Hearts, Minds, and Pockets: the Political Economy of Civilian Support for Kurdish Insurgency in Turkey

Hearts, Minds, and Pockets: The Political Economy of Civilian Support for Kurdish Insurgency in Turkey

Tolga Sinmazdemir∗

March 28, 2016

∗Assistant Professor, Department of Political Science and International Relations, Bogazici University, Istan- bul; Research Oﬃcer, Department of Government, University of Essex Abstract

Does poverty make ethnic insurgencies more likely? I argue that if (a) civilian support for insurgents plays a decisive role in their success, and if (b) civilians experience a reduction in public goods provision after supporting insurgents, then insurgent attacks are less likely to occur in poorer regions. This is because lower income makes poorer individuals more vulnerable to losing public goods, and therefore makes them less willing to support insurgents. I test this argument using data on attacks by the Kurdish insurgency between 1988 and 1999 in 20 Turkish provinces that are poorer than the national average, and in which Kurds form a signiﬁcant share of the population. I ﬁnd that in provinces with initially high levels of public goods provision, attacks are less likely when per-capita income of the province is a smaller percentage of the national average. Moreover, higher unemployment is associated with fewer attacks.

Keywords: income inequality, unemployment, public goods provision, insurgency, ethnic conﬂict

Word Count: 9971 Do economic grievances play a role in insurgency outbreaks? While there is an ongoing scholarly debate on this question,1 recent counterinsurgency eﬀorts presume that economic grievances do play a critical role in insurgent violence, and economic development assistance has been used extensively as a tool for combating insurgencies in countries such as Iraq and

Afghanistan (Beath et al. 2011). The logic underlying these eﬀorts is straightforward: Civilian support is critical for successful insurgency, and economically better-oﬀ civilians are less suscep- tible to the appeal of armed groups. Therefore, economic development should make insurgent violence less likely.2

The literature on the political economy of insurgencies does not provide a clear empirical support for this logic: On the one hand, recent studies document that public goods provision lowers violence in countries involved in civil wars and insurgency (Beath et al. 2011; Berman et al. 2011b), and individuals with lower income are more likely to become involved in violence

(Humphreys and Weinstein 2008; Justino 2009). On the other hand, there are subnational studies that find no relationship between unemployment and insurgent attacks (Berman et al. 2011a), and between poverty and support for political violence (Blair et al. 2012). In this paper, I contribute to this literature, and propose an explanation of insurgent violence that links together the effects of income and of public goods provision on civilians’ willingness to assist militants. My explanation builds on an additional body of research that finds that populations associated with militants suffer negative economic consequences in the aftermath of violence

(Abadie and Gardeazabal 2003; Benmelech et al. 2010).

Following these findings in the literature, I hypothesize that if violence leads to a reduction in public goods provision, then the effect of this reduction on civilian support for insurgents depends on the initial income of the local population, and on the initial level of public goods provision by the government. Specifically, I argue that provision of public goods prior to the outbreak of violence makes poorer populations less likely to assist militants because the prospect of losing public goods, partially or even entirely, increases their opportunity cost for involvement in an insurgency. Since assistance from non-combatants, for example in the form of material support, shelter, or information, is vital for the insurgents’ ability to fight (Justino

1See e.g. Cederman et al. (2011) for a recent review. 2See e.g. U.S. Army’s Counterinsurgency Field Manual (2007). See also Blair et al. (2012) for other examples of the same logic motivating economic assistance from U.S., U.K. and international organizations to Pakistan.

3 2009; Kalyvas 2007), we should observe less violence where the local population is poorer, and where violence can lead to a reduction in public goods provision.

To test this argument, I collect province-level data on per capita income, unemployment, public goods provision, and attacks by the Kurdish insurgency (Partiya Karkeren Kurdistan

(PKK) which in Kurdish means Kurdistan Workers’ Party) between 1988 and 1999 in 20 provinces of Turkey. PKK was founded in late 1970s as a Marxist-Leninist group, aiming at a separate state for the minority Kurds living in Turkey, in response to several assimilationist measures by the Turkish state such as bans on the use of Kurdish language in schools, in publications and in names of places that are predominantly Kurdish.3 The only oﬃcial language of the country has been (and still is) Turkish. The ﬁrst major attacks of PKK took place in 1984.

Following the capture of its leader Abdullah Ocalan in 1999, the demands of the insurgents have changed into political and cultural autonomy for Kurds living in the southeastern parts of

Turkey (Kapmaz 2011).

Studies of insurgencies have demonstrated that an active and prolonged insurgency requires public support (Kalyvas 2006; Petersen 2001; Weinstein 2007). I interpret the presence of PKK attacks as an indicator of civilian support for insurgents in areas where these attacks took place. I follow this interpretation for four reasons: First, as I show in my empirical analysis, insurgent attacks are more likely in provinces with Kurdish majorities, and support for Kurdish insurgents should be higher in these Kurdish majority provinces relative to ethnically more mixed provinces. Second, my empirical analysis shows that attacks are more likely in provinces with lower turnout in elections. Since turnout can be seen as an indicator of support for the government and the Turkish political system, this negative relationship between attacks and turnout suggests that attacks are more likely to happen in provinces with lower support for the government, and hence higher support for the insurgents. Third, and perhaps most importantly, alternative interpretations of insurgent attacks (e.g. attacks proxying for support for the government) are not consistent with patterns in the data. More speciﬁcally, my analysis shows that insurgent attacks are more likely in provinces with lower public goods provision. If attacks are a proxy for support for the government, this pattern implies that lower public goods

3For a history of the Kurds, see McDowall (1996) and Van Bruinessen (1992). For in-depth analyses of the Kurdish question in Turkey, see Barkey and Fuller (1998); Kirisci and Winrow (1997); Van Bruinessen (2000). For a study of PKK insurgency, see Marcus (2007).

4 provision is associated with higher support for the government. This is a highly implausible scenario given the fact that these provinces are poor and beneﬁt from public goods extensively.

Fourth, prior research on PKK violence supports my interpretation of attacks as a proxy for support for the insurgents. For instance, using election data from 1999, 2004 and 2007, Gergin

(2012) ﬁnds that more popular support for the insurgents (measured as the vote share of the political party aﬃliated with PKK) is associated with more insurgent attacks.

My econometric analysis of insurgent attacks supports the hypothesis that the prospect of adverse economic consequences due to insurgent violence affects civilian collaboration with the militants, and the ultimate level of insurgent attacks, conditional on the initial level of public goods provision and income of the local residents: In provinces with high levels of investment in public goods, insurgent attacks are less likely where the income per capita is lower. However, when investment in public goods is low, insurgent attacks are more likely in provinces where income per capita is lower. Moreover, in provinces with high levels of investment, an increase in unemployment is associated with a lower rate of attacks. My argument hinges on the premise that insurgent violence leads to a subsequent decline in public goods provision in the affected areas. To demonstrate that this premise is empirically supported, I also provide qualitative and quantitative evidence on the adverse effects of insurgent violence on different types of public goods provision such as health services and education in these provinces.

In what follows, ﬁrst I review the literature that my paper relates to and highlight the contribution of this paper. Then, I outline the logic of my hypotheses in greater detail and also provide qualitative and quantitative evidence that supports the logic of my hypotheses.

Having speciﬁed my hypotheses, I describe the data collected for the empirical tests, and present regression results of the spatial and temporal distribution of civilian killings by PKK both at the province and the lower district level. I conclude with a summary and discussion of the implications of my results. All tables and ﬁgures are in the Appendix.

Poverty, Public Goods Provision, and Conﬂict

Motivated by the robust positive correlation between low per-capita income and civil war risk at the cross-national level (Fearon and Laitin 2003; Hegre and Sambanis 2006), there is a growing

5 body of work that uses sub-national data to understand further the links between local economic conditions and the likelihood of civil wars and insurgencies.4 Part of this literature documents a positive relation between lower standards of living and risk of political violence in diﬀerent conﬂicts (see Barron et al. 2004 for village-level violence in Indonesia, see Bohara et al. 2011; Do and Iyer 2010; Murshed and Gates 2005 on Maoist insurgency in Nepal). However, some other recent works report that better economic conditions do not necessarily lower the risk of violence.

Using data on insurgencies in Afghanistan, Iraq and Philippines, Berman et al. (2011a) show that there is no positive correlation between unemployment and insurgent attacks. Similarly,

Hegre et al. (2009) show in their study of the Liberian civil war that war events are more frequent in richer locations. Most recently, the link between economic conditions and support for terrorism has also been questioned. In their survey of Pakistanis, Blair et al. (2012) ﬁnd that support for militant politics is stronger among middle class citizens relative to poor citizens.

Hence, the question of the links between local economic conditions and political violence is far from settled.

Secondly, there is an earlier strand of cross-national research on the link between economic inequality and political violence.5 Originally motivated by Tedd Gurr (1970)’s relative deprivation theory, this line of research has argued that economic grievances either due to unfulfilled aspirations or uneven distributions of land or income are a significant cause of various forms of political violence (Mueller and Seligson 1987; Wickham-Crowley 1992). However, recent cross-national research on civil wars has failed to find a link between measures of economic inequality and civil war onset (Collier and Hoeffler 2004; Fearon and Laitin 2003). Scholars have responded to this non-finding by showing the positive effect of horizontal inequalities among culturally-defined groups on violence in different settings (Cederman et al. 2011; Stewart 2008).

Thirdly, my paper is also related to the recent empirical tests of the ”hearts and minds approach” in counterinsurgency, which emphasizes the importance of civilian attitudes towards the government in defeating insurgencies.6 Recently, scholars have subjected this theory to

4For a review of this particular literature, see Sambanis (2004). For a review of the literature on civil wars in general, see Blattman and Miguel (2010); Kalyvas (2007). 5For a review of earlier works in this literature, see Lichbach (1989). For a more recent review of the works on the link between economic inequality and poverty and political violence, see Sambanis (2004). 6For early works, see Galula (1964) and Taber (1965). More recent examples that highlight the importance of civilian support include Petraus (2006) and Sepp (2005). For a recent review of this literature, see Berman et al. (2011b).

6 empirical testing by studying the eﬀect of development assistance and public goods provision on insurgent violence and civilian perceptions of government in diﬀerent settings. Berman et al. (2011b) show that small scale reconstruction projects have reduced violence especially after the increase in troop strength in 2007 in Iraq. Similarly, Beath et al. (2011) report that in

Afghanistan, involvement of villages in development aid programs improve the perception of economic well-being and attitudes towards the government, and it also lowers security incidents recorded by International Security Assistance Force (ISAF) one year or more after project implementation. In contrast, Crost et al. (2014) ﬁnd that in Philippines, municipalities that are eligible for development projects have experienced an increase in violence for the duration of the projects.

Finally, this paper is also related to the small but growing literature on economic determinants of political violence and terrorism in Turkey.7 While some recent studies claim that economic deprivation due to low income and unemployment is one of the causes of terrorist incidents in Turkey (Feridun and Sezgin 2008, Yildirim and Ocal 2013), other studies ﬁnd no relationship between economic measures such as low growth and income inequality, and the incidents of separatist terrorism in Turkey (Derin-Gure 2011, Derin-Gure and Elveren 2014).

In this paper, my contribution to these studies is threefold: First, I present a theoretical argument that offers a potential explanation for the contradictory findings in the literature on the link between local economic conditions and the likelihood of violence. More specifically,

I argue that violence may be inversely related to measures such as poverty or unemployment if public goods provision is extensive, and future violence has the potential of disrupting this provision. Second, in order to test this argument, I present a measure of group-level inequality between Kurds and the rest of the country using GDP per capita values in provinces where Kurds form a signiﬁcant share of the population and where most of the insurgent attacks took place.

Using this measure, I provide new quantitative evidence of the eﬀect of horizontal inequalities on ethnic violence from a case that has been mostly neglected in the literature on civil wars and insurgencies although it is one of the longest running and deadliest conﬂicts of the post World

War II era.8 Finally, I contribute to the literature on successful counterinsurgency by showing

7For a concise summary of this literature, see Derin-Gure and Elveren (2014). 8For instance, according to Fearon (2004)’s coding of duration of civil wars, PKK insurgency ranks 5th in terms of its duration and deaths combined.

7 that high levels of public goods provision may reduce insurgent violence overall but it may also create a new set of incentives among civilian population, and as a result of these incentives, regions with less unemployment and higher income may produce more violence relative to poorer regions even if economic grievances continue to play a role in civilians’ decision to assist insurgents.

Logic of the Argument

Supporting and assisting insurgents can be costly for civilian population for two reasons. First, insurgent violence can inhibit economic activity and reduce income. Extant scholarship provides evidence of this adverse effect in a variety of places affected by violence.9 There is evidence of similar negative effects of violence on individuals’ income also in the context of Kurdish insurgency in Turkey. First, several reports both from local and international sources show that government-imposed restrictions on the movement of people as well as bans on the use of land and pastures with the aim of defeating the insurgents have severely hurt agricultural production and animal husbandry in the region (Human Rights Watch 2002; Turkish Parliament 1998). In addition, there is also qualitative evidence that villagers dependent on agriculture have not been able to work their fields in a timely manner simply due to lack of security.10 Moreover, a substantial number of individuals have been also affected by village evictions implemented by the government to cut the logistical support of the insurgents. As a result of these evictions, the villagers have lost access to their land, and their agricultural equipment, stored produce and livestock were burnt and destroyed (Human Rights Watch 1993; 2005).11

Second, assisting militants can be costly for civilians also because violence can lead to a

9For instance, in the context of the Israeli-Palestinian conflict, Benmelech et al. (2010) show that suicide attacks lead to a higher unemployment and to lower wages in home districts of Palestinian suicide bombers in the aftermath of the attacks. Similarly, Abadie and Gardeazabal (2003) report that terrorism in Basque country has led to a 10 percentage point decline in GDP per capita relative to neighboring regions. For research on the adverse effects of violence on income, poverty and wealth at the household level in several civil-war countries, see the papers at the Households and Conflict Network at http://www.hicn.org/wordpress/?page_id=28. 10For example, a Human Rights Watch Report (1993) has the following quote from a local resident: “The products in the field - cotton and rice and tobacco - can’t be properly harvested because people can’t leave their houses in the evenings. It’s too dangerous. Some crops need to be harvested quickly. Even at night. These crops are the most important factors in the economy.” 11According to the report of the Turkish Parliament (1998) on this issue, in total, by 1997, 820 villages and 2345 hamlets were evicted. The total number of individuals who were forced to emigrate as a result of these evictions is 378335.

8 reduction in public goods provision. This can happen due to a deliberate government effort to reward those who refrain from assisting militants, and punish those who are associated with them.12 Alternatively, public goods provision in the sectors of education, health and infrastructure can drop in the aftermath of violence also due to physical damage or safety concerns of public servants.13 There is both quantitative and qualitative evidence that shows that reductions of public goods provision especially in the sectors of education, health and infrastructure have also occurred in the regions affected by Kurdish insurgency in Turkey. For instance, Turkish Parliament’s aforementioned report (1998) documents that in 1997, in 11 provinces of the region, 2202 out of a total of 5093 public schools at the primary and secondary level were closed either due to lack of security or due to insufficient number of teachers.14 In the same provinces, out of 387 health centers, 87 were closed also due to lack of security or insufficient number of health personnel.15 There are also interviews with villagers living in the area, who have told that the village schools were shut down because no teacher came for fear of the ongoing insurgency (Human Rights Watch 1994). Similarly, extensive infrastructure damage such as the destruction of electricity grids and telephone lines have also been reported by the administrators who have worked in the region and by the villagers (Human Rights Watch 2005;

Turkish Parliament 1998). Finally, there is also anecdotal evidence showing that withdrawing public health services has been used by local authorities as a method of punishing civilians that cooperate with the insurgents.16

If support for militants and violence is costly both in terms of income and public goods provision, then the cost of supporting militants should also depend on the income of the local population and the level of public goods provision prior to the outbreak of violence. In the case

12For example, in the case of Afghanistan, Berman et al. (2011b) report survey evidence which highlights the conditionality of the allocation of reconstruction funds under Commander’s Emergency Response Program (CERP) such that the implementation of a reconstruction project under CERP would be halted if the population increased its support for anti-government elements. 13See Humphreys (2003) who reports that ﬁve years after the end of war in Liberia, the capital Monrovia did not have electricity and running water. Bruck (2001) reports destruction of primary schools and hospitals and health posts due to civil war in Mozambique. For evidence of the destruction of educational infrastructure and buildings in several civil-war countries, see Buckland (2005). 14The provinces and the number of closed schools are as follows: Diyarbakir (368), Hakkari (92), Siirt (22), Sirnak (197), Tunceli (272) , Van (147), Batman (186), Bingol (281), Bitlis (191), Mardin (285), Mus (161). 15The provinces and the number of closed health centers are as follows: Diyarbakir (59), Hakkari (11), Siirt (20), Sirnak (17), Tunceli (15), Van (15), Batman (17), Bingol (23), Bitlis (44), Mardin (33). 16For instance, Human Rights Watch Report (2002, 22) tells the story of two villagers who needed the signature of local gendarmerie commander for their application form for the so-called Green Card that enables individuals to get free consultation and hospital accommodation. One of the individuals had ”PKK member but not on the wanted list” written on his form while the others’ form was torn up by the gendarmerie.

9 of no initial public goods provision, poorer civilians should be more likely to support militants due to stronger economic grievances and their lower opportunity cost of violence. If the initial level of public goods provision is relatively high, and violence leads to a drop in the provision of public goods, then poorer civilians should be less likely to support militants than wealthier civilians. This is because poorer civilians are more aﬀected by the reduction in public goods provision than wealthier individuals.

To clarify further how this argument works, let’s look at a numerical example that compares the incentives of two members of a minority (let’s call them m and n) with ethnic grievances for supporting an insurgency in a country that ignores their identity-based political demands such as greater autonomy, recognition of cultural rights or preferential policies. Suppose the average income in the country is 10 dollars per capita while individual m‘s income is 5 dollars and individual n‘s income is 8 dollars. The state redistributes income through taxation and uses all tax proceeds on public goods provision. When the tax rate in this country is equal to

100 percent,17 both m and n enjoy an income of 10 dollars after taxation and the provision of public goods. An individual considering supporting an insurgency has to weigh the potential costs and benefits. The benefits are the possibility of achieving his political demands, or at least part of them. The costs are due to a reduction in income as a result of fighting, and the potential loss of public goods, if such goods are provided by the government.

Hence, in this example, following the onset of insurgent violence, due to the loss of public goods, individual m would be left with 5 dollars minus the costs of fighting while n would be left with 8 dollars minus the costs of fighting. This means that the total cost of violence for m would be 5 dollars (=10-5) plus the costs of fighting while n’s total loss would be 2 dollars (=10-8) plus the costs of fighting. Therefore, if the costs of fighting are proportional to individual income

(let’s say each individual that lives in an area where insurgent violence is prevalent would lose

50 percent of their income due to destruction of economic activity), m’s total cost would be 7.5

(= 5 + (50 % * 5)) dollars while n’s total cost would be 6 (= 2 + (50 % * 8)) dollars. Hence, n would prefer to support the insurgents as long as he values the political goals of the insurgents more than 6 dollars while m would prefer to support them only if he values those goals for more than 7.5 dollars. Because of this, although n has a higher income, he has stronger incentives to

17The logic of the argument does not depend on the assumption of full redistribution. As it will become clear in a moment, it would stand as long as the tax rate is greater than 50 percent.

10 support the insurgents.

Note that this claim depends critically on the level of redistribution and public goods provision in a country. If the tax rate and public goods provision is low, individual m, who has a lower income than individual n, would also have a lower total cost of violence. For instance, if the tax rate is 0 percent, then the only losses would be due to destruction of economic activity and reduction in income. In that case, m’s total cost would be 2.5 (50 % * 5)) dollars while n’s cost would be 4 dollars. Therefore, n has stronger incentives to support the insurgents only if the tax rate and hence public goods provision is high.

Table 1 lists the total cost of violence for m and n under the following scenarios: when public goods are provided through full redistribution and when there are no public goods.

As in the numerical example above, the total cost for each individual is calculated under the assumption that ﬁghting destroys half of an individual’s income. As the table shows, when there is no redistribution and public goods provision, total cost of violence for individual n is higher. However, when the tax rate is equal to 100 percent and all tax proceeds are used for public goods provision, total cost of violence for n is less than the cost for m (6 vs. 7.5). This relation between incomes and costs of violence would remain the same as long as we compare them for m and n when the tax rate for ﬁnancing the provision of public goods is less than 50 percent and when it is higher than 50 percent.

There are 3 hypotheses that can be drawn from the analysis in Table 1: First, when no or few public goods are provided, poorer individuals are more likely to support militants because of the lower opportunity costs of violence and stronger economic grievances. Thus, in the Turkish context, we should observe more violence in poorer provinces when no or few public goods are provided. Second, regardless of their initial income, all individuals are less likely to support militants when public goods are provided than they are when they receive no or few public goods. Therefore, we should observe less violence in provinces with more public goods provision compared to similar provinces with less public goods provision. Third, when public goods are provided extensively, poorer individuals become less likely to support militants than wealthier individuals. As a result, we should observe less violence in poorer provinces with high levels of public goods than in wealthier provinces that enjoy the same high level of public goods.

I test these hypotheses by using sub-national data on the Kurdish insurgency in Turkey.

11 Data Structure and Description of Variables

The Turkish census does not ask about individuals’ ethnic identity.18 Nevertheless, the 20 eastern and southeastern provinces of Turkey are known to be Kurds’ historical territories in

Turkey.19 In 10 of these provinces, Kurds’ population share is estimated to be more than 50 percent (Gunter 1990; McDowall 1996).20 Because of data limitations for the key independent variables, the start year of my sample is 1988. Since PKK has declared a ceaseﬁre in 1999 following the capture of its leader Abdullah Ocalan, I chose 1999 as the end year of the sample.

Dependent Variable

The dependent variable is based on my coding of the incidents reported in Ozdag (2009), which reports all PKK attacks with three or more civilian fatalities. There have been 385 such attacks between 1984 and 2009. In these 385 attacks, 2434 civilians are killed. These ﬁgures exclude casualties among the members of the state security forces targeted by the insurgents. Thus, the total number of killings perpetrated by the insurgency until 2010 is estimated to be equal to 12340 (Sener 2010). Because of this data limitation, instead of coding a dependent variable measuring the incidence of violent conﬂict on the basis of yearly and cumulative number of fatalities,21 I opted for coding a binary dependent variable Conflict which is equal to 1 if there is at least one attack with 3 or more civilian fatalities in a province-year.

There is a significant amount of variation among provinces in terms of the total number of attacks and their distribution over time between 1988 and 1999.22 While there are provinces which have not experienced a single attack, there are also some which had more than 10 attacks in a year. Since there is a wide variation in the number of attacks over time and space, I also repeat my empirical tests using other definitions of the dependent variable. More specifically, I coded the dependent variable Conflict as equal to 1 if there are at least 2, 3, 4 and 5 attacks in a province-year. As an alternative way of capturing different intensities of violent conflict,

18The last census in which people were asked their mother tongue has been conducted in 1965. 19See for instance Gunter (1990). 20These 20 provinces are listed in the Appendix. The provinces of Batman and Sirnak enter the sample in 1990, and Ardahan and Igdir enter in 1992 because they are created in that year out of the territories of preexisting provinces. 21See for instance Fearon and Laitin (2003) and Gleditsch et al. (2002) for this coding practice. 22See Figure 1 in the Appendix.

12 I also coded Conflict as equal to 1 if there is at least one attack in a year with 10 or more civilian fatalities.

Independent Variables of Interest

My independent variables of interest are measures of income for the Kurds and public goods provision in these provinces. I measure income by collecting province-level data on GDP in

Turkish liras in 1987 prices. By using these data and the yearly population ﬁgures, I calculated

GDP per capita values for each province-year. Then, I calculated the ratios of province-level

GDP per capita to the national GDP per capita for each year in the sample. Hence, this variable, called GDP/cap Ratio, measures per capita income disparity between a province and the national average.23 The ratios of GDP/cap to the national average is less than 1 for all province-years, and the values vary between .15 and .88.24 Hence, the provinces where Kurds form a signiﬁcant portion of the population are poorer than the national average of the country for all years in the sample, and this makes the Kurdish insurgency an apt case for testing my hypotheses on the relation between low income, public goods provision and the onset of an ethnic insurgency in a region historically associated with an ethnic minority.

I proxy province-level public goods provision by using yearly expenditure data provided by the Turkish Development Ministry on public investment projects funded by the state in the sectors of education, health, infrastructure, housing, manufacturing, transportation, tourism, agriculture, mining, and energy. By using the expenditure data on each of these sectors and the yearly population ﬁgures, I calculated the total public investment levels per capita and also public investment levels per capita in each of these 10 sectors separately.25

Controls

My control variables are log of population, turnout rate in the latest parliamentary elections, and dummies that measure whether the Kurdish population forms the majority of the total population in a province and whether these provinces share a border with another country.

23My empirical results do not change if I use GDP/cap values of these provinces directly as my independent variable. 24See Figure 2 in the Appendix. Summary statistics for the variables included in the province-level tests are in Table 2 in the Appendix. 25For descriptions of the projects that are covered under each sector, see Appendix.

13 Large populations have been identified several times as significant determinants of civil war in quantitative analyses (Fearon and Laitin 2003, Hegre and Sambanis 2006). I include a dummy called Kurdish Majority that denotes Kurdish-majority provinces in order to control for the effect of the intensity of Kurdish presence on the likelihood of attacks. It is well known that the mountainous area at which the borders of Turkey, Iraq and Iran intersect has provided an opportunity for lightly armed groups to cross the border into Turkey and carry out attacks.

Therefore, I also include a dummy called Border that measures whether a province shares a border with a neighboring country. Since insurgent activity has become especially tense and strong after 1992 (Barkey 2007)26, I also include a dummy called P ost − 1992.

I control for turnout for the following reason: There are two alternative interpretations of the dependent variable Conflict. It can be a measure of the willingness of the local Kurdish population to support insurgent violence. The literature on civil wars and insurgencies has suggested that insurgent activity requires cooperation of the local population for various reasons such as providing shelter, financial sources or simply not denouncing them.27 However, it can also capture the level of civilian support for the government in these provinces, and the insurgents may have been launching attacks in places where there is high civilian support for the government in order to intimidate the civilians and win them over to their side. The empirical analysis would be a proper test of my hypotheses if the former interpretation is correct rather than the latter. In Turkish parliamentary elections, there is a 10 percent electoral threshold so that candidates of parties that receive less than 10 percent of the national vote are not represented in the parliament no matter what share of votes the candidate receives in his or her electoral district. Hence, the parties that represent the Kurds of this region have been left out of the parliament during the period I cover in my analysis. Because of this, the turnout rate in parliamentary elections seems to be the natural choice to proxy government support in these provinces. If the dependent variable Conflict captures civilian support for the insurgents as a result of which they are better able to carry out attacks, we should expect the coefficient estimate for turnout to be negative, since this implies that insurgents are more likely to attack in places where government support is low. If Conflict captures civilian support for the government, then we should expect the coefficient estimate for turnout to be positive.

26See Figure 1 in Appendix 27See e.g. Fearon and Latin (2003); Justino (2009); Kalyvas (2007).

14 Finally, to cite the rest of the data sources, the data for GDP and yearly population ﬁgures come from the Turkish Statistical Institute. The data for turnout rates come from an online source that records election results in Turkey (www.belgenet.net). I consulted several scholarly works28 on the Kurdish people to code the variable Kurdish Majority.

Analysis

In order to analyze the determinants of the likelihood of insurgent attacks, I estimate the following logit model of conﬂict onset:

P (Conflicti,t = 1|Conflicti,t−1 = 0) = G(β1GDP/capRatioi,t−1 + β2Log(P ublicInvestment(cap))i,t−1+

+β3(GDP/capRatio ∗ Log(P ublicInvestment(cap)))i,t−1 + δXi,t−1 + i,t)

where GDP/cap Ratio measures GDP/cap in each province as a percentage of the national average, P ublic Investment is the annual public expenditure per capita in the 10 sectors I listed above, and X is a vector of controls that include log of population, turnout rate, and dummies for whether the province shares a border with another country, whether more than 50 percent of the population is Kurdish and whether the observation is from post-1992 period.

I exclude from my sample the years which are preceded by a year of conﬂict for two reasons.

First, my theoretical argument is about how the prospect of violence and the subsequent loss of public goods in a particular region affects civilian support for insurgents in that region prior to the onset of violence. Hence, my argument applies to the determinants of conflict onset, not to conflict incidence. Second, a potential source of bias in estimating this model is endogeneity due to reverse causation. I presented qualitative and quantitative evidence on the adverse effects of insurgent violence on different types of public goods provision such as health services and education in provinces where the Kurdish insurgency has been active. Therefore, it is reasonable to expect that GDP per capita and investment levels at the province level are negatively affected because of insurgent attacks. This bias should be more severe for the years

28Izady (1992), McDowall (1996), and Gunter (1990) all have very similar ﬁgures for the geographic distribution of Kurdish population.

15 in which there is an ongoing conflict in a particular province than the years before the start of the conflict. Therefore, I exclude the years which are preceded by a year of conflict to mitigate this potential bias due to reverse causation. In addition, I also use the independent variables with one year lag.

Another major concern in estimating this model is that the likelihood of having an attack may depend on the number of prior conflicts in a province. In order to control for this, X also includes a variable that counts the number of previous conflicts. Finally, the estimates may also suffer from duration dependence: The number of years passed without conflict may also have a significant effect on the likelihood of conflict onset in a province. Therefore, I also estimate the model by including a variable which counts the number of years passed since the last time there has been conflict in a province, and two natural cubic spline functions of this variable.29

In light of my analysis above, my theoretical expectations are the following: First, βˆ1 < 0 because poorer provinces should be more likely to support insurgents if no or very few public goods are provided. Second, βˆ2 < 0 because provision of public goods should reduce the popular support for insurgents. Third, βˆ3 > 0 because if public goods are provided extensively, wealthier provinces should be more likely to support insurgents than poorer provinces.

Empirical Results

Table 3 presents the results of a logit analysis using the variable Conflict as the dependent variable for diﬀerent thresholds of yearly attacks. Models 1, 3, 5, 7 and 9 use 1, 2, 3, 4 and

5 yearly attacks as the threshold for the onset of conflict, respectively. The even-numbered models also control for duration dependence by adding the variables that measure the number of years preceding onset of conflict and two natural cubic splines. First, if the insurgents target provinces in which they are not supported, electoral turnout should be positively related with the likelihood of attacks. However, the results show that this is not the case: In all specifications, the variable of T urnout has a negative sign, and it has a significant effect in several specifications.

This implies that insurgents are more likely to attack in provinces in which support for the government is low, and therefore the number of attacks at the province level can be interpreted

29I decide on the number of splines based on Wald tests (Beck et al. 1998). The results remain almost identical with three cubic splines.

16 as a proxy for the support of the civilians living in these provinces for the insurgents.

The extant research on the relationship between the distribution of PKK attacks and the level of civilian support for the insurgency also confirms this interpretation. For instance, using election data from 1999, 2004 and 2007, Gergin (2012) finds that more popular support for the insurgents (measured as the vote share of the political party affiliated with PKK) is associated with more insurgent attacks.

To move to the interpretation of the results for my independent variables of interest, first, it should be noted that the coefficient estimates have the expected sign with all definitions of conflict onset. GDP/cap Ratio and Log(Investment(cap)) have both negative coefficient estimates, while the coefficient estimate for the interaction term is positive in all specifications.

Moreover, my sample size is relatively small and there is a high degree of correlation between the interaction term and the variables that constitute it. Both of these factors increase variance of the estimator of a coefficient (Woolridge, 2003). Despite these potential effects making it harder to get significance, the interaction term is positive and significant in specifications which define conflict onset as the presence of 3 or more and 4 or more attacks in a year. Hence, the results are consistent with my hypothesis that conditional on high levels of public goods provision, the effect of GDP/cap on conflict onset is positive at the province-level.

Log(Investment(cap)) has a negative coefficient estimate, and it is also significant in several specifications. This result indicates that insurgent attacks are less likely in provinces with higher public goods provision, and it provides further support to my interpretation of the attacks as a proxy for civilian support for the insurgents. If attacks are instead a proxy for support for the government, this finding implies that higher public goods provision is associated with lower support for the government. This is a very unlikely scenario given the fact that these provinces are poor relative to the rest of the country and have a lot to benefit from public goods provision.

To move to the rest of the results, the effect of the variable Kurdish Majority is positive and significant in most of the specifications. Provinces with majority Kurdish populations are more likely to have attacks, which testifies to the fact that PKK is an ethnic insurgency, and it is supported by individuals with identity-based grievances. This finding also corroborates my interpretation of the presence of insurgent attacks as a proxy for civilian support for the

17 insurgents. Support for Kurdish insurgents should be higher in these Kurdish majority provinces relative to ethnically more mixed provinces. Finally, results with most of the speciﬁcations indicate that provinces that are less populated have a higher chance of insurgent attacks.

Table 4 shows the results of a logit analysis with the alternative definition of conflict onset as the presence of an attack with 10 or more civilian fatalities in a province. The first two models include all conflict onsets while the latter two only focus on the first onset. First, it must be noted that the variable T urnout has again a negative sign in all specifications. This implies that large scale attacks (attacks with at least 10 fatalities) have been more likely in provinces where the support for the government is lower. Therefore, these large scale attacks can also be interpreted as a proxy for civilian support for the insurgency. Second, the results with respect to my variables of interest are stronger: The coefficient estimates have the expected sign, and the coefficient estimate of the interaction term is positive and significant in all specifications.

Therefore, in line with the results that focus on all attacks, when I focus on large scale attacks, the conclusion is the same: When investment in public goods is low, attacks have been more likely in poorer provinces. However, when investment in public goods is high, insurgent attacks have been less likely in poorer provinces.

My explanation for these results is that when investment level in public goods is high, the prospect of a reduction in investment makes poorer individuals less likely to support insurgents than wealthier individuals. If this explanation is correct, then I should get similar results if I test my hypotheses with public investment data in specific sectors that affect the welfare of poor individuals more directly relative to other sectors. For instance, construction of public schools and hospitals should be more beneficial to the poor individuals living in the region than the construction of a large dam or the renewal of an oil refinery. Similarly, the provision of drinking water or the construction of an irrigation system should be more helpful to poor individuals living in rural areas than the construction of a road to a touristic site or the construction of public housing for the police or the gendarmerie. Therefore, in light of the projects covered under each sector, reductions in investment in education, health, infrastructure, and agriculture should directly increase the opportunity cost of supporting insurgents while reductions in the remaining sectors of energy, mining, tourism, manufacturing, transportation, and housing30

30As I report in the Appendix, housing expenditures cover construction of public housing for police, gendarmerie, judges, prosecutors and civil servants. These are not public housing projects for the civilians living in

18 should not have the same eﬀect.

Hence, as an alternative test of my theoretical argument, I also estimate the same econometric model using the public investment data on each of the 10 sectors separately. In light of the discussion above, I expect empirical tests with education, health, infrastructure and agriculture to show similar results to the tests with the total investment data, whereas the tests that focus on the remaining sectors should fail to do so. I estimate the model using 1 attack, and 3 and 5 yearly attacks per year as the threshold for the onset of conﬂict. I present the results in

Tables 5-7 in the Appendix. Interestingly, the coefficients have the expected sign for all three definitions of the dependent variable only for the sectors of education, health, infrastructure and agriculture: In each case, the coefficient estimate of the interaction effect is positive while the coefficient estimates of GDP/cap Ratio and Log(Investment(cap)) are negative. In tests that use 1 attack as the threshold for the onset of conflict, the interaction effect is significant at 10 percent for the model that use data on investment in education. For the tests that use 3 attacks as the threshold for the onset of conflict, the effects are statistically significant with education and infrastructure. In tests that use 5 attacks as the threshold, the effects are significant with public investment in education and health.

In order to see if these statistically significant results translate into substantively mean- ingful effects (Brambor et al. 2005), Figure 3 plots the estimated average marginal effect of

GDP/cap on the probability of conflict onset across all investment values in the sample for education, health and infrastructure. I report marginal effects based on the coefficient estimates of the model that uses 3 yearly attacks as the threshold for conflict onset.31 My hypothesis is that when no or very few public goods are provided, poorer provinces are more likely to have insurgent attacks. However, when there is high public goods provision, poorer provinces should be less likely to have insurgent attacks. Therefore, in each of these sectors, I expect the effect of an increase in GDP/cap Ratio to be negative for low levels of public investment and I expect this effect to be positive for high levels of public investment. This also means that the line depicting the estimated effect across different values of investment should have a positive slope.

The plot for education provides strong support for my hypothesis: When public investment the area. 31The plots for models that use 1 attack and 5 yearly attacks as the threshold are substantively very similar to the plots I present here.

19 in education is low, conflict onset is less likely in wealthier provinces. However, when public investment in education is high, conflict onset is more likely in wealthier provinces. The plot for health shows that the marginal effect of income is indistinguishable from zero for low levels of public investment in health. However, when public investment in health is high, wealthier provinces are more likely to have conflict, which is consistent with my hypothesis. Similarly for investment in infrastructure, the marginal effect of income is indistinguishable from zero for low levels of public investment in infrastructure. However, as the level of investment in infrastructure increases, the effect becomes positive and significant. This means that when investment in infrastructure is high, wealthier provinces are more likely to have conflict, which is also consistent with my hypothesis.

In sum, the estimated marginal effects support my hypotheses, and these effects are also consistent across the three different sectors of public investment. More specifically, lower GDP per capita makes conflict onset less likely when initial public investment levels in education, health or infrastructure are high.

District-Level Tests: Unemployment and Attacks

One remaining concern regarding the credibility of the results in the previous section is whether the positive interaction effect between GDP/cap and public investment captures the fact that violence in these provinces are motivated only by economic gain. The theoretical literature on civil wars and political violence has shown that greater national wealth can increase the efforts devoted to fighting over the capture of economic resources.32 Recently, scholars have documented empirically the positive effect of external aid and price increases in capital intensive goods on violence.33 Hence, there are both theoretical and empirical reasons to suspect that among two provinces with high public investment, attacks may be more likely in the wealthier province than in the poorer province simply because there is a bigger economic pie to fight for in the former relative to the latter. Hence, the killings may be motivated by local competition over the control of economic activity between groups that favor and oppose the insurgents, and the insurgents may target the rivals of their civilian supporters in this competition.

32See e.g. Garﬁnkel and Skaperdas (2007); Grossman (1999). 33See e.g. Nunn and Qian, (2012) and Dube and Vargas, (2013).

20 If the attacks in provinces with high public investment are indeed motivated by local competition over the control of economic activity, then we should observe a negative relation between individual measures of opportunity cost of violence and the likelihood of attacks in these provinces. This is because individuals with lower opportunity costs of violence should have stronger incentives to resort to violence for control of economic activity (Collier and Hoeﬄer,

1998, 2004). Hence, measures of individual economic well-being in this region would help whether we can rule out this alternative explanation. Unemployment data is one such measure.

If poorer provinces are less likely to have insurgent attacks under high levels of public goods provision because there is less to ﬁght for in these provinces relative to richer ones, then we should observe a positive relation between higher unemployment and violence in provinces with high public goods provision. However, as I argue, if poorer provinces with high public goods provision are less likely to have insurgent attacks because poorer individuals are more vulnerable to losing public goods and thus less likely to support insurgents, then higher unemployment should be negatively associated with violence when public goods provision is high. This is because places with higher unemployment have more individuals vulnerable to losing public goods. Therefore, the observable implications of my argument and of the explanation based on economic gain are opposite with respect to the link between unemployment and insurgent attacks.

In light of the above discussion, I present empirical results using unemployment data at the district level. Districts are the second-level administrative divisions within provinces in

Turkey. Unfortunately, the district-level unemployment data are available only for census years of 1985 and 1990.34 Therefore, I took unemployment ﬁgures from the censuses in 1985 and 1990 for each district from 20 provinces in the sample, and calculated the changes in unemployment rate for each of these districts from 1985 to 1990. In order to test the relation between changes in unemployment and insurgent activity in the region, I also calculated the total number of attacks per 1000 people in each district for two 5-year periods, 1985-89 and 1990-94, and then took the diﬀerence between the two.

To analyze the relation between unemployment and insurgent violence at the district-level,

I estimate the parameters of the following equation:

34The next census after 1990 was conducted in 2000.

21 ∆ Attacks/capi = β ∆ Unemploymenti + δ Xi + i

where i denotes one of 135 districts in 20 provinces.35 ∆Attacks/cap is the diﬀerence in the total number of attacks per 1000 people between the ﬁve-year periods of 1985-89 and 1990-

94, ∆Unemployment is the change in unemployment rate in a district from 1985 to 1990, and

Xi includes the change in the number of high school graduates per capita from 1985 to 1990, the change in turnout rate between parliamentary elections in 1987 and 1991, and dummies for whether the district became a central district after becoming part of a newly formed province, and whether the district lost areas to a newly formed district between 1985 and 1994.

To test my theoretical claim, I also need data on public goods provision at the district-level.

Since district-level data for public investment is not available, I employ the following strategy:

There is an ongoing multi-sector regional development program in the region called ”South- eastern Anatolia Project” (S.A.P.). The projects under S.A.P. cover several sectors including irrigation, hydraulic energy, agriculture, rural and urban infrastructure, forestry, education and health. 7 out of 20 provinces in my sample are covered by S.A.P. while the rest are not. There- fore, I use membership in S.A.P. as a proxy for high public goods provision, and divided the sample of districts into those that are part of S.A.P. and those that are not. If my hypotheses are correct, we should expect to see a negative relation between changes in unemployment and changes in the rate of attacks in districts that are covered by S.A.P. but not in districts excluded from S.A.P. If the explanation based on economic gain is correct, then we should expect to see a positive relation between changes in unemployment and changes in the rate of attacks both in districts covered by S.A.P. and in districts excluded from S.A.P.

Table 9 presents the results of empirical tests for both of these samples. In Panel 1, among the S.A.P. districts, the relation is negative and signiﬁcant in all speciﬁcations: Total number of attacks per capita in 1990-94 period has decreased from its 1985-89 level in districts in which unemployment has increased from 1985 to 1990. In contrast to the results in Panel 1, in Panel

3533 districts that are created out of the existing districts after 1985 are left out of the analysis because of missing data. See the Appendix for the list of districts included in the analysis. Summary statistics of the sample for district-level tests are in Table 8 in the Appendix.

22 2, among the non-S.A.P. districts, the relation is still negative but it is not significant in any of the specifications. Hence, the first-difference estimates of the effect of unemployment on the rate of insurgent attacks provide further support to my theoretical argument: Among districts with high public investment, attacks have decreased in districts where a greater share of the population has become unemployed and thus vulnerable to losing public goods. Among districts with low public investment, there is no statistically significant relationship between changes in the number of attacks and unemployment rate. Moreover, echoing earlier results, the negative relationship between unemployment and insurgent attacks is stronger among Kurdish-majority districts. These results further suggest that the insurgency is primarily motivated by ethnic grievances, but economic factors also have a significant effect on the likelihood of insurgent attacks.

Conclusion

In this paper, I present and test an argument that links income and public goods provision to insurgent violence in ethnic conflict. More specifically, I argue that if insurgent violence leads to a reduction in public goods provision, then extensive provision of public goods prior to the outbreak of violence makes poorer populations less likely to assist militants of their own ethnic group. This is because poorer individuals’ opportunity cost for involvement in an ethnic insurgency is higher than wealthier individuals due to their greater vulnerability to losing public goods. Since assistance from non-combatants is vital for the insurgents’ ability to fight,

I hypothesize that we should observe less violence where the local population is poorer, and where violence can lead to a signiﬁcant reduction in public goods provision.

I test this hypothesis by focusing on the case of the Kurdish insurgency in Turkey and find empirical support for it: In poor provinces with high levels of public investment, lower GDP per capita decreases the risk of insurgent attacks against civilians. This relationship is especially strong for investment in public goods that are most relevant for the welfare of poor civilians, such as health services, education, and infrastructure. I support this finding by analyzing the link between insurgent violence and a direct measure of individual vulnerability to the loss of public goods. I find that increases in unemployment rate are also associated with lower number

23 of attacks in districts with high levels of public goods provision.

These results suggest that distinguishing between alternative means of economic improve- ment for potential supporters of insurgents (income, public goods, employment) and focusing on the interactions among each of these provides a better understanding of the economic incentives involved in civilians’ decisions to support ethnic insurgencies. A potential avenue for further research is to study these interactions in a cross-national context. Another possibility is to collect data on the distribution of the killings of members of the army and police, and study whether the same relation between income, public investment and violence also extends to the killings of state security forces.

24 References

Abadie, Alberto, and Javier Gardeazabal. 2003. ‘The Economic Costs of Conﬂict: A Case

Study of the Basque Country.’ American Economic Review, 93(1): 113-32.

Barkey, Henri J., and Graham Fuller. 1998. Turkey’s Kurdish Question. Lanman, MD:

Rowman and Littleﬁeld Publishers.

Barkey Henri. 2007. “Turkey and the PKK: A Pyrrhic Victory?” In Democracy and

Counterterrorism: Lessons from the Past, eds. Robert Art and Louise Richardson. Washington,

DC: United States Institute of Peace, 343-81.

Barron, Patrick, Kai Kaiser, and Menno Pradhan. 2004. “Local Conﬂict in Indonesia:

Measuring Incidence and Identifying Patterns.” World Bank Policy Research Working Paper:

August, 3384.

Beath Andrew, Fotini Christia, and Ruben Enikolopov. 2011. “Winning Hearts and Minds

Through Development: Evidence from a Field Experiment in Afghanistan.” Massachusetts

Institute of Technology Political Science Department Working Paper: 2011-14.

Beck, Nathaniel, Jonathan N. Katz, and Richard Tucker. 1998. “Taking Time Seriously:

Time-Series-Cross Section Analysis with a Binary Dependent Variable.” American Journal of

Political Science 42(4): 1260-88.

Benmelech, Efraim, Claude Berrebi, and Esteban Klor. 2010. “The Economic Cost of

Harboring Terrorism.” Journal of Conﬂict Resolution 54(2): 331-353.

Berman, Eli, Michael Callen, Joseph H. Felter, and Jacob N. Shapiro. 2011a. “Do Working

Men Rebel? Insurgency and Unemployment in Iraq, Afghanistan and Philippines”. Journal of

Conﬂict Resolution. 55 (4): 496-528.

Berman, Eli, Joseph H. Felter, and Jacob N. Shapiro. 2011b. “Can Hearts and Minds be

Bought? The Economics of Counterinsurgency in Iraq.” Journal of Political Economy 119 (4):

766-819.

Blair, Graeme, C. Christine Fair, Neil Malhotra, and Jacob N. Shapiro. 2012. “Poverty and Support for Militant Politics: Evidence from Pakistan. ” American Journal of Political

Science. 57 (1): 30-48.

25 Blattman Chris and Edward Miguel. 2010. “Civil War.” Journal of Economic Literature

48:1, 3-57.

Brambor, Thomas, William Roberts Clark, and Matt Golder. 2005. “Understanding

Interaction Models: Improving Empirical Analyses.” Political Analysis. 13: 1-20.

Bruck, Tilman. 2001. Macroeconomic Eﬀects of the War in Mozambique. University of

Oxford Queen Elizabeth Working Paper Series 11.

Buckland, Peter. 2005. Reshaping the Future: Education and Post-Conﬂict Reconstruc- tion. Washington, DC: World Bank.

Cederman, Lars-Erik, Kristian Skrede Gleditsch, and Nils B. Weidmann. 2011. “Horizon- tal Inequalities and Ethnonationalist Civil War: A Global Comparison.” American Political

Science Review 105(3): 478-95.

Collier, David and Anke Hoeﬄer. 1998. “On Economic Causes of Civil War.” Oxford

Economic Papers 50 (4): 563-73.

Collier, David and Anke Hoeﬄer. 2004. “Greed and Grievance in Civil War”. Oxford

Economic Papers 56: 563-595.

Crost, Benjamin, Joseph Felter, and Patrick Johnston. 2014. “Aid under Fire: Develop- ment Projects and Civil Conﬂict.” American Economic Review, 104(6): 1833-56.

Derin-Gure, Pinar. 2011.“Separatist Terrorism and The Economic Conditions in South-

Eastern Turkey”. Defence and Peace Economics 22 (4): 393-407.

Derin-Gure, Pinar, and Adem Yavuz Elveren. 2014. “Does Income Inequality Derive The

Separate Terrorism in Turkey?”, Defence and Peace Economics 25 (3): 311-327.

Do, Quy-Toan, and Lakshmi Iyer. 2010. “Geography, Poverty and Conﬂict in Nepal.”

Journal of Peace Research 47(6): 735-48.

Dube, Oeindrila, and Juan F. Vargas. 2013. “Commodity Price Shocks and Civil Conﬂict:

Evidence from Colombia.” Review of Economic Studies, 80: 1384-1421.

Fearon, James and David Laitin. 2003.“Ethnicity, Insurgency and Civil War”. American

Political Science Review. 97(1): 75-90.

Fearon, James. 2004. Why Do Some Civil Wars Last So Much Longer Than Others?

26 Journal of Peace Research 41(3): 275-301.

Feridun, Mete and Selami Sezgin. 2008. “Regional Underdevelopment and Terrorism: The

Case of South Eastern Turkey”. Defence and Peace Economics, 19 (3): 225-233.

Franzese, Robert J. JR., and Cindy D. Kam. 2007. Modeling and Interpreting Interactive

Hypotheses In Regression Analysis. Ann Arbor: University of Michigan Press.

Galula, David. 1964. Counterinsurgency Warfare: Theory and Practice. NewYork:

Praeger.

Garﬁnkel, Michelle, and Stergios Skaperdas. 2007. “Economics of Conﬂict: An Overview”

In Handbook of Defense Economics, Volume 2, Defense in a Globalized World, eds. Todd

Sandler and Keith Hardley. Amsterdam and Oxford: Elsevier and North Holland, 649-710.

Gergin, Nadir . 2012. “The Nexus between the Ballot and Bullet: Popular Support for the PKK and Post-election Violence in Turkey”. Ph. D. Thesis. Virginia Commonwealth

University.

Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg, and

Havard Strand. 2002. “Armed Conﬂict 1946-2001: A New Dataset”. Journal of Peace Research,

39(5): 615-37.

Grossman, Herschel. 1999. “Kleptocracy and Revolutions.” Oxford Economic Papers,

51(2): 267-83.

Gunter, Michael. 1990. The Kurds In Turkey: A Political Dilemma. Boulder: Westview

Press.

Gurr, Ted. 1970. Why Men Rebel. Princeton, NJ: Princeton University Press.

Hegre, Havard, and Nicholas Sambanis. 2006. “Sensitivity Analysis of Empirical Results on Civil War Onset.” Journal of Con!ict Resolution, 50(4): 508-35.

Hegre Havard, Gudrun Ostby and Clionadh Raleigh. 2009. “Poverty and Civil War Events:

A Disaggregated Study of Liberia”. Journal of Conﬂict Resolution 53: 4, 598-623.

Households and Conﬂict Network. http://www.hicn.org/wordpress/?page_id=28 (accessed September 8, 2012).

Human Rights Watch/Helsinki. 1993. Kurds of Turkey: Killlings, Disappearances and

27 Torture. http://www.hrw.org/sites/default/files/reports/TURKEY933.PDF (accessed, July 24, 2012)

Human Rights Watch. 1994. Forced Displacement of Ethnic Kurds from Southeast- ern Turkey. http://www.hrw.org/sites/default/files/reports/TURKEY94O.PDF (accessed, July 24, 2012)

Human Rights Watch. 2002. Displaced and Disregarded: Turkey’s Failing Village Return

Program. http://www.hrw.org/reports/2002/turkey/Turkey1002.pdf (accessed, July 24, 2012)

Human Rights Watch. 2005. Still Critical: Prospects in 2005 For Internally Displaced

Kurds in Turkey http://www.hrw.org/sites/default/files/reports/turkey0305.pdf (accessed, July 24, 2012)

Humphreys, Macartan. 2003. “Economics and Violent Conﬂict” Columbia University.

Unpublished Manuscript.

Humphreys, Macartan and Jeremy M. Weinstein. 2008 “Who Fights? The Determinants of Participation in Civil War.” American Journal of Political Science 52 (2): 436-55

Izady, Mehrdad. 1992. The Kurds: A Concise Handbook. Washington, DC: Taylor and

Francis.

Justino, Patricia, and Philip Verwimp. 2006. “Poverty Dynamics, Violent Conﬂict and

Convergence in Rwanda.” Households in Conﬂict Network Working Paper 16.

Justino, Patricia. 2009. “Poverty and Violent Conﬂict: A Micro-Level Perspective on the

Causes and Duration of Warfare.” Journal of Peace Research 46: 315-33.

Kalyvas, Stathis. 2006. The Logic of Violence in Civil Wars. New York: Cambridge

University Press.

Kalyvas, Stathis. 2007. “Civil Wars.” In Handbook of Political Science, eds. Carles Boix and Susan Stokes. New York: Oxford University Press, 416-434.

Kalyvas, Stathis. 2008. Promises and Pitfalls of an Emerging Research Program: The

Microdynamics of Civil War. In Stathis N. Kalyvas, Ian Shapiro and Tarek Masoud (eds),

Order, Conﬂict, Violence. Cambridge University Press, 1-14.

28 Kapmaz, Cengiz. 2011. Ocalan’in Imrali Gunleri. Istanbul: Ithaki.

Kirisci Kemal and Gareth Winrow. 1997. The Kurdish Question and Turkey: An Example of a Trans-state Ethnic Conﬂict. London: Frank Cass and Co.

Lichbach, Mark Irving. 1989. “An Evaluation of ‘Does Economic Inequality Breed Political

Conﬂict?’ Studies”. World Politics, 41(4): 431-70.

Marcus Aliza 2007. Blood and Belief: the PKK and the Kurdish Fight for Independence

New York: New York University Press.

McDowall, David. 1996. A Modern History of the Kurds. London: Tauris.

Muller, Edward N., and Mitchell Seligson. 1987. “Inequality and Insurgency”. American

Political Science Review 81(2): 425-52.

Murshed, S. Mansoob, and Scott Gates. 2005. “Spatial-Horizontal Inequality and the

Maoist Insurgency in Nepal.” Review of Development Economics 9(1): 121-34.

Nunn, Nathan, and Nancy Qian. 2012. “Aiding Conﬂict: The Impact of US Food Aid on

Civil War.” Harvard University. Unpublished Manuscript.

Ozdag, Umit. 2009. Pusu ve Katliamlarin Kronolojisi: PKK’nin Gerceklestirdigi Toplu

Katliamlar. Ankara: Kripto.

Petersen, Roger. 2001. Resistance and Rebellion: Lessons from Eastern Europe. Cam- bridge: Cambridge University Press.

Petraus, David. 2006. “Learning Counterinsurgency: Observations from Soldiering in

Iraq.” Military Review January-February, 45-55.

Sambanis, Nicolas. 2004. “Poverty and the Organization of Political Violence.” Brookings

Trade Forum 2004: 165-211.

Sener, Nedim. 2010. “26 Yilin Kanli Bilancosu.” Milliyet, June 24, 2010, http://www. milliyet.com.tr/26-yilin-kanli-bilancosu/guncel/haberdetay/24.06.2010/1254711/default. htm (accessed, September 8, 2012)

Sepp, Kalev. 2005. “‘Best Practices’ in Counterinsurgency.” Military Review May-June,

8-12.

29 Stewart, Frances. ed. 2008. Horizontal inequalities and Conﬂict: Understanding Group

Violence in Multiethnic Societies. Houndmills, UK: Palgrave Macmillan.

Taber, Robert. 1965. The War of the Flea: A Study of Guerilla Warfare Theory and

Practice. New York: Lyle Stuart.

Turkish Parliament. 1998. Dogu ve Guneydogu Anadolu’da Bosaltilan Yerlesim Birimleri

Nedeniyle Goc Eden Yurttaslarimizin Sorunlarinin Arastirilarak Alinmasi Gereken Tedbirlerin

Tespit Edilmesi Amaciyla Kurulan Meclis Arastirma Komisyonu Raporu. http://www.tbmm. gov.tr/sirasayi/donem20/yil01/ss532.pdf (accessed, June 13, 2012)

U.S. Army. 2007. Field Manual 3-24, Counterinsurgency Field Manual. Chicago: Univer- sity of Chicago Press.

Van Bruinessen, Martin. 1992. Agha, Shaikh and State: The Social and Political Structures of Kurdistan. New Jersey: Zed Books.

Van Bruinessen, Martin. 2000. Kurdish Ethno-nationalism versus nation-building states: collected articles. Istanbul: Isis Press.

Weinstein, Jeremy. 2007. Inside Rebellion: The Politics of Insurgent Violence. Cam- bridge: Cambridge University Press.

Wickham-Crowley, Timothy. 1992. Guerillas and Revolution in Latin America: A Com- parative Study of Insurgents and Regimes since 1956. Princeton, NJ: Princeton University

Press.

Woolridge, Jeﬀrey. 2002. Introductory Econometrics: A Modern Approach. 2nd ed.

South-Western College Publications.

Yildirim, Julide, and Nadir Ocal. 2013. “Analysing the Determinants of Terrorism in

Turkey Using Geographically Weighted Regression”. Defence and Peace Economics, 24 (3):

195-209.

30 Appendix

List of Provinces

ID Province

1 Adiyaman

2 Agri

3 Ardahan

4 Batman

5 Bingol

6 Bitlis

7 Diyarbakir

8 Elazig

9 Erzincan

10 Hakkari

11 Igdir

12 Kars

13 Malatya

14 Mardin

15 Mus

16 Sanliurfa

17 Siirt

18 Sirnak

19 Tunceli

20 Van

31 List of Provinces and Districts

ID Province District ID Province District

1 Adiyaman Merkez 2 Adiyaman Besni

3 Adiyaman Celikhan 4 Adiyaman Gerger

5 Adiyaman Golbasi 6 Adiyaman Kahta

7 Adiyaman Samsat 8 Agri AgriMerkez

9 Agri Diyadin 10 Agri Dogubeyazit

11 Agri Eleskirt 12 Agri Hamur

13 Agri Patnos 14 Agri Taslicay

15 Agri Tutak 16 Batman Batman

17 Batman Gercus 18 Batman Kozluk

19 Batman Sason 20 Bingol Bingol

21 Bingol Genc 22 Bingol Karliova

23 Bingol Kigi 24 Bingol Solhan

25 Bitlis Adilcevaz 26 Bitlis Ahlat

27 Bitlis Hizan 28 Bitlis Mutki

29 Bitlis Tatvan 30 Diyarbakir Bismil

31 Diyarbakir Cermik 32 Diyarbakir Cinar

33 Diyarbakir Cungus 34 Diyarbakir Dicle

35 Diyarbakir Diyarbakir 36 Diyarbakir Ergani

37 Diyarbakir Hani 38 Diyarbakir Hazro

39 Diyarbakir Kulp 40 Diyarbakir Lice

41 Diyarbakir Silvan 42 Elazig Agin

43 Elazig Baskil 44 Elazig Merkez

45 Elazig Karakocan 46 Elazig Keban

47 Elazig Maden 48 Elazig Palu

49 Elazig Sivrice 50 Erzincan Cayirli

51 Erzincan Merkez 52 Erzincan Ilic

53 Erzincan Kemah 54 Erzincan Kemaliye

55 Erzincan Refahiye 56 Erzincan Tercan

32 ID Province District ID Province District

57 Hakkari Cukurca 58 Hakkari Merkez

59 Hakkari Semdinli 60 Hakkari Yuksekova

61 Kars Aralik 62 Kars Ardahan

63 Kars Arpacay 64 Kars Cildir

65 Kars Digor 66 Kars Gole

67 Kars Hanak 68 Kars Igdir

69 Kars Kagizman 70 Kars Merkez

71 Kars Posof 72 Kars Sarikamis

73 Kars Selim 74 Kars Susuz

75 Kars Tuzluca 76 Malatya Akcadag

77 Malatya Arapkir 78 Malatya Arguvan

79 Malatya Darende 80 Malatya Dogansehir

81 Malatya Hekimhan 82 Malatya Merkez

83 Malatya Poturge 84 Malatya Yesilyurt

85 Mardin Derik 86 Mardin Kiziltepe

87 Mardin Merkez 88 Mardin Mazidagi

89 Mardin Midyat 90 Mardin Nusaybin

91 Mardin Omerli 92 Mardin Savur

93 Mus Bulanik 94 Mus Malazgirt

95 Mus Merkez 96 Mus Varto

97 Sanliurfa Akcakale 98 Sanliurfa Birecik

99 Sanliurfa Bozova 100 Sanliurfa Ceylanpinar

101 Sanliurfa Halfeti 102 Sanliurfa Hilvan

103 Sanliurfa Merkez 104 Sanliurfa Siverek

105 Sanliurfa Suruc 106 Sanliurfa Viransehir

107 Siirt Baykan 108 Siirt Besiri

109 Siirt Eruh 110 Siirt Kurtalan

111 Siirt Pervari 112 Siirt Siirt

113 Siirt Sirvan 114 Sirnak Beytussebap

115 Sirnak Cizre 116 Sirnak Idil

33 ID Province District ID Province District

117 Sirnak Silopi 118 Sirnak Sirnak

119 Sirnak Uludere 120 Tunceli Cemisgezek

121 Tunceli Hozat 122 Tunceli Mazgirt

123 Tunceli Nazimiye 124 Tunceli Ovacik

125 Tunceli Pertek 126 Tunceli Pulumur

127 Tunceli Merkez 128 Van Baskale

129 Van Catak 130 Van Ercis

131 Van Gevas 132 Van Gurpinar

133 Van Muradiye 134 Van Ozalp

135 Van Merkez

34 Public Investment Projects in Each Sector36

Education: School building projects at the primary, secondary and tertiary levels.

Health: Building of new public hospitals, health centers, dispensaries, and daycare.

Infrastructure:37 Projects for drinking water, sanitation, and sewage.

Housing: Public housing for police, gendarmerie, judges, prosecutors and civil servants working in the area.

Manufacturing: Construction of state-owned factories and storehouses for several agricultural products such as sugar and tobacco as well as mechanical products such as engines and generators.

Transportation: Construction of roads and airports in the area.

Tourism: Construction of national parks and roads for improved access to touristic sites.

Mining: Projects for mineral exploration as well as the construction and renewal of mines and oil reﬁneries.

Agriculture: Irrigation projects included in the regional development program Southeast- ern Anatolia Project (S.A.P.)

Energy: Construction of hydroelectric dams as part of the regional development program

Southeastern Anatolia Project (S.A.P.)

36I report the projects that constitute the majority of expenditures in each category. The documents that report these expenditures in their entirety can be reached (in Turkish) at http://www2.kalkinma.gov.tr/kamuyat/ il-arsiv.html. 37The investments that I call infrastructure are summed under the category ”other” in the Turkish State Plan- ning Organization documents. However, I use the word ”infrastructure” because of the nature of the expenditures in this category.

35 Table 1: Public Goods and Economic Cost of Violence for Two Individuals Who Support Insurgents in a Country with the Per Capita Income of 10 dollars Income Cost of Violence Public Goods No Yes Individual m 5 2.5 7.5 Individual n 8 4 6 Note: The numbers are calculated for the case when ﬁghting leads to the destruction of half of each individual’s income.

36 Table 2: Summary statistics (Provinces) Variable Mean Std. Dev. Min. Max. N 1attack 0.40 0.49 0 1 224 2attacks 0.24 0.43 0 1 224 3attacks 0.17 0.38 0 1 224 4attacks 0.13 0.34 0 1 224 5attacks 0.08 0.27 0 1 224 attack10 0.11 0.31 0 1 224 GDP/cap Ratio 0.41 0.19 0.15 0.88 220 log(Investment (cap)) 20.13 1.99 16.48 26.30 218 Turnover 83.40 6.59 70.02 95.27 216 37 log (Population) 12.90 0.63 11.33 14.10 220 Kurdish Majority 0.51 0.50 0 1 224 Border 0.42 0.50 0 1 224 Post-1992 0.63 0.49 0 1 224 1attack is a binary variable equal to 1 if there is at least one attack in a province-year. 2attacks, 3attacks, 4attacks, 5attacks are binary variables that use 2,3,4,5 yearly attacks as the threshold for conﬂict. attack10 is a binary variable equal to 1 if there is at last one attack with 10 or more fatalities in a province-year. GDP/cap Ratio is the ratio of province GDP/cap to the national GDP/cap. log(Investment (cap)) is the log of public investment per capita in a province. Turnover is the turnout rate in parliamentary elections. Kurdish Majority is a binary variable that denotes whether Kurds are more than 50 percent of the population in a province. Border is a dummy that denotes whether the province is bordering a neighboring country. Post-1992 is a dummy that denotes observations after 1992. 38

Number of Attacks

0 5 10 15 0 5 10 15 0 5 10 15 0 5 10 15 1985 1985 1985 1985 1990 1990 1990 1990 Adiyaman iue1 ubro tak ih3o oeFtlte vrTm yPoic (1988-1999) Province by Time over Fatalities more or 3 with Attacks of Number 1: Figure Sanliurfa Bitlis Igdir 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Diyarbakir Kars Agri Siirt 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Year Ardahan Malatya Sirnak Elazig 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Erzincan Batman Tunceli Mardin 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Hakkari Bingol Mus Van 1995 1995 1995 1995 2000 2000 2000 2000 39

GDP/cap as a Percentage of National GDP/cap

.2 .4 .6 .8 .2 .4 .6 .8 .2 .4 .6 .8 .2 .4 .6 .8 1985 1985 1985 1985 iue2 rvnelvlGPprcpt sapretg ftentoa D e aia(1988-1999) capita per GDP national the of percentage a as capita per GDP Province-level 2: Figure 1990 1990 1990 1990 Adiyaman Sanliurfa Bitlis Igdir 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Diyarbakir Kars Agri Siirt 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Year Ardahan Malatya Sirnak Elazig 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Erzincan Batman Tunceli Mardin 1995 1995 1995 1995 2000 2000 2000 2000 1985 1985 1985 1985 1990 1990 1990 1990 Hakkari Bingol Mus Van 1995 1995 1995 1995 2000 2000 2000 2000 Table 3: Income Inequality and Insurgent Attacks (1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

GDP/cap Ratio -0.05 -0.09 -0.12 -0.12 -0.31 -0.34 -0.55** -0.75** -0.24 -0.20 (0.08) (0.08) (0.16) (0.16) (0.24) (0.22) (0.27) (0.31) (0.27) (0.32)

Log(Investment(cap)) -0.09 -0.14 -0.60** -0.64** -1.67*** -1.36*** -1.97*** -2.48* -1.50** -1.37** (0.43) (0.47) (0.29) (0.32) (0.51) (0.45) (0.73) (1.31) (0.66) (0.65)

GDP/capRatio∗ 0.00 0.01 0.01 0.01 0.02* 0.02** 0.03** 0.04** 0.02 0.01 Log(Investment(cap)) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.02) (0.01) (0.02)

Kurdish 1.41* 1.29 2.11*** 2.03*** 4.58** 4.73** 5.33 4.64 3.98* 3.04 Majority (0.80) (0.83) (0.68) (0.67) (1.92) (1.95) (3.41) (3.52) (2.13) (2.07)

T urnout -0.09* -0.10** -0.01 -0.03 -0.18* -0.08 -0.38*** -0.54** -0.22** -0.22** (0.05) (0.05) (0.08) (0.10) (0.10) (0.13) (0.12) (0.25) (0.10) (0.11) 40 Log(P opulation) -0.74 -0.62 -0.77 -0.72 -2.54* -2.64* -0.57 -0.62 -0.19 0.38 (0.62) (0.71) (0.78) (0.90) (1.34) (1.39) (1.15) (1.16) (0.70) (1.08)

Border 0.28 0.22 0.42 0.38 -0.53 -0.34 -3.80* -4.69*** -1.10 -1.58 (0.93) (1.04) (0.78) (0.78) (0.92) (0.94) (2.13) (1.81) (1.30) (1.25)

P ost − 1992 -0.89 -0.74 0.22 0.32 1.24 0.74 -0.31 2.61 0.19 4.12** (1.19) (1.18) (1.31) (1.32) (1.25) (1.15) (1.80) (2.41) (1.37) (2.00)

Numberof 0.01 -0.15 -0.13 -0.24 -1.03 -0.55 -1.74 -5.60* -0.70 -3.68*** P rior Conflicts (0.36) (0.35) (0.49) (0.69) (0.82) (0.80) (1.29) (3.02) (0.54) (1.15)

Y ears W ithout -0.31 0.68 -5.49* -0.80 -4.08*** Conflict (0.54) (3.13) (3.29) (2.70) (1.21) Observations 126 126 148 148 159 159 169 169 180 180 Province-clustered standard errors in parentheses.GDP/capRatio, Log(Investment(cap)), Log(P opulation), T urnout are lagged one year. Models 1, 3, 5, 7 and 9 use 1, 2, 3, 4 and 5 yearly attacks as the threshold for the incidence of conflict, respectively. The even-numbered models also include the variables that measure the number of years preceding onset of conflict and two natural cubic splines. ∗∗∗, ∗∗, ∗ indicate statistical significance at 1%, 5 % and 10 %, respectively. Table 4: Income Inequality and Insurgent Attacks (1) (2) (3) (4) GDP/cap Ratio -0.29 -0.34 -0.37* -0.59** (0.19) (0.21) (0.21) (0.25)

Log(Investment)(cap)) -1.58*** -1.64*** -1.44** -1.72** (0.55) (0.59) (0.61) (0.69)

GDP/cap Ratio∗ 0.02** 0.02* 0.02** 0.03*** Log(Investment(cap)) (0.01) (0.01) (0.01) (0.01)

Kurdish 2.48** 2.32** 2.81** 2.94** Majority (0.99) (1.00) (1.21) (1.41)

T urnout -0.18** -0.17** -0.14* -0.05 (0.08) (0.08) (0.08) (0.10)

Log(P opulation) -1.19* -0.93 -2.34** -2.60** (0.68) (0.71) (0.98) (1.24)

Border 1.21 1.22 0.73 1.14 (1.29) (1.27) (1.34) (1.45)

P ost − 1992 0.03 0.63 -0.03 -0.24 (1.00) (1.00) (1.34) (1.33)

Number of -0.36 -0.90 P rior Conflicts (0.97) (1.27)

Y ears W ithout 0.70 -0.06 Conflict (0.77) (0.82) Observations 187 187 126 126 Province-clustered standard errors in parentheses. GDP/capRatio, Log(Investment(cap)), Log(P opulation), T urnout are lagged one year. The dependent variable is whether there is an attack with 10 or more fatalities in a province-year. Models 3 and 4 use only the first onset of conflict in each province. Models 2 and 4 add the variables measuring the number of years preceding conflict and two natural cubic splines. ∗∗∗, ∗∗, ∗ indicate statistical significance at 1%, 5 % and 10 %, respectively.

41 Table 5: Income Inequality, Public Investment, and Conﬂict Onset (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Education Health Infrastr. Housing Manufact. Mining Energy Agriculture Transport. Tourism GDP/cap Ratio -0.14* -0.03 -0.15 -0.06* -0.07 0.05 0.04 -0.08 -0.02 0.04 (0.08) (0.03) (0.09) (0.03) (0.10) (0.05) (0.03) (0.05) (0.08) (0.05)

Log(Investment(cap)) -0.43 -0.32* -0.68 -0.45*** -0.16 0.06 0.11 -0.29 -0.23 0.02

42 (0.33) (0.18) (0.60) (0.17) (0.22) (0.20) (0.11) (0.21) (0.21) (0.12)

Interaction 0.01* 0.00 0.01 0.01* 0.01 -0.00 -0.00 0.01 0.00 -0.00 (0.01) (0.00) (0.01) (0.00) (0.01) (0.00) (0.00) (0.00) (0.00) (0.00) Observations 121 128 122 123 127 128 127 126 125 125 The dependent variable is equal to 1 if there is at least 1 attack in a province-year. GDP/capRatio, Log(Investment(cap)) are lagged one year. All models include the following control variables: log of population (lagged), turnout rate in parliamentary elections, dummies for Kurdish population being the majority in a province and a province being on the border, the number of preceding years without conflict and two natural cubic splines of this variable. Province-clustered standard errors are in parentheses. ∗∗∗, ∗∗, ∗ indicate statistical significance at 1%, 5 % and 10 %, respectively. Table 6: Income Inequality, Public Investment, and Conflict Onset (Three or More Attacks) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Education Health Infrastr. Housing Manufact. Mining Energy Agriculture Transport. Tourism GDP/cap Ratio -0.48** -0.07 -0.52** 0.01 -0.27*** 0.06 0.10 -0.14* 0.12** 0.04 (0.21) (0.12) (0.25) (0.07) (0.10) (0.06) (0.07) (0.08) (0.06) (0.06)

Log(Investment(cap)) -1.70** -0.60 -2.16*** -0.39* -0.65*** -0.17 0.11 -0.67** 0.09 -0.23

43 (0.70) (0.43) (0.79) (0.23) (0.24) (0.28) (0.13) (0.30) (0.16) (0.18)

Interaction 0.03** 0.01 0.03*** 0.00 0.02*** 0.00 -0.00 0.01* -0.00 0.00 (0.01) (0.01) (0.01) (0.00) (0.01) (0.00) (0.00) (0.01) (0.00) (0.00) Observations 151 160 153 155 159 161 158 159 156 158 The dependent variable is equal to 1 if there are at least 3 attacks in a province-year. GDP/capRatio, Log(Investment(cap)) are lagged one year. All models include the following control variables: log of population (lagged), turnout rate in the latest parliamentary elections, dummies for Kurdish population being the majority in a province and a province being on the border, the number of preceding years without conflict and two natural cubic splines of this variable. Province-clustered standard errors are in parentheses. ∗∗∗, ∗∗, ∗ indicate statistical significance at 1%, 5 % and 10 %, respectively. Table 7: Income Inequality, Public Investment, and Conflict Onset (Five or More Attacks) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Education Health Infrastr. Housing Manufact. Mining Energy Agriculture Transport. Tourism GDP/cap Ratio -0.57*** -0.20* -0.25 -0.14* 0.00 0.09 0.03 -0.06 0.14 0.01 (0.21) (0.12) (0.37) (0.07) (0.17) (0.09) (0.06) (0.09) (0.13) (0.06)

Log(Investment(cap)) -1.90*** -0.83* -1.48 -1.03*** -0.25 0.53 0.11 -0.15 0.18 0.05

44 (0.66) (0.50) (1.28) (0.24) (0.39) (0.37) (0.28) (0.31) (0.19) (0.27)

Interaction 0.03*** 0.02** 0.02 0.01 0.00 -0.01 -0.00 0.01 -0.01 0.00 (0.01) (0.01) (0.02) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Observations 171 180 173 175 180 182 179 180 177 179 The dependent variable is equal to 1 if there are at least 5 attacks in a province-year. GDP/capRatio, Log(Investment(cap)) are lagged one year. All models include the following control variables: log of population (lagged), turnout rate in the latest parliamentary elections, dummies for Kurdish population being the majority in a province and a province being on the border, the number of preceding years without conflict and two natural cubic splines of this variable. Province-clustered standard errors are in parentheses. ∗∗∗, ∗∗, ∗ indicate statistical significance at 1%, 5 % and 10 %, respectively. n nrsrcue oflc ne sdfie sa es tak napoic nayear. a in province a in attacks health 3 education, least probability in at capita the as per on defined investment GDP/cap is public higher onset logged Conflict of of effect values infrastructure. different marginal and average for onset the conflict display of plots The 3: Figure

Change in the Probability of Conflict Onset Change in the Probability of Conflict Onset Change in the Probability of Conflict Onset −.02 −.01 0 .01 .02 −.02 −.01 0 .01 .02 −.02 −.01 0 .01 .02 14 7 7 Marginal EffectofHigherProvince−LevelGDP/cap Marginal EffectofHigherProvince−LevelGDP/cap Marginal EffectofHigherProvince−LevelGDP/cap 15 9 9 Public InvestmentinInfrastructurepercapita(log) 16 Public InvestmentinEducationpercapita(log) Public InvestmentinHealthpercapita(log) 11 11 17 13 13 18 45 19 15 15 20 17 17 21 19 19 22 21 21 23 24 23 23 Table 8: Summary statistics (Districts) Variable Mean Std. Dev. Min. Max. N ∆ Attacks/cap 0.03 0.06 -0.21 0.31 135 ∆ Unemployment 1.36 2.27 -6.50 8.11 135 ∆ Education 0.01 0.01 -0.01 0.05 135 ∆ T urnout -11.63 5.02 -25.04 2.24 135 KurdishMajority 0.49 0.50 0 1 135 SAP 0.4 0.49 0 1 135 Lost87 0.10 0.31 0 1 135 Lost90 0.15 0.36 0 1 135 46 Lost92 0.02 0.12 0 1 135 Central 0.03 0.17 0 1 135 ∆ Attacks/cap is the diﬀerence in the total number of attacks per 1000 people in each district for two 5-year periods, 1985-89 and 1990-94. ∆ Unemployment is the change in the unemployment rate from 1985 to 1990. ∆ Education is the change in the number of high school graduates per capita from 1985 to 1990 ∆ T urnout is the change in turnout rate between parliamentary elections in 1987 and 1991. KurdishMajority is a binary variable that denotes whether Kurds are more than 50 percent of the population in a district. SAP is a binary variable that denotes membership in Southeastern Anatolia Project (S.A.P.) Lost87, Lost90 and Lost92 are binary variables that denote whether the district has lost areas to a newly formed district in 1987, 1990 and 1992, respectively. Central is a binary variable that denotes whether the district has become a central district between 1985-94. Table 9: Attacks and Unemployment (First-Diﬀerence Estimates) Dependent Variable: Change in Number of Attacks between 1990-94 and 1985-89 Panel 1: S.A.P. Districts

All Maj. Kurd All Maj. Kurd All Maj. Kurd ∆ Unemployment -0.008∗∗∗ -0.010∗∗∗ -0.008∗∗∗ -0.010∗∗∗ -0.007∗∗∗ -0.009∗∗∗ (0.003) (0.003) (0.003) (0.003) (0.003) (0.003)

∆ T urnout 0.001 0.001 0.001 0.001 (0.002) (0.002) (0.002) (0.003)

∆ Education -1.653 -0.921 (1.802) (2.198) N 54 37 54 37 54 37

47 Panel 2: Non-S.A.P. Districts

All Maj. Kurd All Maj. Kurd All Maj. Kurd ∆ Unemployment -0.001 -0.001 -0.001 -0.002 -0.002 -0.002 (0.001) (0.003) (0.001) (0.003) (0.002) (0.003)

∆ T urnout 0.002∗ 0.001 0.002∗ 0.001 (0.001) (0.001) (0.001) (0.001)

∆ Education 0.344 1.492 (0.437) (1.704) N 82 30 81 29 81 29 District-clustered standard errors in parentheses. Panel 1 reports results of OLS regressions of change in total number of attacks per 1000 people between 1985-89 and 1990-94 on change of unemployment from 1985 to 1990 for districts that are covered by Southeastern Anatolia Project (SAP). Panel 2 reports the same estimates for districts not covered by the project. All speciﬁcations include dummies for districts losing villages to a newly formed district between 1985 and 1994, and dummies for districts that become central districts in a newly formed province in the same period. ∗∗∗, ∗∗, ∗ indicate statistical signiﬁcance at 1%, 5 % and 10 %, respectively.