THE UNIVERSITY OF CHICAGO ECONOMIC SHOCKS AND INSURGENT STRATEGY: EVIDENCE FROM PAKISTAN A BACHELOR THESIS SUBMITTED TO

THE UNIVERSITY OF CHICAGO ECONOMIC SHOCKS AND INSURGENT STRATEGY: EVIDENCE FROM PAKISTAN A BACHELOR THESIS SUBMITTED TO THE FACULTY OF THE DEPARTMENT OF ECONOMICS FOR HONORS WITH THE DEGREE OF BACHELOR OF THE ARTS IN ECONOMICS BY MICHAEL GALPERIN CHICAGO, ILLINOIS MAY 216

Economic Shocks and Insurgent Strategy Michael Galperin May 216 Abstract Empirical studies of civil conflict often rely on aggregate measures of violence, such as the number of deaths or the number of attacks, to attempt to infer the determinants of conflict. This broad tactic ignores substantial variation in insurgent strategic choice. Using new data on political violence in Pakistan, I examine whether conventional rebel violence and asymmetric rebel violence respond differently to economic shocks. Using the same data, I also perform a general analysis of the determinants of civil conflict overall. Negative shocks to local agricultural revenue are significantly related to increases in insurgent violence, which is consistent with an opportunity cost theory of rebel violence. These results are robust to a number of specifications and outcome measures of violence. I find that the responses of asymmetric and conventional violence to income shocks are similar in magnitude, indicating that rebels reduce all types of conflict activity in response to positive economic shocks rather than substituting to potentially less costly forms of insurgency. I conclude by offering potential directions for future research in evaluating differential effects of violence types to economic shocks. 1 Introduction A broad literature has shown that there is tremendous variation in the forces driving civil war. These include opportunistic greed on the part of rebels (Collier and Hoeffler, 24; Weinstein, 27); ethnic, religious, or linguistic grievances (Horowitz, 1985; DeVotta, 24); ideology (Staniland, 215); political exclusion (Cederman et al., 21); state weakness (Fearon and Laitin, 23); and community forces (Petersen, 21). This variety presents a fundamental estimation problem for empirical studies of conflict: absolute measures University of Chicago. Last updated November 217 as writing sample for application to graduate school. I thank Ethan Bueno De Mesquita for his invaluable advisorship and expertise, which was instrumental at every level of this research project. I also thank Victor Lima and Grace Tsiang for their extremely valuable advice and help throughout this project, and Stephane Bonhomme and Daniel Roberts for many thoughtful comments. 1

of conflict intensity, such as the number of deaths, do not provide information on rebel strategic choice or motivation. This inability has led to considerable disagreement among empiricists, and has encouraged efforts to formalize standard regression strategies that include the correct independent variables. 1 Much less attention has been paid, however, to how results differ across different types of conflict. Most empirical studies of conflict estimate their effects either by pooling together disparate types of conflict incidents, such as guerilla violence, terrorism, or conventional attacks by insurgent groups, or by studying these types of conflict in isolation by using datasets that record only a specific type of political violence (Bueno de Mesquita et al., 215). For this reason, little is understood empirically about the causes of variation in insurgent strategy. This variation is tremendous in scope, and has both analytically important consequences for research and operationally valuable implications for counterinsurgency. In this paper, I use new data on political violence in Pakistan to measure the causes of variation in insurgent strategy. I distinguish between conventional violence, which includes highly organized attacks on the state by militant groups and is characterized by high levels of rebel mobilization, and asymmetric violence, which requires far fewer insurgents and includes incidents such as guerilla attacks and terrorist acts. My analysis tests for these differential effects by evaluating the responses of each violence type to exogenous economic shocks. Following Dube and Vargas (213), I proxy for economic variables by interacting the regional production of primary commodity exports with their international price, which is taken as exogenous. Much of my analysis focuses on cotton and rice, two of the largest Pakistani commodity exports. 2 While these exports are a substantial share of Pakistani GDP, the country s low absolute export volume relative to the total global trade in cotton and rice makes it plausible to assume that shocks to international prices are truly exogenous. The distinction between conventional and asymmetric violence, while still a simplification of the myriad of strategies used by rebels, is an improvement over existing literature which studies right-hand-side variables extensively but treats variation in outcome variables as unimportant. In doing so, existing papers measure broad trends in insurgent violence but fail to account for patterns of endogenous substitution by rebels in choosing their strategies. In contrast, I argue that analyzing the differential violence types addressed in this paper is essential to more deeply understand the causes of variation in rebel violence. While the absolute level of rebel violence is also a product of insurgent strategy (Weinstein, 27), the type of violence employed by rebels is intimately tied to their strategic and operational goals. 1 For a review, see Hegre and Sambanis (26). 2 Rice is Pakistan s #2 primary commodity export from 1988 through 1992, and is the #1 export afterwards. Cotton is the #1 export from 1988 through 1992, before being overtaken by rice. Afterwards, in the time window for my analysis, it does not drop below the #4 primary commodity export level. Commodity data is from Bazzi and Blattman (214). 2

The paper has two main findings. First, negative local shocks to income, as proxied by shocks to local agricultural revenue, are significantly related to higher levels of conflict. This is consistent with an opportunity cost model of insurgent violence, in which positive shocks to income decrease violence by increasing the opportunity cost of rebellion. Second, and more surprisingly, the effects of these shocks are remarkably similar on conventional and asymmetric conflict when the outcome measure of violence is disaggregated to consider both violence types. This stands in supposed contrast to accounts of conventional and asymmetric insurgencies different operational strategies and goals, and suggests that economic shocks are relevant concern in determining the level of conflict across multiple types. More importantly, this result can be interpreted as running counter to the notion that rebel leaders endogenously substitute from one type of conflict to another in response to transitory changes in mobilization an assumption commonly invoked in the political science literature on political violence (see, e.g. Fortna, 215). Rather, the lack of a substantial difference in effects on asymmetric and conventional violence is evidence that rebel leaders scale all types of violence in response to economic shocks. The paper proceeds as follows. Section 2 provides a review of the empirical literature on the determinants of civil conflict, and establishes the theoretical motivations for testing for heterogeneity in the responses of different political violence types to economic shocks and environmental conditions. Section 3 discusses data sources and provides descriptive statistics. Section 4 describes my estimation methodology and provides regression results, and Section 5 interprets them in the context of insurgent strategy and concludes. 2 The Determinants of Civil Conflict 2.1 Economic Incentives and Rebel Types Economic theory yields two competing predictions of income shocks effects on the incidence of conflict. In one prediction, positive income shocks increase the opportunity cost of participating in insurgency, and therefore make conflict less likely by decreasing labor supplied to violence. This opportunity cost effect mirrors a fundamental prediction from the seminal Becker (1968) crime model: improvement of the outside alternative to violence should decrease the amount of violence. On the other hand, positive economic shocks increase the value of contestable goods, and could thus make conflict more likely as the increased spoils of violence present incentives for violent appropriation. This is the rapacity effect. Dal Bo and Dal Bo (211) formalize this distinction in an equilibrium trade model which predicts that positive shocks to labor intensive industries decrease conflict, while positive shocks to capital intensive industries increase 3

it. Much of the empirical literature connecting economic shocks to violence uses large, cross national datasets to evaluate this relationship. A fairly robust and agreed upon conclusion is that higher national incomes and rates of economic growth are negatively correlated with conflict (Blattman and Miguel, 21; Miguel et al., 24). Beyond this general conclusion, however, there is considerable disagreement on the mechanism linking economic shocks to rebellion. For example, Fearon and Laitin (23) take per capita income as a proxy for state strength, and argue that state weakness (proxied by low per capita GDP) encourages insurgency by increasing the probability that rebellion is successful. This stands in contrast to an influential paper by Collier and Hoeffler (24), who claim that low per capita incomes make rebellion less costly by reducing the income that insurgents forego by choosing to rebel. There is also considerable disagreement on the role of natural resources in financing or encouraging rebellion. A common proxy variable for the value of natural resources, and one that I employ, is the value of primary commodity exports. Collier and Hoeffler document a significant link between civil conflict and high levels of dependence on primary commodity exports in constituting a country s wealth. This relationship, taken as evidence of the rapacity effect, is interpreted by the authors as evidence that rebels are greedy and are motivated primarily by material wealth, although this interpretation is disputed by some (see, e.g. Humphreys, 25). Oil is commonly invoked in particular as a natural resource linked to conflict; while the causal link of oil to conflict is fairly robust across studies, the connection of agricultural commodities to conflict is less so (Ross, 24). More recently, Bazzi and Blattman (214) have used new data on international commodity prices to evaluate the relationship between economic shocks and conflict. The authors find that shocks to commodity prices do not affect the incidence of conflict, but may affect its duration. Despite these results, a number of problems endemic to large, cross national studies of conflict limit their predictive power. Regressions of civil war onset on various country level characteristics must address the issue of comparing a highly transitory variable (the onset of civil war) with factors that are often less likely to change drastically (e.g. population), forcing researchers to make arbitrary choices of the time scale on which they evaluate wars. 3 Such regressions are also unable to incorporate country-specific elements, such as government policies, which are instrumental in shaping the dynamics of war but difficult to quantify or compare across nations. Finally, since the specification of cross country regressions studying conflict requires the use of time scales that are often quite broad (for example, Collier and Hoeffler use civil war 3 For example, Fearon (25) shows that the influential result of Collier and Hoeffler (24) is sensitive to the time scale on which the dependent variable, civil war onset, is considered. While Collier and Hoeffler use 5-year periods in their analysis, the result is not robust to other specifications of timeframe. 4

onset within a five-year window as their outcome variable), such studies are often subject to significant endogeneity concerns. Civil war often has devastating effects on countries economies, and frequently also drastically affects common regressors such as ethnic composition and population. This endogeneity issue has led some researchers to study conflict by attempting to instrument for economic variables (see, e.g. Miguel et al., 24). One increasingly popular response to the issues inherent to cross national conflict studies is to instead study within country variation in violence due to localized shocks. Interestingly, results at the subnational level frequently differ from national level results. Berman and Couttenier (213) find that although income shocks do not affect conflict at the national level, local shocks (including those occurring as a result of shocks to commodity prices) are significantly related to the incidence and intensity of conflict. There is also mixed support at the subnational level for the opportunity cost rebellion mechanism. While Berman et al. (211a) finds no relationship between unemployment and insurgent violence in Afghanistan, Iraq, and the Philippines, Berman et al. (211b) finds that government counterinsurgency efforts aimed at increasing employment in Iraq are robustly linked to decreasing conflict. An especially relevant empirical within country study is Dube and Vargas (213), who use municipalitylevel data on Colombia to show that while positive shocks to coffee production (a labor intensive industry) decrease conflict, positive shocks to oil production (which is capital intensive) increase it. This distinction allows the authors to empirically test for the existence both of opportunity cost effects of income shocks on conflict in labor intensive industries such as coffee, and of opposite-direction rapacity effects in capital-intensive industries such as oil. The authors proxy for economic shocks to individual economic sectors by interacting local production of oil and cotton with their exogenous international commodity prices. Thus, variations in the international commodity price can be interpreted as causing revenue shocks to local economies. The authors find consistent evidence that labor intensity determines the balance of the opportunity cost and rapacity effects. For labor intensive goods, a positive shock to income reduces conflict by increasing the implicit costs of foregoing market labor. For capital-intensive goods, the increase in contestable value from a positive income shock increases conflict by making it more potentially lucrative. Using district level data on the production of rice and cotton, I perform a similar analysis on Pakistan. 2.2 Dependent Variables Despite the mixed results of this empirical literature, one well-documented phenomenon is the sensitivity of empirical results on civil conflict to differential empirical specifications or codings of outcome variables. 5

Difficulties in these codings include the definition of a threshold at which a given conflict is deemed a civil war, information on when conflicts start and end, and clear identification of the distinct actors involved in conflict (Sambanis, 24). These difficulties often influence empirical results; for example, Fearon (25) demonstrates that the hugely influential result of Collier and Hoeffler is undermined by small changes in sample framing. While studies frequently attempt to make up for potential data deficiencies or coding errors by running a number of robustness checks, most of these tests are ad hoc and seldom use a variety of data sources to test their predictions. This risks faulty conclusions due to coding practices inherent to the data (Hegre and Sambanis, 26). However, concerns over the proper coding of outcome variables mask a deeper conceptual problem in the empirical study of civil conflict. Studies that treat conflict as a unitary action without meaningfully distinguishable types ignore the substantial variation in insurgent strategy that occurs within conflict, and in particular fail to measure patterns of endogenous substitution in rebel strategy in response to external shocks to mobilization. While this distinction has been largely ignored by the economics literature on civil conflict, endogenous variation in insurgent strategy shaped by external contingencies is a well established phenomenon among political scientists. The tremendous variation in the degree of insurgent violence, for example, is taken as a result of strategic choices stemming from insurgent groups initial endowments of wealth and social capital (Weinstein, 27). Insurgents also frequently vary in the degree to which they collude with state governments, owing either to strategic choices of group survival or materialistic calculations of potential gain (Staniland, 212; Seymour, 214). In the case of terrorism, insurgent groups may endogenously choose to engage in small-scale asymmetric conflict instead of confronting states directly as a strategy aimed at coercing states into ceding territory (Pape, 23). These examples are but a small sample of the significant dimensions along which insurgent conflict varies and none of these analytically important factors are captured by empirical studies that do not attempt to identify insurgent strategic choices. Furthermore, the attempt to create datasets that correctly set thresholds to classify events as civil wars, while useful for large cross national studies, ignores this important variation in the technology of conflict pursued by insurgent groups. Even among studies using datasets that carefully specify civil wars, this represents a potentially significant sample selection problem. For example, it is unclear whether empirical studies that use outcome data only on certain types of violent incidents have relevance for understanding civil violence in general, as this strategy for inference samples on the dependent variable by only including the conflict type of interest (Bueno de Mesquita et al., 215; Ashworth et al., 28) Bueno de Mesquita (213) formalizes the notion of insurgent strategic substitution in a game theoretic 6

model of insurgent tactical choice, which I incorporate into my design. In the model, rebel leaders observe insurgent mobilization among the population and choose between conventional action, which is highly labor intensive; asymmetric conflict, which requires fewer participants but is less effective at high levels of mobilization, and withdrawal from conflict. The model predicts that while the degree of conventional violence is linear in mobilization, the effect of mobilization on asymmetric violence is nonmonotone. At very low levels of mobilization, rebel leaders choose to withdraw from conflict rather than fighting, and at high levels, they choose to substitute to conventional violence rather than pursuing less effective asymmetric strategies. This differential effect highlights the need for empirical studies that distinguish between different types of political violence. In the particular case of insurgent substitution among violence types, a study that measures only total numbers of conflict incidents may measure violence as decreasing overall when insurgents are merely substituting among different strategic possibilities. The main innovation of this paper is to measure these differential effects empirically. 3 Data and Descriptive Statistics 3.1 Data Sources and Coding For data on conflict outcomes, I use the BFRS Political Violence in Pakistan Dataset (Bueno de Mesquita et al., 215). The dataset records over 28, conflict events in Pakistan between 1988 and 211, and is coded from regional editions of The Dawn, the largest syndicated newspaper in Pakistan. An innovation of the BFRS data, and one that is important for my analysis, is that the data distinguishes among different types of politically violent acts. In particular, each observation in the BFRS dataset reports information on the specific violent actor(s) involved, the number of casualties, the location of the event, and a classification of the type of violence reported (e.g. terrorism ), whenever this information is available. This represents a significant improvement over previous databases recording political violence, which rely mostly on death toll measures to report on the severity of conflict while ignoring the specific details of the technologies of violence used by rebels in each case. Furthermore, the data provides conflict information at the event level, a degree of specificity that is unusual among datasets on political violence that rely on aggregate measures. Using the violence categories included in the data, I follow Bueno de Mesquita et al. (215) in grouping conflict events into a number of categories in order to distinguish between conventional and asymmetric insurgent violence. Militant Attacks, used in the first model to evaluate broad patterns of violent response to economic shocks, are coded as including attacks on state targets, conventional attacks on military, paramil- 7

itary, police or intelligence targets, and guerilla attacks on military, paramilitary, police or intelligence targets. I continue by evaluating differential effects on two different violence categories. Militant Asymmetric Attacks include terrorist attacks by militants and guerilla attacks on military, paramilitary, police or intelligence targets. Militant Conventional Attacks include conventional attacks on military, paramilitary, police or intelligence targets and militant attacks on the state. The unit of my analysis is the district year. After cleaning, this results in 242 unique district year observations in the baseline specification. Agricultural data is taken from the UN Food and Agriculture Organization s Global Agro-ecological Zones (GAEZ) database (FAOUN, 212). In particular, I use total production levels for cotton and rice, assumed to be fixed over the timespan of my analysis. I then aggregate agricultural data at the district level using the shapefile boundaries from the U.S. Census Bureau population data. These production measures are interacted with international commodity price data, released publicly by Bazzi and Blattman (214). Panel A of Figure 1 plots the evolution of commodity prices over time. There is considerable variation in the commodity price for both goods. There is also evidence of differential trends: for example, the price of rice increases sharply after the year 2, while the price of cotton oscillates but does not seem to have a clear trend upwards or downwards. Panel B of Figure 1 plots the evolution of each commodity s export weight over time, defined as the proportion of total Pakistani export value represented by the commodity. While cotton is a hugely important Pakistani export at the beginning of the sample period, its relative importance in exports declines significantly over time. On the other hand, the export weight of rice remains relatively constant and high in absolute terms. 3.2 Descriptive Statistics and Figures For ease of interpretation, Figure 5 in Appendix A provides a descriptive map of Pakistan, subdivided into provinces and showing surrounding countries. Table 1 reports descriptive statistics for the outcome variables used in my analysis, which are the number of attacks and the number of deaths in a given district in a given year. Roughly 38.5% of district years have at least one militant attack in a given year, and roughtly 54.7% of district years have at least one reported death due to political violence. There is substantial variability in the degree of violence across district years: while some years in some districts are entirely peaceful, other districts experience as many as 317 attacks in a single year. Figure 2 shows the number of recorded militant attacks in each district over the entire time window of the BFRS data. As the map shows, there is also substantial variablility in conflict intensity by geographic 8

Figure 1: Cotton and Rice Commodity Prices by Panel A: Commodity Prices by Commodity Price 5 1 15 2 25 3 Rice Cotton 199 1995 2 25 Panel B: Export Weights by Export Weight..5.1.15 199 1995 2 25 Rice Cotton Note: Panel (A) shows the evolution of international commodity prices for rice and cotton over time. The price of rice is given by the thick black line, and the price of cotton is given by the dashed line. Panel (B) plots the evolution of each commodity s export weight over time, measured as the share of total export value represented by the given commodity. International commodity price data is taken from Bazzi and Blattman (214). 9

Table 1: District-Level Violence Statistics Variable Mean (Std. Dev.) Min. Max. N Number of attacks 2.772 (11.741) 317 2662 Number of attacks, if at least one attack occurred 7.26 (18.72) 1 317 124 1{At least one attack occurred}.385 (.487) 1 2662 Number of Deaths 11.779 (57.997) 151 2662 Number of Deaths, if at least one death occurred 21.535 (77.81) 1 151 1456 1{At least one death occurred}.547 (.498) 1 2662 Note: This table shows average violence statistics by district year. The expression 1{At least one attack occurred} is the indicator function, and evaluates to 1 if an attack occurred in the given district in the given year, and otherwise. region. In particular, high counts of violent incidents are recorded in the Federally Administered Tribal Areas (FATA) to the country s northwest, as well as the Azad Jammu and Kashmir (AJK) territory in the northeast bordering the disputed Kashmir territory. Comparatively lower levels of violence are reported in Gilgit-Baltistan and immediately south of FATA, with intermediate levels throughout most of Balochistan and Punjab. The spatial distribution of violence also differs considerably between conflict types if militant attacks are disaggregated into conventional and asymmetric events. Figure 6 in Appendix B demonstrates this heterogeneity graphically. There is also substantial geographic variability in the production intensity of agricultural commodities. Figure 3 shows cotton and rice production along with Pakistani administrative district boundaries. The most intense production occurs in the central and northern part of Pakistan, with particularly intense production occurring in northern Punjab. Notably, GAEZ does not report agricultural data in AJK, as this is considered part of the disputed Kashmir territory by some accounts. This lapse in data is unfortnate, especially given the high incidence of conventional militant attacks in AJK reported by the BFRS data. As described in Appendix E, I correct for this omission in some regression specifications by imputing crop production data to missing value districts from adjacent nonmissing districts. The imputation does not significantly change the regression results. In addition to the substantial geographic variation exhibited in Figure 2, the level of militant violence is highly variable over time. Figure 4 demonstrates this phenomenon by plotting the evolution of violence over time separately for each Pakistani province. Violence generally increases over time for most provinces, especially in FATA and Khyber Pakhtunkhwa. Two notable exceptions are AJK and Gilgit-Baltistan, which see increases in violence followed by subsequent declines. In the case of Gilgit-Baltistan, this observed pattern might be due to a low number of observed violent incidents in general in that province. I account for the possibility that secular trends in the overall level of violence differ across provinces by including 1

province level linear time trends in my empirical specification. Similar plots to Figure 2, disaggregated to separately record different violence types, are given in Appendix C. In general, there asymmetric and conventional violence are highly correlated; provinces with high levels of one violence type in a given year typically are also high in the other type. This suggests that the observed levels of both violence types in a given province are likely to at least in some part be the result of actions by the same organizations in other words, it is likely that some organizations engage both in conventional and asymmetric violence, and that the violence types are complements from the perspective of rebel leaders. Differential responses of the violence types to economic shocks are therefore likely to represent endogenous substitution on the part of rebel leaders, who strategically choose their violence type in response to external shocks to mobilization. If, on the other hand, the effect of shocks is the same for asymmetric and conventional conflict, then rebel leaders are not substituting but instead decreasing activity across all types. Province AJK Balochistan FATA Gilgit Baltistan KPK Punjab Sindh Table 2: Descriptive Statistics by Province Cotton Prod. Rice Prod. Ruggedness Num. Eth. Frac. (Tons/) (Tons/) (1 km) Districts 19.332 13.919.472 (8.811) (11.251) (.87) 6.981 12.52.215.16 (1.926) (3.613) (.47) (.19) 11.355.413.16.329 (3.198) (.162) (.7) (.25) 5.454 2.18.863 (4.997) (1.872) (.31) 15.912 3.971.139.343 (2.943) (1.171) (.28) (.56) 19.511 138.78.213.43 (1.81) (14.739) (.3) (.19) 53.554 83.385.284.3 (9.661) (11.38) (.56) (.1) Note: This table reports means of key district level variables by province, with standard errors in parentheses. Due to low data availability, all variables in this table are assumed fixed over the time period of analysis. Ethnic fractionalization is the probability that two randomly chosen individuals from a given district will belong to different ethnolinguistic groups, and ranges between and 1. Ethnic fractionalization data is unavailable for AJK and Gilgit-Baltistan. Also, the reported agricultural production data for these regions is imputed from adjacent districts, as described in Appendix E. 7 25 7 5 23 36 18 Pakistan s provinces also vary significantly along other dimensions which may contribute significantly to the observed level of conflict. Table 2 provides descriptive statistics at the province level for a number of variables of interest. There is substantial variation in crop production across regions of Pakistan; Punjab is by far the highest-producing region of both cotton and rice over the sample period. In addition, the 11

relationship between cotton and rice production varies by province FATA, for example, is a relatively large producer of cotton but produces little rice, while Sindh produces far more rice than cotton. Also reported in Table 2 are descriptive statistics on Ethnic Fractionalization and Terrain Ruggedness, two frequently cited determinants of conflict intensity. Appendix D details the sources of these data. Because of the lack of available time-series data on ethnic fractionalization and the mostly time invariant properties of terrain ruggedness, these variables are not included in my primary regression specification, as they are absorbed into the included district level fixed effects. Furthermore, any regression that includes fixed effects at the province level and uses the variables in Table 2 as regressors to estimate conflict is unable to report anything beyond correlational estimates; in particular, the variation in ethnic fractionalization and terrain ruggedness is not necessarily exogenous with respect to the level of conflict. For this reason, I include results from this regression type in an Appendix (D), but do not report them in the main body of the paper. The results should be interpreted with this caveat in mind. 4 Methodology and Results I model insurgent violence as a function of economic shocks, district and time fixed effects, and province level linear time trends. Variables are aggregated to the district year level for the analysis. I estimate two regression designs. The first is intended to measure the effect of these variables on aggregate measures of conflict, such as the total number of militant attacks in a district in a given year. This specification allows for identification of broad trends in insurgent violence. The within country nature of my analysis offers a potential improvement over cross national studies of conflict, as many unobservable characteristics are likely to remain relatively constant over Pakistan s small geographic area. The second regression design aims to disaggregate this effect into effects on conventional and asymmetric violence. This decomposition is the main innovation of this paper, and allows testing for heterogeneity in the response of violence to its economic and population-level determinants. I make a number of assumptions throughout the analysis. First, due to the unavailability of time series data for these variables, I assume crop production to be fixed over the sample period. As Dube and Vargas (213) argue, while this assumption is an unfortunate consequence of poor data availability, it may also improve the validity of regression results by not recording the endogenous response of crop production to civil conflict. Second, I assume that Pakistan is a price-taker in the international market for cotton and rice, and that price shocks can be considered exogenous. This assumption follows established trends in the 12

Figure 2: Militant Attacks by District Total Number of Militant Attacks, 1988-211 - 6 7-14 15-27 28-43 44-81 82-125 126-28 29-344 345-67 68-1238 Notes: This figure shows the total number of recorded militant attacks recorded in the BFRS dataset, over the entire timespan covered by the data. Militant attacks include attacks on state targets, conventional attacks on military, paramilitary, police or intelligence targets, and guerilla attacks on military, paramilitary, police or intelligence targets. Shapefile information for administrative boundaries is taken from the U.S. Census Bureau s Demobase Pakistan Project (U.S. Census Bureau, 21). 13

Figure 3: Cotton Production in Pakistan with District-Level Boundaries Total Cotton Production Total Rice Production Tons High : 6.954 Tons High : 9.18481 Low : Low : Notes: This figure shows cotton and rice production throughout the country of Pakistan, as reported in raster data from the UN Global Agro-ecological Zones (GAEZ) database. Shapefile information for administrative boundaries is taken from the U.S. Census Bureau s Demobase Pakistan Project (U.S. Census Bureau, 21). 14

Total Militant Attacks by, by Province 2 Azad Jammu and Kashmir 3 Balochistan 15 1 5 2 1 5 4 Federally Administered Tribal Areas 4 Khyber Pakhtunkhwa 3 2 1 3 2 1 Punjab Sindh 15 3 1 5 2 1 Gilgit Baltistan 7.5 5. 2.5. Figure 4: Province-Specific Violence Trends Notes: This figure shows the evolution of violence over time for every province in Pakistan. Violence is measured as the total number of militant attacks recorded in the BFRS dataset in the given year. Militant attacks include attacks on state targets, conventional attacks on military, paramilitary, police or intelligence targets, and guerilla attacks on military, paramilitary, police or intelligence targets. The trends given are optimal bandwidth Nadaraya-Watson kernel regression estimates, with symmetric 95% confidence intervals. 15

conflict literature. On average over the time frame of my analysis, Pakistani exports of cotton are 4.5% of global export volume, and exports of rice are 8.8%. Both of these values fall below the 1% cuttoff used by Bazzi and Blattman (214) in their analysis. 4 4.1 Model 1: The Effect of Economic Shocks on Political Violence Given provinces j, districts i within these provinces, and time t indexed by year, the primary initial OLS specification for testing for broad violence trends is the following: y ijt = α i + β t + δ j t + (Crops i P t ) λ + ɛ ijt (1) Here, y ijt is a measure of conflict. I consider two outcome variables: the number of attacks in a given district in a given year, and the total number of deaths from conflict in a given district in a given year. The coefficients α i and β t represent district and time fixed effects, respectively. The term δ j t is a province specific linear time trend, and is intended to capture differential trends in insurgent violence over time at the province level for reasons external to changes in agricultural revenue. As Figure 4 demonstrates, there is substantial variation in insurgent violence across provinces and over time, and the temporal trend in violence differs across provinces. This differential trend motivates my inclusion of this trend. The economic coefficient of interest is λ, the regression coefficient for the variable (Crops i P t ). This variable interacts the district level production of crops with their time varying international price. Therefore, λ measures the effect of a unit increase in revenue from crops, which in my analysis are cotton and rice, on the conflict measure y ijt. I take changes in the international commodity price P t to be exogenous, and therefore interpret changes in the variable (Crops i P t ) as local income shocks at the district year level. I consider two crops, cotton and rice, in my analysis. For this reason, the variable (Crops i P t ) and the coefficient λ are actually two-component vectors (one component for each crop), and λ = (λ c, λ r ) records the effect of shocks to cotton and rice revenue separately. 4.1.1 Results Table 3 records the results of OLS estimation of Equation 1. Column (1) is a simple baseline specification that does not include the province-specific time trend δ j t. Columns (3) and (5) correct for missing crop 4 There are some years in which Pakistani exports of rice exceed the 1% threshold used by Bazzi and Blattman (214), though this share never exceeds 15%. An interesting direction for further study would be a paper that accounts directly for the possibility of endogeneity by instrumenting for shocks in the international price of rice. 16

data in some districts by imputing these values from adjacent districts, in a procedure that is detailed in Appendix E. Columns (4) and (5) lag the price shock to revenue by one year. Column (5) is my preferred specification. The regression yields two main results. First, increased agricultural revenue due to rice production is significantly and negatively related to conflict. This result is robust across all specifications, and is consistent with an opportunity cost model of insurgent choice in which positive shocks to income increase the opportunity cost of rebellion, and therefore decrease its amount. The magnitude of this effect is large: in the preferred specification, a standard deviation increase in the international price of rice is responsible, in the average rice producing district, for approximately.91 fewer militant attacks in a given year. This decrease is approximately equal to 32.9% of the average district year level of violence. Second, shocks to agricultural revenue from the production of cotton are not estimated to have statistically significant effects on the degree of violence in a given district year. A number of factors may explain the difference in this effect. First, the variance of the price of rice is slightly higher than the price of cotton, 5 so the larger economic shocks due to this fluctuation may be responsible for higher responsiveness of outcome violence measures. Second, and more importantly, the export weight of cotton declines substantially over the analysis period as shown in Figure 1, while the export weight of rice remains relatively constant over time. The sharp decline in the importance of cotton exports in total Pakistani export volume over the time period is consistent with the result that international cotton price shocks have less importance in determining the degree of violence. 4.2 Model 2: Differential Responses by Conflict Type Theory predics that the responses of conventional and asymmetric violence to shocks to mobilization may be different in magnitude or even opposite in sign (Bueno de Mesquita, 213). I test this theory by estimating the following equation relating asymmetric and conventional violence to economic shocks and local characteristics: y ijt = α i + β t + δ j t + Conventional i ρ + (Crops i P t ) λ + [Conventional i (Crops i P t )] η + ɛ ijt (2) As in Equation 1, y ijt is a measure of violence in district i, belonging to province j, at time t. The model includes district-level fixed effects α i and time fixed effects β t, as well as the province-specific linear time 5 The variance of the log price of rice is.274. The variance of the log price of cotton is.25. 17

Table 3: Results of OLS Differences in Differences Regression Dependent Variable: (1) (2) (3) (4) (5) Cotton Production * log Price.1 -.6 -.6.8.15 (.2) (.26) (.25) (.27) (.26) Rice Production * log Price -.64 -.61 -.5 -.61 -.53 (.23) (.21) (.23) (.23) (.23) Dependent Variable: Number of Deaths Cotton Production * log Price.18 -.11 -.5.124.127 (.92) (.13) (.124) (.136) (.129) Rice Production * log Price -.371 -.375 -.343 -.436 -.399 (.19) (.176) (.168) (.214) (.23) Province level linear time trend no yes yes yes yes Imputed crop values for KPK and GB no no yes no yes Price shock lagged one year no no no yes yes Observations 2596 242 2596 242 2596 Adjusted R 2 Panel 1.229.268.249.264.246 Adjusted R 2 Panel 2.228.31.296.299.295 Notes: p <.5, p <.1, p <.1. Standard errors are clustered at the division level, and are reported in parentheses. Divisions are administrative divisions of Pakistan which are supersets of districts but subsets of provinces. Columns (2) through (5) differ from the baseline specification by including province-level linear time trends in the regression. Columns (4) and (5) differ from the baseline specification by lagging the price shock by one year. Columns (3) and (5) differ from the baseline specification by including districts for which crop data is initially missing, but is imputed according to the crop production of neighboring districts. Appendix E describes this imputation process. 18

trend δ j t. As in Equation 1, the variable ( ) Crops ij P t is interpreted as a local shock to agricultural revenue, and is decomposable into shocks from cotton and rice production. The methodology for estimating Equation 2 differs slightly from that required to estimate Equation 1. In particular, rather than using a dataset with one observation for every district year, the data is first stacked to contain two observations per district year: one measuring violent outcomes from asymmetric violence, and one measuring these outcomes from conventional violence. The dummy variable Conventional i indicates whether a given observation records conventional or asymmetric violence. In this construction, the estimate of the coefficient λ will record the effect of local income shocks on asymmetric violence. The estimate of η will measure the violence premium from conventional violence that is, it is a measure of the difference between the effects of local income shocks on conventional conflict and the effects on asymmetric conflict. Therefore, the sum λ + η gives the effect of agricultural price shocks on the level of conventional violence. Table 4 provides estimates from estimation of Equation 2. As in Table 3, columns (2) through (5) include province-specific linear time trends, columns (3) and (5) impute missing crop data for some districts, and columns (4) and (5) lag shocks to crop prices by one year. Column (5) is my preferred specification. Estimation of Equation 2 yields a number of conclusions. First, in every specification, positive shocks to local revenue from the production of rice are significantly and negatively correlated with the incidence and intensity of asymmetric violence. I interpret this as evidence of an opportunity cost mechanism in the production of violence. The effects are far less significant for price shocks to cotton than they are due to rice, which again may be due to the relatively larger importance of rice in Pakistani exports. The scaling parameter ρ is negative and significant in all specifications, and reflects the lower frequency of conventional attacks relative to asymmetric attacks. Second, the effect of shocks to agricultural revenue on the degree of violence is remarkably similar for conventional and asymmetric conflict. Indeed, in every specification the estimates of the conventional premium, η, are statistically identical to zero. This is a surprising result given the substantial qualitative differences between the technology of asymmetric and conventional violence in particular, asymmetric attacks typically require smaller levels of mobilization than conventional attacks, and insurgencies relying primarily on asymmetric tactics are on average less explicilty regimented and bureaucratic than primarily conventional insurgencies. Furthermore, the goals of asymmetric and conventional fighters may be different while conventional insurgencies are typically explicitly concerned with wresting territorial control from incumbent governments (DeVotta, 24), asymmetric insurgencies are often aimed at producing smaller scale territorial concessions (Pape, 23). 19

Table 4: OLS Estimation of Differential Effects by Violence Type Dependent Variable: (1) (2) (3) (4) (5) Conventional Dummy (ρ) -2.49-2.49-1.844-2.414-1.847 (.684) (.684) (.83) (.684) (.85) Cotton Production * log Price (λ c ).4 -.4 -.4.3.6 (.11) (.13) (.13).14) (.13) Rice Production * log Price (λ r ) -.38 -.31 -.25 -.31 -.26 (.11) (.11) (.12) (.11) (.12) Conventional Cotton Revenue (η c ).2.2.2.2.2 (.2) (.2) (.2) (.2) (.2) Conventional Rice Revenue (η r ).1.1 -..1 -. (.1) (.1) (.1) (.1) (.1) Dependent Variable: Number of Deaths Conventional Dummy (ρ) -1.63-1.63-1.157-1.625-1.152 (.654) (.655) (.713) (.656) (.715) Cotton Production * log Price (λ c ).15.7.7.2.21 (.12) (.15) (.14) (.14) (.13) Rice Production * log Price (λ r ) -.64 -.57 -.5 -.61 -.54 (.23) (.19) (.19) (.22) (.21) Conventional Cotton Revenue (η c ) -.1 -.1 -. -.1 -. (.3) (.3) (.2) (.3) (.2) Conventional Rice Revenue (η r ).1.1..1. (.2) (.2) (.2) (.2) (.2) Province level linear time trend no yes yes yes yes Imputed crop values for KPK and GB no no yes no yes Price shock lagged one year no no no yes yes Observations 2596 242 2596 242 2596 Adjusted R 2 Panel 1.178.19.169.187.167 Adjusted R 2 Panel 2.231.35.299.33.298 Notes: p <.5, p <.1, p <.1. Standard errors in parentheses. Columns (2) through (5) differ from the baseline specification by including province-level linear time trends in the regression. Columns (4) and (5) differ from the baseline specification by lagging the price shock by one year. Columns (3) and (5) differ from the baseline specification by including districts for which crop data is initially missing, but is imputed according to the crop production of neighboring districts. Appendix E describes this imputation process. 2

This result can be interpreted as evidence against the theory that rebel leaders make significant endogenous strategic substitutions in violence type in response to exogenous shocks to mobilization. Rather, while the level of conventional violence is lower on average than the level of asymmetric violence, the two violence types respond very similarly to economic shocks. Under the assumption of a unitary rebel leader choosing between two different types of violence given a level of mobilization, this would imply a broad organizational response to these shocks across all strategies rather than internal substitution based on the labor intensivity of different violence types. 5 Discussion of Results and Conclusion I begin by discussing the results relating economic variables to conflict. Negative shocks to agricultural revenue are significantly related to increased conflict incidence, as measured by the number of attacks, and conflict intensity, as measured by the number of deaths. The effect of revenue shocks due to cotton is much less significant, both in magnitude and statistically, than the effect of shocks due to rice. I argue that this discrepancy is due to the large export share of rice in Pakistan relative to the export share of cotton. In particular, this difference is especially severe towards the end of the sample period, when violence is most severe; while rice exports remain relatively constant, exports of cotton fall almost to zero, as demonstrated in Figure 1. For this reason, I argue that the difference in the effects of cotton and rice revenue on the degree of violence acts almost as a placebo exercise that demonstrates that the observed relationship between crop price fluctuations and the degree of violence depends on the crop s actual importance in Pakistani exports. The very low export weight of cotton towards the end of the estimation sample implies that shocks to the international price of cotton should have less bearing on local economies. As the results in Table 4 demonstrate, the effect of local revenue shocks on violence persists at almost the same level of magnitude for both asymmetric and conventional conflict. This homogeneity is surprising given the differences in the technology of conflict employed in conventional and asymmetric insurgencies, and provides a number of promising directions for future empirical research. While the distinction between asymmetric and conventional violence is useful for the purposes of my study, these broad categories are far from capturing the immense degree of heterogeneity in the strategies pursued by rebel groups. A finer characterization of differential violence types may provide deeper insights on the question of which types of violence respond particularly strongly to economic shocks. In particular, a relevant consideration that I largely exclude from my analysis is collusion between insurgent groups and the state, a consideration that 21

certainly has implications for the degree of rebel violence, but may affect conventional and asymmetric groups differently (Staniland, 212; Carey et al., 215). The main empirical difficulty in attempting to distinguish between different rebel strategies, however, is that the analyst is usually only able to observe the outcome of rebel strategy (for example, an actual attack) but not the strategic calculations that produce it. Data that classifies insurgent violence in great detail may therefore be hugely valuable in proxying for a wide range of rebel strategic choices. A second limitation of the current study is the paucity of high quality data available on Pakistan. In particular, the lack of time series agricultural data on cotton and rice production forces the assumption that these production levels are fixed. In reality, there is reason to believe that there exist substantial endogeneities between the level of conflict and agricultural income; for example, destructive attacks may make farming a dangerous activity, or lower the value of affected land. While this time series data did not exist for almost the entire duration of this project, it has very recently become available on a provisional basis as part of a collaboration between the government of Pakistan, the USDA, and a number of independent and academic institutions. 6 Incorporation of this data could potentially allow for approaches that directly address endogeneities in agricultural production as they respond to conflict levels for example, the level of rainfall could be used to instrument for production levels or economic variables under the assumption that these shocks do not affect conflict, as in Miguel et al. (24). Further studies could also incorporate other exports, such as wheat and sugar, to test for the persistence of these effects across agricultural categories. Despite these limitations, the current study does provide useful insight into the relationship between economic shocks and the level of conflict. In particular, positive shocks to income are robustly and negatively related to the number of attacks. This is consistent with an opportunity cost model of conflict, in which positive shocks to income decrease labor supplied to violence by increasing the opportunity cost of leaving conventional work. My results replicate the results of Dube and Vargas (213) in this regard, which find a similar relationship between income shocks due to changes in the international price of coffee and the level of insurgency in Colombia. As mentioned earlier in this report, Dube and Vargas also distinguish between opportunity cost effects, and the opposite rapacity effect in which positive shocks increase the value of contestable goods, and therefore increase conflict. The authors test for this effect by using changes in the international commodity price of oil, a capital intensive good. As Pakistan is also an exporter of oil, data on these exports could allow testing for similar rapacity effects in a different institutional setting, and offers another promising direction for future research. 6 See the Pakistan Agriculture Information System, accessible at http://dwms.fao.org/~test/home_en.asp 22