0 if not, and

Y_{i, t} \sim 0 with probability 1 − p_{i} and

Y_{i, t} {\sim F}_{i, t} with probability p_{i}

where F_{i, t} is a cumulative distribution function for a binary choice model.

Let p_i = G_i where G_i is also a cumulative distribution function for a binary variable. Then, because the zero outcome is generated by both the binary choice models G_i and F_{i, t}, we have

Y_{i, t} = 0 with probability ({1 − G}_{i}) {+ G}_{i} ({1 − F}_{i, t}) and

Y_{i, t} {= 1 with probability G}_{i} F_{i, t}

under the assumption that the two distributions are independent. Here, G_i determines the distribution of underlying conflict risk and F_{i, t} determines the distribution of event risk. Note that the underlying risk is not observable. We only observe whether a location experiences a conflict event at a given time or not.

The independence assumption is not a reasonable one in a civil conflict context. Locations with high underlying risk will most probably have high probability of experiencing conflict events. I therefore adopt a bivariate standard normal distribution to model the two correlated cumulative distribution functions. As a result, the above model becomes

Y_{i, t} = 0 with probability [1 - Φ_{2} {(βX}_{i, t} {,γZ}_{i, t};ρ)]

Y_{i, t} = 1 with probability Φ_{2} {(βX}_{i, t} {,γZ}_{i, t};ρ)

where Φ ₂ is the bivariate standard normal cumulative distribution function and ρ is the correlation coefficient. Z is the vector of covariates that affect the base risk, and X is the vector of covariates that affect the probability of diffusion. This is very similar to the binary choice model Xiang (2010) employs in the context of international war.

The subscripts i and t designate time and space units to be used. In terms of designating space, the literature hosts two approaches. The first approach makes use of administrative divisions (Holtermann, 2016; Weidmann & Ward, 2010; O’Louglin & Witmer, 2012; Zhukov, 2012) to identify distinct spaces. Scholars justify the distinctiveness of administrative units by referring to their socio-economic differences. However, as Schutte & Weidmann (2011: 147) argue ‘administrative boundaries may have little relevance in civil wars, since they can be crossed easily by armed forces’. But this concern is even more valid for the second approach which divides the area of study into grid cells of equal size and treats each grid cell as a distinct space (Schutte & Weidmann, 2011; Raleigh & Hegre, 2009; Townsley, Johnson & Ratcliffe, 2008). Unfortunately, there is not much discussion in the works that employ geographical grids on what makes two adjacent grid cells distinct spaces from the perspective of conflict actors and/or for the sake of conflict dynamics. To deal with the ad hoc nature of this gridding exercise, two alternatives have been offered so far. The more commonly adopted alternative is to repeat the analyses with multiple grid sizes. While this practice may help ease concerns about the robustness of estimated associations, it does not provide any theoretical justification for the relevance of resulting geographical units. Moreover, exact event coordinate information is very hard to come by. Most conflict event databases, including the GED, report the coordinates of the administrative unit in which the event takes place. The second alternative is offered in a recent study by Schutte (2017) where he employs a point process model (PPM) in which, rather than relying on predefined spatial units, suitable spatial windows are heuristically defined from the distribution of events over territory. As Schutte (2017: 454) argues, these windows offer a better alternative to ad hoc spatial grids but are so far limited to cross-sectional analysis as ‘the introduction of a temporal dimension provides additional challenges’.

Schutte & Weidmann (2011) argue that spatially aggregating conflict activity leads to a loss of information, and they favor using as fine a spatial resolution as possible with the data to hand. In this study, I follow up on their advice and estimate my model under the highest spatial resolution possible with my data, namely, the county level. This is quite a high resolution especially considering that I am able to sustain it across all events throughout the whole span of the conflict and that it cannot be attained by any of the other available datasets. Moreover, my socio-economic controls are able to match the geographical resolution of my event data. To make sure that my results are not specific to any time aggregation, I estimate my model under monthly, quarterly, half-yearly and yearly time denominations.

My X vector of diffusion covariates contains control variables for within-conflict dynamics. To control for the geotemporal interdependencies, I include the length of the last peace spell; the percentage of past periods with conflict incidences; one-period lagged state casualties; one-period lagged insurgent casualties; one-period lagged inverse-road-distance-weighted state casualties in other counties; and one-period lagged inverse-road-distance-weighted insurgent casualties in other counties. These last two are the control variables of interest for this study as they are the spatio-temporal lags whose association with the dependent variable will inform us about the extent and nature of geotemporal spillovers. Unlike existing studies, I do not limit geotemporal interdependencies to immediately neighboring administrative units or units within a fixed distance. Instead, my spatio-temporal lags take into account all events of the previous period after assigning them weights according to distance. Accordingly, the one-period lagged inverse-road-distance-weighted insurgent casualties for county i at time t is

\sum_{k \neq i} ({1/d}_{ki}) {(state casualties)}_{k, t-1}

where d_ki is the road distance between county k and county i in kilometers. Similarly,

\sum_{k \neq i} ({1/d}_{ki}) {(insurgency casualties)}_{k, t-1}

is the spatio-temporally lagged insurgency casualties for county i at time t⁸. I use road distances to calculate these spatial lags in order to account for the role of the road networks in facilitating the spread of violence (Zhukov, 2012).

I also control for area, border status and percentage of rural mountainous terrain across counties. Finally, I include season and year dummies to capture the seasonal and yearly variations.

My Z vector of covariates for base risk includes a rich set of controls which depict the socio-economic situation at the start of the armed conflict. I control for percentage of farmers with no land; percentage of farmers with more than 100 acres of land; number of mosques per village; percentage of villages with no drinking water; percentage of villages with no electricity; unemployment rate; literacy rate and level of urbanization across counties as measures of state capacity and reach, and as indicators of economic development and welfare. Economic grievances and low state capacity create fertile environments for insurgencies (Holtermann, 2012). I also control for the percentage of villages whose names were changed by the state as a measure of ethnic discriminatory policies. Ethnic discrimination is identified as an important factor that increases the vulnerability of a country to experience conflict (Metternich, Minhas & Ward, 2017; Cederman et al., 2013; Cederman, Wimmer & Min, 2010). Relatedly, I also control for the ethnic distribution of the population across counties to account for the potential support base of the PKK.

Finally, I control for the percentage of mountainous area, border status, and population across counties (Fearon & Laitin, 2003; Do & Iyer, 2010; Weidmann & Ward, 2010; Zhukov, 2012; Holtermann, 2016; Cederman, Girardin & Gleditsch, 2009; Cederman et al., 2013).

I provide a detailed discussion on the control variables in Section A4 in the Online appendix along with descriptive statistics and a visual representation of their predictive power in terms of the base risk of conflict across counties.

Results

Table I presents the estimated parameters of the bivariate probit model under monthly, quarterly, half-yearly, and yearly time aggregations.

Table I.

Partial observability bivariate probit results

	Unit of observation: County-month (1984–2018) Number of obs.: 266,845	Unit of observation: County-quarter year (1984–2018) Number of obs.: 88,734	Unit of observation: County-half year (1984–2018) Number of obs.: 44,367	Unit of observation: County-year (1984–2018) Number of obs.: 21,862
Within-conflict dynamics
Length of last peace spell	–0.004**	–0.012**	–0.054**	–0.037*
Length of last peace spell	(–13.16)	(–12.35)	(–5.63)	(–2.13)
Percentage of past time periods with conflict events	0.035**	0.030**	0.017**	0.012**
Percentage of past time periods with conflict events	(10.06)	(16.60)	(10.59)	(9.42)
Inverse-distance-weighted state casualties in other counties in t-1	1.284**	0.619**	1.280**	0.827**
	(6.43)	(6.25)	(4.82)	(5.92)
Inverse-distance-weighted insurgent casualties in other counties in t-1	0.084^†	–0.069**	-0.119**	–0.158**
	(1.90)	(–2.09)	(–2.74)	(–4.27)
State casualties in the county in t-1	0.102**	0.047**	0.060*	0.158**
State casualties in the county in t-1	(6.08)	(5.34)	(2.19)	(2.73)
Insurgent casualties in the county in t-1	0.020*	0.012**	0.025**	0.067**
Insurgent casualties in the county in t-1	(2.41)	(2.99)	(3.21)	(2.79)
County area in km squares	0.0001**	0.0001**	0.0001*	0.0001*
County area in km squares	(4.04)	(3.52)	(2.50)	(2.39)
Percentage of mountainous terrain in rural area	0.002*	0.003**	0.001	0.001
Percentage of mountainous terrain in rural area	(2.02)	(3.07)	(0.22)	(0.20)
Border status	0.037	0.028	0.056	0.115
Border status	(0.56)	(0.55)	(0.45)	(0.88)
Estimated parameters for seasonal and year dummies are not reported.
Base risk
Percentage of mountainous terrain in rural area	0.002	0.001	0.006	0.004
Percentage of mountainous terrain in rural area	(0.80)	(0.27)	(0.99)	(1.09)
Percentage of villages whose names were changed by the state	0.007**	0.008*	0.006*	0.004*
	(2.66)	(2.27)	(2.02)	(1.97)
Percentage of landless farmers	0.013**	0.008	0.011^†	0.007**
Percentage of landless farmers	(3.02)	(1.45)	(1.88)	(2.94)
Percentage of farmers with more than 100 acres of land	–0.023**	–0.021**	–0.014*	–0.010**
Percentage of farmers with more than 100 acres of land	(–3.45)	(–3.09)	(–2.17)	(–2.98)
Number of mosques per village	-0.590**	–0.260^†	–0.423*	–0.223*
Number of mosques per village	(–4.32)	(–1.68)	(–2.47)	(–2.49)
Percentage of villages with no drinking water	0.005	–0.002	0.003^†	0.002
Percentage of villages with no drinking water	(1.15)	(–0.70)	(1.93)	(1.27)
Percentage of villages with no electricity	0.004^†	0.003	0.005^†	0.003^†
Percentage of villages with no electricity	(1.81)	(1.46)	(1.86)	(1.95)
Border status	0.478	0.798**	0.341	0.159
Border status	(1.32)	(3.26)	(1.09)	(0.88)
Percentage of ethnically Kurdish population	0.031**	0.081**	0.013*	0.010**
Percentage of ethnically Kurdish population	(4.49)	(4.98)	(1.98)	(3.21)
Urbanization rate	–0.009	–0.004	–0.009	–0.002
Urbanization rate	(–1.39)	(–0.56)	(–1.46)	(–0.62)
Literacy rate	0.008	–0.013	0.008	0.004
Literacy rate	(0.95)	(–1.03)	(1.06)	(0.87)
Unemployment rate	0.107*	0.094*	0.085	0.041
Unemployment rate	(2.51)	(2.22)	(1.54)	(1.74)
Population in 10 thousand	0.016**	0.012**	0.022**	0.012**
Population in 10 thousand	(2.77)	(2.84)	(2.99)	(3.05)
Wald Chi-square	3,786.69	3,604.00	687.50	269.11
Correlation (rho) between the two equations(robust standard error)	0.341	0.684	–0.231	–0.912
	(0.191)	(0.131)	(0.325)	(0.282)
Wald test of rho=0: Chi-square	2.70*	11.56**	0.47	10.45**

Standard errors are adjusted for 648 counties.

z-values in parentheses.

^† p < 0.1, * p < 0.05, ** p < 0.01.

The estimated coefficients for the within-conflict-dynamics equation reveal the importance of geotemporal interdependencies. The longer a county stays peaceful the less likely it becomes to experience violence. Similarly, counties with more troubled histories, measured by percentage of past times with conflict events and with casualties in the county itself in the previous period, are more likely to experience conflict events.

As hypothesized, results indicate that geotemporal spillovers are conditioned by the losses of the sides. The estimated parameters for the inverse-distance-weighted-lagged-casualties indicate that it is the state force casualties in an area that are heralds of conflict events in nearby locations. On the other hand, events with insurgent casualties have a significant dampening impact on future events in the vicinity especially in the long run. These results are consistent with the asymmetric power structure in civil conflicts and with a rebel strategy of hit-and-run attacks whose likelihood of spilling over to nearby locations increases as state forces incur more military losses and decreases as insurgents themselves get hit.

The estimated parameters of the base risk equation point towards the importance of state reach and control. Mosques as preachers of state ideology, and big landowners as the agents of the state in the rural areas seem to make an important dampening impact on the underlying conflict risk.

As expected, base risk is significantly higher for counties with higher Kurdish population percentages. Those locations which had been subject to discriminatory state policies and had their names Turkified by the state have a higher likelihood of being part of the conflict area.

Difficult terrains are not significantly associated with the base risk but they facilitate the diffusion of violence in the short run.

To make sure that these results are not specific to the bi-probit model, I replicated the foregoing analyses with a simple logit model as well. The estimated associations between the incidence probability and the control variables remain similar in direction but get statistically stronger.⁹ I present the results in Table A5 in the Online appendix.

Figure 3 plots the average marginal effects of spatio-temporally lagged casualties in order to give a better understanding of their relative substantive significance. The marginal effect of a variable corresponds to the average expected change in the estimated probability of observing a fatal conflict event in county i at time t given a marginal change in that control variable while other control variables are evaluated at their mean values. As can be seen, the spatio-temporal lag of state casualties has a high marginal effect on the probability of diffusion. The average marginal effect of a unit increase in inverse-distance-weighted state casualties on the probability of observing a fatal conflict event at county i in

Figure 3.

Average marginal effects with 95% CIs

time t ranges from 2.5% to 7.1% across different temporal aggregations. On the other hand, the average marginal effect of a unit increase in inverse-distance-weighted insurgency casualties ranges from (-1.3) % to 0.0% depending on the time unit.¹⁰

Robustness of results

In this section, I employ the Coarsened Exact Matching (CEM) (Blackwell et al., 2009) technique as a robustness check on my results.

Matching is a nonparametric method of controlling for some or all of the confounding influence of pretreatment control variables. The key goal of matching is to prune observations in order to better approximate experimental conditions in observational data. More specifically, a statistical matching criterion based on the similarity of confounding factors or their effect on the probability of treatment is used to pair treated and untreated observations, and only those matched observations are retained in the dataset while the unmatched ones are discarded. I argue that matching is a good robustness check for the results I obtain from the split population bi-probit model because it in fact offers another way of defining the relevancy of observations based on their similarity in terms of treatment likelihood. Those unmatched observations are left out as irrelevant for the purpose at hand.

The predicted base risk gives me a composite pretreatment variable subsuming the socio-economic pre-conflict variables across which I can match my observations. Note that matching techniques are usually employed on cross-sectional data with binary treatments, whereas I have a panel dataset and my treatment, i.e. the spatio-temporally lagged casualties, is by design a continuous variable. The panel nature of the data is easily incorporated by treating groups of observations under each time unit as a separate cross-section subset and by conducting matching within each of these subsets.¹¹ The continuous treatment variable is a more challenging obstacle. Matching methods have rarely been applied to a continuous treatment. Researchers have proposed techniques to generalize the binary case to continuous treatments (Hirano & Imbens, 2004; Bia & Mattei, 2007; Fong, Hazlett & Imai, 2018); nevertheless, these techniques come with some untrivial assumptions (Fong, Hazlett & Imai, 2018) which are not suitable for my purposes.

Instead, I adopt the more common approach in applied social sciences and binarize my treatment. My binary treatment variable takes on the value 1 if location i at time t has a high inverse-road-distance-weighted average of conflict event incidences in the neighborhood at time t-1 and zero otherwise. Formally, I define my treatment variable as

{Treatment}_{i, t} = 1 if \sum_{k \neq i} ({1/d}_{ki}) {(incidence)}_{k, t-1} {> 90}^{th} percentile

{Treatment}_{i, t} = 0 otherwise

where d_ki is the road distance between locations k and i in kilometers.

The choice of the cutoff value is based on the distribution of the treatment variable and is chosen to distinguish between observations with high and low spatio-temporal incidence lags. To make sure that my results do not depend on my choice of cutoff value, I repeated the exercise with different cutoff percentiles such as the 75th and the 95th, and obtained substantively similar results.¹²

I then use the CEM to match my observations. I then run a probit regression with clustered errors at the county level on my matched sample. Table II presents the results.¹³ As can be seen, the estimated coefficients still associate events with state force casualties with higher risks of future events in the vicinity, while events with

Table II.

Results of the probit regression on matched sample

	Unit of observation:	Unit of observation:	Unit of observation:	Unit of observation:
	County-month (1984–2018)	County-quarter (1984–2018)	County-halfyear (1984–2018)	County-year (1984–2018)
	Number of obs.: 94,206	Number of obs.: 33,029	Number of obs.: 15,739	Number of obs.: 11,528
Length of last peace spell	–0.006**	–0.015**	–0.063**	–0.076**
Length of last peace spell	(–4.52)	(–5.96)	(–4.82)	(–4.22)
Percentage of past time periods with conflict events	0.041**	0.034**	0.020**	0.019**
	(8.81)	(12.13)	(15.64)	(18.12)

Inverse-distance-weighted state casualties in other counties in t-1	1.387**	0.648**	0.890**	0.671**
	(4.99)	(3.97)	(9.33)	(8.90)
Inverse-distance-weighted insurgent casualties in other counties in t-1	0.182	0.064	–0.107**	–0.103**
	(1.51)	(1.14)	(3.86)	(4.16)
State casualties in the county in t-1	0.096**	0.011	0.041**	0.035**
State casualties in the county in t-1	(3.04)	(0.98)	(3.57)	(3.16)
Insurgent casualties in the county in t-1	0.017	0.015*	0.012**	0.022**
Insurgent casualties in the county in t-1	(1.40)	(2.20)	(3.20)	(4.19)
County area in km squares	0.0001**	0.0002**	0.0001**	0.0001**
County area in km squares	(3.24)	(5.85)	(4.71)	(3.67)
Percentage of mountainous terrain in rural area	0.001	0.002^†	0.0003	0.001^†
Percentage of mountainous terrain in rural area	(0.79)	(1.75)	(0.36)	(1.73)
Border status	–0.086	–0.091	0.040	0.096^†
Border status	(–0.94)	(–0.91)	(0.68)	(1.74)

Estimated parameters for seasonal and year dummies are not reported.

Robust errors clustered at the county level.

z-values in parentheses.

^† p < 0.1, * p < 0.05, ** p < 0.01.

insurgency casualties are associated with lower risks of mayhem in the neighborhood in the long run.

Conclusion

Kalyvas (2008) applauds the emergence of the literature on micro-dynamics of civil conflicts as a very exciting development that deepens our understanding of political violence. But he also points out some recurrent flaws in the literature stemming from ‘insufficient theorization, superficial engagement with the case at hand and reliance on off-the-shelf datasets’ ( Kalyvas, 2008: 398). My starting point in this study echoes Kalyvas’ criticisms specifically for the emerging literature on within-country diffusion of civil conflicts.

I argue that existing works fail to acknowledge that conflict events can have incapacitating effects on warring sides. I tackle this shortcoming by hypothesizing that the geotemporal interdependency among conflict events is conditioned by the impact these events have on the operational capability of the sides of the conflict. I then test my hypotheses on a new and detailed event dataset on the long-running civil conflict in Turkey. TPCONED has been in the making for more than a decade. It relies on a wide range of local sources from different political affiliations and views, and as such offers comprehensive, accurate, high resolution and unbiased coverage of the long-running civil conflict between the Turkish state and the rebel organization PKK. I release TPCONED with the hope that it will be a valuable resource especially for in-depth studies on micro-level dynamics of civil conflicts.

One important point to note here is that the hypotheses tested in this study are about the diffusion patterns of irregular civil conflicts, and not about their outcomes or their level of violence. While the first hypothesis does imply that an irregular civil conflict is likely to continue and geographically expand as long as rebels are able to inflict damages on state forces, this implication does not allow us to conclude on a rebel victory in such conflicts. In fact, we know from cross-country studies that irregular civil conflicts do last long, but because of the power asymmetry that characterizes them, are mostly won by incumbents (Balcells & Kalyvas, 2014; Kalyvas & Balcells, 2010).

Similarly, while the second hypothesis does imply that states can curb the geographical spread of conflict events by inflicting damages on the insurgents, this implication does not mean that states can resolve conflicts by going on a killing rampage. Military coercion, as the theoretical arguments and the empirical results here indicate, can be effective in containing or even extinguishing violence in the short run, but it must be remembered that civil conflicts are population-centric contests with social, political and economic dimensions. Insurgent organizations will have the ability to generate political support and continue their violent campaign as long as states fail in their counterinsurgency efforts to address those social, economic and/or political problems and to secure the loyalties of civilians. The Turkish case itself is a good example. Turkish security forces dealt a major blow to the PKK through their use of extensive military coercion in the 1990s and even captured its leader Abdullah Öcalan in 1999. The PKK then declared a unilateral ceasefire. Many commentators heralded this as the PKK’s defeat and as the end of the conflict (Bila, 2000; Cemal, 1999). However, because subsequent Turkish governments failed to take the necessary steps to address the political, economic and social problems underlying the conflict, the organization easily recovered in the early 2000s and violence resumed.

Finally, results must be tempered by the study’s limitations. First, it must be admitted that, rather than absolute numbers, the ratio of casualties to group size in each location at each time period provides a better measure of the extent of the damage on the operational capabilities of the sides. Not surprisingly though, for strategic purposes, no army or insurgent organization reveals such information, hence working with ratios remains as an ideal.

Second, it must be emphasized that this is a single-case study. While the media coverage of ongoing conflicts around the world offers ample anecdotal evidence of the generalizability of results,¹⁴ a comparative study is yet to be conducted.

Footnotes

Replication data

The dataset, codebook, and do-files for the empirical analysis in this article, along with the Online appendix, can be found at http://www.prio.org/jpr/datasets. All statistical analyses were conducted using Stata15. TPCONED is available at the University of Warwick Research Archive Portal .

Acknowledgments

I thank the anonymous reviewers, and the Editor of JPR, Alex Braithwaite, and Vincenzo Bove and Kristian Skrede Gleditsch for helpful feedback on previous versions of the article.

ORCID iD

Arzu Kibris

Notes

References

Ayres

R William

Saideman

Stephen

(2000) Is separatism as contagious as the common cold or as cancer? Testing international and domestic explanations. International Interactions 6(3): 91–113.

Balcells

Laia

Kalyvas

Stathis N

(2014) Does warfare matter? Severity, duration, and outcomes of civil wars. Journal of Conflict Resolution 58(8): 1390–1418.

Bara

Corinne

(2018) Legacies of violence: Conflict-specific capital and the postconflict diffusion of civil war. Journal of Conflict Resolution 62(9): 1991–2016.

Beardsley

Kyle

Gleditsch

Kristian Skrede

Nigel

(2015) Roving bandits? The geographical evolution of African armed conflicts. International Studies Quarterly 59(3): 503–516.

Beiser

Janine

(2016) International information flows, government response and the contagion of ethnic conflict. In: Wilson

Alan

(ed.) Global Dynamics: Approaches from Complexity Science. Chichester: John Wiley & Sons, 214–230.

Bia

Michela

Mattei

Alessandra

(2007) Application of the generalized propensity score. Evaluation of public contributions to Piedmont enterprises. POLIS Working Paper 80. University of Eastern Piedmont.

Bila

Fikret

(2000) Bu nasıl diplomasi? Cumhuriyet 23 February.

Blackwell

Matthew

Iacus

Stefano

King

Gary

Porro

Guiseppe

(2009) CEM: Coarsened exact matching in Stata. Stata Journal 9(4): 524–546.

Bohnet

Heidrun

Cottier

Fabien

Hug

Simon

(2018) Conflict-induced IDPs and the spread of conflict. Journal of Conflict Resolution 62(4): 691–716.

10.

Braithwaite

Alex

(2010) Resisting infection: How state capacity conditions conflict contagion. Journal of Peace Research 47(3): 311–319.

11.

Braithwaite

Alex

Chu

Tiffany S

(2018) Civil conflicts abroad, foreign fighters, and terrorism at home. Journal of Conflict Resolution 62(8): 1636–1660.

12.

Braithwaite

Alex

Johnson

Shane D

(2012) Space–time modeling of insurgency and counterinsurgency in Iraq. Journal of Quantitative Criminology 28(1): 31–48.

13.

Braumoeller

Bear

Goertz

Gary

(2002) Watching your posterior: Comment on Seawright. Political Analysis 10(2): 198–203.

14.

Buhaug

Halvard

Gleditsch

Kristian Skrede

(2008) Contagion or confusion? Why conflicts cluster in space. International Studies Quarterly 52(2): 215–233.

15.

Carter

Timothy Allen

Veale

Daniel Jay

(2015) The timing of conflict violence: Hydraulic behavior in the Ugandan civil war. Conflict Management and Peace Science 32(4): 370–394.

16.

Cederman

Lars-Erik

Girardin

Luc

Gleditsch

Kristian Skrede

(2009) Ethnonationalist triads: Assessing the influence of kin groups on civil wars. World Politics 61(3): 403–437.

17.

Cederman

Lars-Erik

Wimmer

Andreas

Min

Brian

(2010) Why do ethnic groups rebel? New data and analysis. World Politics 62(1): 87–119.

18.

Cederman

Lars-Erik

Gleditsch

Kristian Skrede

Salehyan

Idean

Wucherpfennig

Julian

(2013) Transborder ethnic kin and civil war. International Organization 67(2): 389–410.

19.

Cemal

Hasan

(1999) Yol ayrımı [Road separation]. Milliyet 4 March.

20.

Clark

David

Regan

Patrick

(2003) Opportunities to fight: A statistical technique for modeling unobservable phenomena. Journal of Conflict Resolution 47(1): 94–115.

21.

Quy-Toan

Iyer

Lakshmi

(2010) Geography, poverty and conflict in Nepal. Journal of Peace Research 47(6): 735–748.

22.

Donnay

Karsten

Dunford

Eric T

McGrath

Erin C

Backer

David

Cunningham

David E

(2018) Integrating conflict event data. Journal of Conflict Resolution 63(5): 1137–1364.

23.

Duffy Toft

Monica

Zhukov

Yuri M

(2012) Denial and punishment in the North Caucasus: Evaluating the effectiveness of coercive counter-insurgency. Journal of Conflict Resolution 49(6): 785–800.

24.

Fearon

James D

Laitin

David D

(2003) Ethnicity, insurgency and civil war. American Political Science Review 97(1): 75–90.

25.

Fong

Christian

Hazlett

Chad

Imai

Kosuke

(2018) Covariate balancing propensity score for a continuous treatment: Application to the efficacy of political advertisement. Annals of Applied Statistics 12(1): 156–177.

26.

Forsberg

Erika

(2008) Polarization and ethnic conflict in a widened strategic setting. Journal of Peace Research 45(2): 283–300.

27.

Gleditsch

Kristian Skrede

Salehyan

Idean

Schultz

Kenneth

(2008) Fighting at home, fighting abroad. Journal of Conflict Resolution 52(4): 479–506.

28.

Grossman

Herschell

(1991) A general equilibrium model of insurrections. American Economic Review 81(4): 912–921.

29.

GTD – Global Terrorism Database (2020) National consortium for the study of terrorism and response to terrorism (https://www.start.umd.edu/data-tools/global-terrorism-database-gtd).

30.

Haushofer

Johannes

Biletzki

Anat

Kanwisher

Nancy

(2010) Both sides retaliate in the Israeli-Palestinian conflict. Proceedings of the National Academy of Sciences 107(42): 17927–17932.

31.

Hegre

Håvard

Østby

Gudrun

Raleigh

Clionadh

(2009) Poverty and civil war events: A disaggregated study of Liberia. Journal of Conflict Resolution 53(4): 598–623.

32.

Hill

Stuart

Rothchild

Donald S

Cameron

Colin

(1998) Tactical information and the diffusion of peaceful protests. In: Lake

David A

Rothchild

Donald S

(eds) The International Spread of Ethnic Conflict: Fear, Diffusion, and Escalation. Princeton, NJ: Princeton University Press, 61–88.

33.

Hirano

Keisuke

Imbens

Guido W

(2004) The propensity score with continuous treatments. In: Gelman

Meng

X-L

(eds) Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives. Chichester: Wiley InterScience, 73–84.

34.

Holtermann

Helge

(2012) Explaining the development-civil war relationship. Conflict Management and Peace Science 29(1): 56–78.

35.

Holtermann

Helge

(2016) Relative capacity and the spread of rebellion: Insights from Nepal. Journal of Conflict Resolution 60(3): 501–529.

36.

Jaeger

David A

Paserman

M Daniele

(2008) The cycle of violence? An empirical analysis of fatalities in the Palestinian-Israeli conflict. American Economic Review 98(4): 1591–1604.

37.

Kalyvas

Stathis

(2006) The Logic of Violence in Civil War. Cambridge: Cambridge University Press.

38.

Kalyvas

Stathis

(2008) Promises and pitfalls of an emerging research program: The microdynamics of civil war. In: Kalyvas

Stathis

Shapiro

Ian

Masoud

Tarek

(eds) Order, Conflict and Violence. Cambridge: Cambridge University Press, 397–421.

39.

Kalyvas

Stathis

Balcells

Laia

(2010) International system and technologies of rebellion: How the end of the Cold War shaped internal conflict. American Political Science Review 104(3): 415–429.

40.

Kibris

Arzu

(2011) Funerals and elections: The effects of PKK terrorism on voting behavior in Turkey. Journal of Conflict Resolution 55(2): 220–247.

41.

Kocher

Matthew Adam

Pepinsky

Thomas B

Kalyvas

Stathis

(2011) Aerial bombings and counterinsurgency in the Vietnam War. American Journal of Political Science 55(2): 201–218.

42.

Lake

David A

Rothchild

Donald

(1998) Spreading fear: The genesis of transnational ethnic conflict. In: Lake

David A

Rothchild

Donald

(eds) The International Spread of Ethnic Conflict. Princeton, NJ: Princeton University Press, 3–34.

43.

Lane

Matthew

(2016) The intrastate contagion of ethnic civil war. Journal of Politics 78(2): 396–410.

44.

Linke

Andrew M

Witmer

Frank DW

O’Loughlin

John

(2012) Space-time granger analysis of the war in Iraq: A study of coalition and insurgent action-reaction. International Interactions 38(4): 402–425.

45.

Lyall

Jason M

(2009) Does indiscriminate violence incite insurgent attacks? Evidence from Chechnya. Journal of Conflict Resolution 53(3): 331–362.

46.

Machiavelli

Nicolo

(n.d.) The Prince. [Online] (https://www.gutenberg.org/files/1232/1232-h/1232-h.htm accessed 11 August 2020).

47.

Mahoney

James

Goertz

Gary

(2004) The possibility principle: Choosing negative cases in comparative research. American Political Science Review 98(4): 653–669.

48.

Maves

Jessica

Braithwaite

Alex

(2013) Autocratic institutions and civil conflict contagion. Journal of Politics 75(2): 478–490.

49.

McColl

Robert W

(1969) The insurgent state: Territorial bases of revolution. Annals of the Association of American Geographers 59(4): 613–631.

50.

Metternich

Nils W

Minhas

Shahryar

Ward

Michael D

(2017) Firewall? Or wall on fire? A unified framework of conflict contagion and the role of ethnic exclusion. Journal of Conflict Resolution 61(6): 1151–1173.

51.

Mikulaschek

Christoph

Shapiro

Jacob N

(2018) Lessons on political violence from America’s post-9/11 wars. Journal of Conflict Resolution 62(1): 174–202.

52.

O’Loughlin

John

Witmer

Frank DW

(2012) Diffusion of violence in the North Caucasus of Russia, 1999–2010. Environment and Planning A: Economy and Space 44(10): 2379–2396.

53.

Pegg

Scott M

(1998) International Society and the De Facto State. Aldershot: Ashgate.

54.

Pulur

Hasan

(2010) Özal 26 yıl önce ‘üç beş çapulcu’ demişti. (Özal said “a handful of outlaws” 26 years ago). Milliyet 5 August.

55.

Raleigh

Clionadh

Hegre

Håvard

(2009) Population size, concentration, and civil war. A geographically disaggregated analysis. Political Geography 28(4): 224–238.

56.

Raleigh

Clionadh

Linke

Andrew

Hegre

Håvard

Karlsen

Joakim

(2010) Introducing ACLED: An Armed Conflict Location and Event Dataset. Journal of Peace Research 47(5): 651–660.

57.

Rüegger

Seraina

(2018) Refugees, ethnic power relations, and civil conflict in the country of asylum. Journal of Peace Research 56(1): 42–57.

58.

Salehyan

Idean

Gleditsch

Kristian Skrede

(2007) Refugees and the spread of civil war. International Organization 60(2): 335–366.

59.

Schutte

Sebastian

(2017) Regions at risk: Predicting conflict zones in African insurgencies. Political Science Research and Methods 5(3): 447–465.

60.

Schutte

Sebastian

Donnay

Karsten

(2014) Matched wake analysis: Finding causal relationships in spatiotemporal event data. Political Geography 41(July): 1–10.

61.

Schutte

Sebastian

Weidmann

Nils B

(2011) Diffusion patterns of violence in civil wars. Political Geography 30(3): 143–152.

62.

Schultz

Kenneth A

(2010) The enforcement problem in coercive bargaining: Interstate conflict over rebel support in civil wars. International Organization 64(2): 281–312.

63.

Sundberg

Ralph

Melander

Erik

(2013) Introducing the UCDP Georeferenced Event Dataset. Journal of Peace Research 50(4): 523–532.

64.

Tellis

Ashley J

Bially

Janice

Layne

Christopher

McPherson

Melissa

Sollinger

Jerry M

(2000) Measuring National Power in the Post-Industrial Age. Santa Monica, CA: Rand Corporation.

65.

Tezcür

Güneş Murat

(2016) Ordinary people, extraordinary risks: Joining an ethnic rebellion. American Political Science Review 110(2): 247–264.

66.

Townsley

Michael

Johnson

Shane D

Ratcliffe

Jerry H

(2008) Space time dynamics of insurgent activity in Iraq. Security Journal 21: 139–146.

67.

Tull

Denis

(2004) The Reconfiguration of Political Order in Postcolonial Africa: A Sase Study from North Kivu (DR Congo). Unpublished PhD Dissertation, University of Hamburg.

68.

Ward

Michael D

Gleditsch

Kristian Skrede

(2010) Location, location, location: An MCMC approach to modelling the spatial context of war and peace. Political Analysis 10(3): 244–260.

69.

Weidmann

Nils B

(2015) Communication networks and the transnational spread of ethnic conflict. Journal of Peace Research 52(3): 285–296.

70.

Weidmann

Nils B

Ward

Michael D

(2010) Predicting conflict in space and time. Journal of Conflict Resolution 54(6): 883–901.

71.

Weinstein

Jeremy M

(2007) Inside Rebellion: The Politics of Insurgent Violence. Cambridge: Cambridge University Press.

72.

Xiang

Jun

(2010) Relevance as a latent variable in dyadic analysis of conflict. Journal of Politics 72(2): 484–498.

73.

Zhukov

Yuri M

(2012) Roads and the diffusion of insurgent violence: The logistics of conflict in Russia’s North Caucasus. Political Geography 31(3): 144–156.