Policy Research Working Paper 9373 The Seasonality of Conflict Jenny Guardado Steven Pennings Development Economics Development Research Group August 2020 Policy Research Working Paper 9373 Abstract This paper investigates whether poor employment prospects and unanticipated, thus also varying the dynamic benefits of potential insurgents help to fuel conflict. The paper of fighting that confound estimates of the opportunity cost provides a new test of this “opportunity cost mechanism” mechanism. Empirically, the paper estimates the effect of using one of the largest shocks to labor demand in agricul- harvest shocks on conflict intensity in Afghanistan, Iraq, tural societies: harvest. Theoretically, the paper shows that and Pakistan using subnational variation in the timing because seasonal harvest shocks are temporary and antic- and intensity of harvest driven by local climatic condi- ipated, they change opportunity costs while keeping the tions. Consistent with the opportunity cost mechanism, dynamic benefits of fighting constant, yielding unbiased the results show that the onset of harvest usually reduces estimates even if those benefits are unobserved. In contrast, the number of insurgent attacks. many other shocks in the conflict literature are persistent This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at spennings@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team THE SEASONALITY OF CONFLICT JENNY GUARDADO* AND STEVEN PENNINGS** Date : 26 August 2020 . An Online Appendix is available at: link. *Assistant Professor. Georgetown University. jgr45@georgetown.edu. (Project developed while visiting the Harris School of Public Policy at the University of Chicago.) ** Research Economist. World Bank Research Department. spennings@worldbank.org. JEL: D74, O13, J43, Q34; Keywords : Conict, Harvest, Opportunity Cost Mechanism. The ndings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank and its aliated organi- zations, or those of the Executive Directors of the World Bank or the governments they represent. We are grateful for comments from Scott Ashworth, Eli Berman, Ethan Bueno de Mesquita, James Fearon, Anthony Fowler, Norman Loayza, Noel Maurer, Matias Iaryczower, John Londregan, Jacob Shapiro, Mateo Vasquez, and participants of the ESOC Lab Meetings at Princeton University, the Conict Seminar at Vanderbilt University and the IADB Washington Area Political Economy Sem- inar, the NYU Alumni Conference, the 2017 NEUDC Conference, the 2017 Cesifo Venice Summer Institute and the 2017 Princeton Political Economy Seminar, the 2015 Third Formal & Comparative Conference (Becker-Friedman Institute), the World Bank MENA Seminar Series, the Harris School of Public Policy Political Economy Lunch, the 2015 HiCN (Households in Conict Network) confer- ence, the Georgetown Political Economy & Development Seminars, and the 2016 Warwick-Princeton Political Economy Workshop. We thank Ethan Bueno de Mesquita and the Oce of Naval Research for nancial support and are grateful to Paula Ganga, Jeong Whan Park, Marissa Barragan, and Saisha Mediratta for their research assistance. Peasant recruitment and desertions in all the [Russian] civil war armies uctuated with the farming seasons. Peasants joined up in the winter, only to desert the following summer. In the central agricultural region, the weekly rate of desertion was up to ten times higher in summer than winter. Figes (1996: 596) 1. Introduction Understanding why civil conict occurs is of utmost policy importance given its toll on human lives and on broader development prospects. While traditional two-state conicts have been relatively rare post-WW2, around 40% of countries have had at least one civil war that has killed more than 1,000 people (Fearon 2008). The best predictor of civil war is low per capita income (Fearon and Laitin 2003), leading to a commonly held view that poverty and poor employment prospects are key drivers of conict (World Bank 2011). Indeed, a number of theoretical models establish a connection between the availability and rewards for work and the onset and intensity of conict, known as the opportunity cost mechanism (Becker 1968; Grossman 1991; Dal Bo and Dal Bo 2011). However, the empirical evidence on the opportunity cost mechanism is mixed. On one hand, a number of studies show that conict intensies with negative income shocks driven by rainfall (Miguel et al. 2004) or commodity prices (Bruckner and Ciccone 2010; Dube and Vargas 2013; Guardado 2016; Hodler and Raschky 2014, among others). But on the other hand, other studies nd a more mixed relationship between commodity prices and conict (Blattman and Bazzi 2014; Crost and Felter 2016), and some even nd that conict intensity increases with labor availability (Berman et al. 2011b) or income windfalls from aid (Nunn and Qian 2014; Crost et 1 al.2014; Weintraub 2016). In this paper, we make three contributions to the opportunity cost literature. First, we argue theoretically that the mixed empirical results in the literature may be be- cause the highly persistent shocks often studied change the unobserved dynamic ben- ets of ghting at the same time as the opportunity costs, confounding empirical estimates. Second, we propose using harvest shocks as a novel alternative, which are both temporary and anticipated and are also one of the largest and most common shocks to labor demand in developing countries. Third, we estimate the eect of 1In a similar vein, Iyengar et al. (2011) and Beath et al. (2017) nd both a positive and negative impact of development assistance programs in Iraq and Afghanistan, respectively. 1 harvest shocks on subnational conict intensity in Iraq, Afghanistan, and Pakistan and nd support for the role of opportunity cost considerations. Conceptually, shock persistence is important because in many conict settings, the costs of ghting are born today, whereas the benets of winning a conict are mostly in the future. For example, for conict over an oil-rich regionmotivated by greedthe future benets are the ow of rents after the oil wells are captured. In this context, a persistent oil price shock that stimulates an oil-exporting economy and increases the opportunity cost of ghting will also increase future prots from capturing oil wells. As both the opportunity costs and future benets of ghting increase, estimates of the strength of the opportunity cost mechanism driven by this shock will be biased toward zero. Given that many commodity price shocks are highly persistent (and are often in- distinguishable from a random walk; see, e.g., Cashin et al. 2000), this argument suggests that the opportunity cost eect captured by commodity prices is likely to be even stronger (more negative) than shown in current studies (e.g., Bruckner and Ciccone 2010; Dube and Vargas 2013; Blattman and Bazzi 2014; Guardado 2018; Hodler and Raschky 2014). We also show that the same concern arises in other models where conict is motivated by grievances or depends on counterinsurgency informants (Berman et al. 2011a), albeit through a dierent mechanism. In these grievance/counterinsurgency models, persistent shocks change the marginal utility of consumption, which then aects the subjective value attached to grievances or tip-o payments. Our theory suggests that the ideal shocks to uncover the strength of the opportu- nity cost mechanism are temporary or anticipated changes in labor demand. Seasonal shocks, like harvest, are both. Temporary or anticipated shocks do not aect the ex- pected benets of ghting, like the rents from capturing an oil well, because those rents are in the future (for temporary shocks) or the eect is already built in (for anticipated shocks). In a grievance/counterinsurgency framework, temporary or an- ticipated shocks are smoothed by the permanent income hypothesis (PIH), and so they do not change the marginal utility of consumption and the subjective value of the grievance/tip-o payment. Harvest has the added practical advantage of being a large shock to labor demand: harvests are common throughout the world in dierent conict settings and their exact timing and size are determined by exogenous local climatic conditions. 2 Our theoretical argument builds on that of Fearon (2008) and Chassang and Padro- i-Miquel (2009), who argue that permanent changes in incomes in a greed-motivated model will not change the level of conict, as both the costs and benets of ght- ing increase. Chassang and Padro-i-Miquel (2009) also argue that temporary shocks will aect conict via the opportunity cost mechanism. Relative to those papers, we generalize this argument using a quantitative regression framework using simulated data, where the future value of winning is an omitted (unobserved) variable. We also introduce a specic denition of the actual and measured strength of the opportunity cost mechanism, which allows us to precisely characterize the size and direction of the bias in opportunity cost regressions with seasonal or persistent auto-regressive shocks. We likewise show that the same results hold (for dierent reasons) in griev- ance/counterinsurgency models not considered by Fearon (2008) and Chassang and Padro-i-Miquel (2009). Beyond theory, there are hundreds of years of narrative evidence that harvest aects conict intensity through the opportunity cost mechanism. For example, during the American Civil War (18611865), desertions from the Confederate Army increased in the months of June and July, the harvesting times for tobaccoan important Southern crop at the time (Giure 1997). The introductory quote also suggests that during the Russian Civil War (19171922), desertion rates in the Red and White armieslargely formed of peasantswere higher during the summer harvest (Figes 1996, cited by Dal bo and Dal bo 2011: 657). The sensitivity of conict intensity to harvest stems from the fact that insurgencies in civil wars are largely fought by part-time ghters (see Appendix Figure A.3). There are many advantages for an insurgency to hire part-time ghters: they have valuable local knowledge, they may be cheaper than full-time ghters, and they can help pro- tect the more skilled professional insurgents from the risks of day-to-day ghting. For example, Vietcong guerrilla members famously worked as farmers during the day but fought US forces at night. Cline (2000) nds that the Moro Islamic Liberation Front in the Philippines lled battalions with part-time ghters on a monthly rotational basis. In Afghanistan, Taliban forces have been known to organize in village cells, each containing around 10 to 50 part-time ghters (Afsar et al. 2008). Part-time ghters were also common among Iraqi insurgents ghting against the US military presence and among the Shining Path insurgency in Peru in the 1980s (McClintock 1998). 3 In the empirical section of this paper, we estimate the eect of the wheat har- vest on subnational conict intensity in Iraq (20042009), Pakistan (20022010), and Afghanistan (20042014). Wheat is the main (legal) crop in all three countries and is generally harvested annually using labor-intensive methods. Unsurprisingly, house- hold survey data show that during harvesting months, agricultural workers tend to have higher employment rates than other rural workers, suggesting a positive seasonal labor demand shock at harvest time. As monthly data on local wages or employment are unavailable, we estimate a reduced-form relationship between the timing and size of the area harvested and the number of attacks in that location. As conict intensity varies widely across settings and years due to other factors, we use a normalized mea- sure of the number of attacks occurring in the month as a share of the annual total in that district, which can be compared across the three samples. We also use several dierent datasets of conict intensitybased on US government records as well as on local media reportsto ensure our ndings are robust to measurement error in recording conict intensity. While the use of transitory and anticipated shocks keep constant the dynamic ben- ets of ghting (as in the theory), our identication strategy also removes remaining concerns about reverse causality and omitted variable bias. To ensure against reverse causality (running from conict to harvest), we use pre-conict data on local climatic conditions that determine the timing and intensity of harvest at the subnational level. Other omitted variables are controlled for by district-by-year xed eects that partial out local characteristicseven those changing slowly over timeand monthly time eects to control for any aggregate political and economic shocks and for the sea- son. Our results are also robust to controls for temperature and precipitation (among others). Our main empirical nding is that in all three countries, the intensity of conict is lower at harvest than at other times of the year, with greater falls in areas with larger areas under cultivation. Specically, we nd that at the mean intensity of wheat cultivation, the onset of harvest reduces the average number of monthly attacks by around 6%22% in Iraq, 20% in Pakistan, and 8%18% in Afghanistan. Moreover, dynamic specications show that most of the reduction in conict is generally in the harvesting month itself; there is little evidence that the harvest may be promoting or nancing conict in subsequent months. Overall, these results provide some support for the opportunity cost playing an important role as a determinant of conict, even 4 for insurgencies thought to be driven by ideology, rather than economic factors, like the Taliban and Al-Qaeda. The remainder of the paper is organized as follows. Section 2 presents our theoret- ical argument illustrating the importance of shock persistence, and the advantages of seasonal shocks in estimating the true strength of the opportunity cost mechanism. Section 3 discusses our empirical approach in estimating harvest's eect on conict intensity. Section 4 presents our empirical results for Iraq, Afghanistan, and Pakistan. Section 5 concludes by discussing some policy implications. 2. Theory: Dynamic Benefits of Seasonal Shocks In this section, we seek to understand why the literature nds mixed results on the importance of the opportunity cost mechanism in explaining conict. We argue that the estimated strength of the opportunity cost mechanism depends on the type of shock studied: highly persistent shocks can lead to systematically biased estimates, while temporary and anticipated seasonal shocks can uncover the true magnitude of the opportunity cost mechanism. The intuition is as follows: in standard dynamic models of conict motivated by greed, the costs of ghting are temporary and are incurred today, whereas the benets of victory are persistent and are received in the future (Chassang and Padro-i-Miquel 2009). Since seasonal harvest shocks are both temporary (only a few weeks a year) and anticipated, they change the costs of ghting today while keeping constant its unobserved dynamic future benets, thus properly capturing the strength of the op- portunity cost mechanism. In contrast, many of the unanticipated and persistent shocks studied in the literature (e.g. persistent commodity prices or development aid) may aect both today's opportunity costs and tomorrow's gain from conict. These income shocks would thus bias estimates of the strength of the opportunity cost mechanism toward zeroparticularly when the dynamic benets are hard to observeexplaining the wide variety of estimates in the literature. This argument is robust to dierent motivations for conict. Below we present a dynamic greed model of conict, where insurgents engage in violence in order to capture a resource that has some monetary value, but we show in Appendix 2 that this argument also holds in models where conict has intrinsic motivations (grievances) or is driven by counterinsurgency information. For each model, we identify the size of the bias and provide a precise denition of the strength of the opportunity cost mechanism namely, the elasticity of time allocated to violence with respect to wages. 5 2.1. Greed Model. What we call the greed model of conict is one of the most popular in the literature (Haavelmo 1954; Hirshleifer 1988, 1989; Garnkel 1990; Skaperdas 1992; Garnkel and Skaperdas 2007; Fearon 2008; Chassang and Padro- i-Miquel 2009; Dal bo and Dal bo 2011), with our paper building on the last three papers. Similar to Chassang and Padro-i-Miquel (2009), we model the gains from victory as dynamic, whereas the costs are static, meaning that temporary but not permanent shocks aect violence, and that winning is decisive. As in Fearon (2008), conict varies at the intensive, rather than extensive, margin. Like Dal bo and Dal bo (2011), our appropriation/ghting technology is concave in labor (reecting con- gestion eects); our production function is non-linear in labor (such that real wages depend on the allocation of labor between working and ghting); and we abstract from the government's response to violence. 2.1.1. Model Setup. We study the problem of a representative insurgent ghter, who has one unit of time to split between labor (L) or violent activities (V ). The benet of time spent ghting is to increase the insurgent's probability of victory, represented 1−φ−1 by a contest success function p(Vt ) = ψVt , where φ >1 measures the strength of the opportunity cost mechanism and ensures positive but diminishing marginal 2 returns to ghting. Output in this economy Yt is produced with labor Lt and a xed factor of pro- duction ¯ (such as land N or a natural resource), which we normalize to one so that 3 Yt = θt Lα t . θt measures productivity in terms of consumption goods, which would capture commodity price shocks, seasonal shocks, or land productivity that varies 4 with rainfall, among others. An increase in θt increases output. We study persistent or seasonal (temporary and anticipated) shocks to θt . The reward for working is the wage, which is the marginal product of labor Wt = αYt /Lt . As such, total labor income Wt Lt is simply αYt . The remainder of output is 2This is a good approximation of the power form of the contest function (Garnkel and Skaperdas 2007, Equation 3) in a low-level insurgency where the strength of the government is −1 large (and constant). That is: ¯ m ) ≈ ψVt1−φ with m = 1 − φ−1 and p(Vt , G) = Vtm /(Vtm + G ¯m +G ψ ≈ 1/(V ¯ m if the government's strength G ¯ m ) ≈ 1/G ¯ m is very large relative to the insurgent's (and constant). For example, with our default calibration, V = 0.065, G ¯ = 1900, and φ = 3 such that the probability of the insurgents winning is ≈0.1% (per quarter). 3That Yt = θt Lα ¯ 1−α ¯ = 1. is, the production function is Cobb-Douglas tN with N 4In this former case, let the price of domestic consumption goods be the numeraire and the international price of the commodity be θt , and assume that the volume of cash crops is produced α α for export is Lt . Then the amount of consumption goods that can be purchased is θt Lt . 6 rents Π to the xed factor of production (land or a natural resource) Πt = (1 − α)Yt , which accrue its owner. The motivation for the insurgent to ght is to capture these rents and to gain Πt . We present two submodels that vary in their dynamics: a two-period model that we can solve analytically and an innite horizon model that we solve numerically. The two models are very similar. The size of the prize of winning, denoted as ΠP rize t+1 ≡ UtW in Lose +1 − Ut+1 equals Πt+1 in the two-period model, but is more complicated in innite horizon model. More formally, the ghter's problem is (1) UtL (θt ) = maxV,L Wt Lt + p(Vt ) ×EβUtW in L +1 (θt+1 ) + [1 − p(Vt )] EβUt+1 (θt+1 ) W age Inc P robW in P robLose −1 s.t. p(Vt ) = ψVt1−φ , Lt + Vt = 1, (taking Wt = αYt /Lt and Yt = θt Lα t as given) 2.1.2. Measuring the Opportunity Cost Mechanism. The rst-order condition of the insurgent's problem is Wt = p (Vt )βEt ΠP rize t+1 , where the right hand side is the gain p (Vt ) times the from ghting an extra hour: the increase in the probability of winning P rize expected discount prize from winning βEt Πt+1 . The left hand side is the foregone wage from working (i.e., the opportunity cost of spending an extra hour ghting). Substituting the functional forms and taking logs, we obtain an expression that can 5 be potentially taken to the data: (2) lnVt = κ1 − φlnWt + φlnEt ΠP rize t+1 . Denition 1. The true strength of opportunity cost mechanism is the T rue elasticity of violence with respect to wages, keeping everything else constant: βOpp ≡ ∂lnVt = −φ. ∂lnWt Denition 2. The measured strength opportunity cost mechanisms is the elasticity of violence with respect to wages, allowing the value of the prize of ghting M eas. dlnVt ∂lnVt ∂lnVt ∂lnΠP rize t+1 to change endogenously: βOpp ≡ = + . dlnWt ∂lnWt ∂lnΠP rize t+1 ∂lnWt 5The constant κ1 = φlnψ (1 − φ−1 ) + φlnβ . In equation 4, κ2 = κ1 − φlnα + φln(1 − α). 7 To illustrate, consider a standard univariate regression of changes in violence on some measure of opportunity cost of ghting, like wages: (3) lnVt = α0 + βOpp lnWt + et . T rue The objective is to estimate βOpp = −φ as in Denition 1. However, the value of P rize the prize for ghting lnEt Πt+1 is typically unobserved, and so it enters the error M eas. term et in Equation 3, yielding an estimate of βOpp , which includes the eect of P rize P rize endogenous changes in lnEt Πt+1 , namely ∂lnVt /∂lnΠt+1 , which is usually positive. As we show in the next section, temporary and anticipated seasonal shocks satisfy ∂lnΠP rize t+1 /∂lnWt = 0 such that βOpp M eas. T rue = βOpp = −φ. In contrast, for persistent P rize shocks, ∂lnΠt+1 /∂lnWt > 0, biasing estimates of the opportunity cost mechanism M eas. T rue upward toward zero: βOpp > βOpp . 2.1.3. Two-Period Analytical Greed Model. The simplest version of this model has only two periods. Conict, labor, and production occur in the rst period t, and the outcome of conict is decided at the end of this period. In the second period t + 1, there is peace and all of the insurgent's time is spent working (Lt+1 = 1). In the two-period model, the insurgent's gain from winning are prots Πt+1 from production W in L P rize in the second period: Ut+1 = Πt+1 + Wt+1 × 1 and Ut+1 = Wt+1 × 1, so Πt+1 ≡ W in L Ut+1 − Ut+1 = Πt+1 . However, prots and wages are driven by shocks to productivity θ. Πt+1 = (1 − α)Yt+1 = (1 − α)θt+1 such that lnEt ΠP rize t+1 = ln(1 − α) + lnEt θt+1 . We assume that we start with a very low fraction of time spent on violent activities, 6 so lnWt ≈ lnα + lnθt (wages are approximately proportional to productivity). Then Equation 2 can be rewritten as (4) lnVt ≈ κ2 − φ lnθt +φ lnEt θt+1 . ”W age” ”P rize win” Proposition 1 Consider an autoregressive process for productivity lnθt+1 = κ3 + 7 T rue ρlnθt , where βOpp = −φ as in Equations 2 and 4. Then the measured strength of the opportunity cost mechanism will M eas. T rue (A) equal its true strength βOpp = βOpp = −φ in the event of temporary shocks, ρ = 0 (such as seasonal shocks); M eas. (B) be zero (βOpp = 0) with permanent shocks ρ = 1 (i.e., lnθt+1 = lnθt ); and 6That Lt ≈ 1 so Lα is, −1 ≈ 1 and Wt = αθt Lt α−1 ≈ αθt . t 7κ = (1 − ρ)lnθ ¯ is the log mean productivity. ¯, where lnθ 3 8 M eas. T rue (C) be biased upward (βOpp > βOpp ) by persistent shocks ρ > 0. Proof: Substituting the shock process, Equation 4 becomes lnVt ≈ (κ2 + κ3 ) − φ(1 − M eas. ρ)lnθt , and hence βOpp = dlnVt /dlnWt = dlnVt /dlnθt = −φ(1 − ρ). The result follows from substituting dierent values of ρ. Proposition 1A shows our main motivation for using seasonal shocks: variation in labor market outcomes driven by temporary shocks, including seasonal shocks, yield the true strength of the opportunity cost mechanismeven if variation in the prize of ghting is unobserved. This argument is related to Chassang and Padro-i-Miquel's (2009) nding that only temporary negative income shocks will increase violence with a dierent model. Proposition 1B is a restatement of Fearon's (2008) and Chassang and Padro-i- Miquel's (2009) result that permanent changes in the level of economic development, or income, increase both the opportunity cost of violence and the spoils of war, leaving the level of violence unchanged. Finally, Proposition 1C applies to many shocks used to assess the role of opportu- nity cost considerations, such as many commodity price shocks, which are well known to be highly persistent. It states that using variation in persistent shocks will gener- ate estimates of the strength of the opportunity cost mechanism that are too small in absolute value. This might explain the wide variety of estimates of the strength of the opportunity cost mechanism in the literature. We quantify this result in the next subsection. 2.1.4. Dynamic Quantitative Greed Model. The second submodel is an innite horizon one and allows for richer dynamics but can only be evaluated quantitatively. Just like in the two-period model, the part-time insurgent allocates time between working and ghting in the rst period t and the outcome of conict is decided at the end of the rst period. If the insurgents win, we assume that victory is decisive and there is peace forever with all time allocated to working (L = 1). The expected future utility of the victorious insurgent UtW +1 in is the present discounted value of future prots from ∞ i the capture of the natural resource ( i=1 β Et Πt+i ) plus the present discounted value ∞ i of future labor income i=1 β (Et Wt+i ). If the insurgents lose, the game resets and the part-time ghter allocates time be- tween working and ghting again, with the utility of the out-of-power ghter UtL +1 dened recursively in Equation 1. The violence continues forever or until the insur- P rize gents win. The gain from winning in the dynamic model is Πt+1 ≡ UtW in L +1 − Ut+1 = ( ∞i=1 β i− 1 Πt+i + ∞ i=1 β i− 1 Wt+i ) − UtL +1 . That is, the value of the prize of winning 9 ∞ today depends on the present value of future prots i=1 β i Et Πt+i but also on other 8 terms that do not have a closed-form solution. Now that we have a full dynamic model, we can fully categorize shocks as seasonal (alternating high and low) versus persistent shocks following an AR(1) process (Equation 5): (5) ¯ + χ1(t = k ), k = 1, 3, 5.. AR(1) : lnθt+1 = κ3 + ρlnθt + et+1. Seasonal : θt = lnθ The model is not analytically tractable, so instead we simulate data when produc- tivity is driven by persistent shocks (like commodity price shocks) or by anticipated temporary seasonal variation, and we estimate a regression of simulated violence on simulated wages. The model is solved by log-linearizing the winning and losing value functions and the rst-order conditions around a non-stochastic steady state (where ¯ ∀t).9See θt = θ Appendix 1.2 for a list of equations. Calibration. We calibrate the strength of the true opportunity cost mechanism −φ = −3 to match the estimated elasticity of violence with respect to wages in Colombia driven by coee price shocks, using data from Dube and Vargas (2013) 10 and estimated using indirect inference. This means that, other things equal, a 1% increase in wages should lead to a 3% fall in time spent allocated to violence. We calibrate relative strength of the insurgent forces ψ so that given the other parameters, 11 the insurgents have a 0.1% chance of winning each quarter. See Appendix 1.3 for further details on the calibration. 8More formally, note that in the innite horizon setup, Equation 1 is a Belman equation (with some abuse of notation for consistency with the two-period model). U L (θ) is the value (discounted W in lifetime expected utility) of an out-of-power ghter. The state is θ . U (θ) is the value of the W in part-time ghter who is in power, which can be written recursively as U (θ) = θ + βEU W in (θ ), where a prime denotes the next period's value. 9This is a rst-order Taylor series approximation of the model's FOCs and value functions but with respect to logXt rather than to Xt . 10The indirect inference approach involves choosing φ so that estimating Equation 3 in the sim- M eas. ulated and actual data generates the same βOpp . Such an approach is needed because we know M eas. βOpp will be biased upward. We cannot calibrate using our empirical results, as we do not observe wages at seasonal frequencies. See Online Appendix 1.2.1. 11That is, ψ = 0.0065. Other parameters: α = 0.5, β = 0.99 and ¯ = 1. θ 10 A1. Persistent shocks ( =0.966); "Greed" Model-Generated Data B. Estimated Strength Opportunity Cost using Simulated Data 0.1 0 True Opportunity Cost (- )=-3 Elasticity of Violence to 1% increase in Wage 0 Regression Coeff. Meas. (AR(1) Simulations) -0.5 Opp Regression Coeff. Meas. Opp (Seasonal Shocks) [= True Opp Cost] -0.1 Persistence Coffee Prices in Data c =0.966 (Vertical Line) -1 Persistence Oil Prices in Data o=0.983 (Vertical Line) -0.2 5 10 15 20 25 30 35 40 -1.5 A2. Seasonal Labor Demand; "Greed" Model Generated Data 0.2 -2 0.1 -2.5 0 Violence (deviation from SS) -3 -0.1 Wages (deviation from SS) Prize from fighting (deviation from SS) [goes in error term as unobserved] -3.5 -0.2 5 10 15 20 25 30 35 40 0 0.2 0.4 0.6 0.8 1 Quarters Persistence of Labor Market Shock Shock (Quarterly) Figure 1. Panel A (LHS) shows the simulated data, and Panel B (RHS) shows the estimated coecient. Results. The results are best illustrated with an example shown in Figure 1. Panel A1 (top) of Figure 1 is simulated random data generated by shocks as persistent as coee prices (ρ = 0.966 quarterly). One can see that around quarter 10, a large shock raises wages substantially (black line), increasing the opportunity cost of ghting and resulting in a fall in time allocated to violence (blue line). However, as this shock is unanticipated and persistent, it also raises the value of the prize of winning (red line). The higher prize of winning increases violence, other things equal. Combining the higher wages and the larger prize, the graph shows that violence still falls overall but not as much as it should based on the opportunity cost mechanism. For example, by period 25, wages are up by 10%, but violence has only fallen slightly more than 20% rather than the 30% fall expected by the opportunity cost mechanism (−3 × 10%). Simulations driven by seasonal shocks are shown in Panel A2 (bottom) of Figure 1. The gure shows that wages go up and down with the seasons (black lines) and violence goes down and up (blue lines) with three times the volatility, as expected, based on −φ = −3. In contrast with results in Panel A1, the value of the prize from winning is unaected by the anticipated and temporary seasonal shock and is thus completely constant (red line). What does this mean for the estimated strength of the opportunity cost mechanism? M eas. Given that the value of winning is typically unobserved, researchers estimate βOpp in univariate Equation 3 with the prize of winning going in the error term et . Panel B 11 M eas. of Figure 1 plots βOpp estimated using simulated data generated by AR(1) shocks of dierent degrees of persistence (blue curve). The true strength of the opportunity cost M eas. mechanism (−φ = −3) is plotted in red. βOpp is close to 3 when shocks are almost M eas. temporary, while βOpp increases sharply as shocks become persistent. Namely, as M eas. ρ → 1 (more permanent), βOpp → 0, as in Proposition 1B above. Unfortunately, the region on the right of the persistence parameter space with the largest bias is where a number of shocks studied in the literature reside, such as coee and oil price shocks (marked with pink and green vertical lines, respectively). This means that labor market shocks with the persistence of coee prices lead to an upward bias of around 1/4 (i.e. M eas. βOpp = −2.2 rather than T rue βOpp = −3).12 For oil prices, things are even worse due to higher persistence, as there would be an upward bias of almost M eas. 1/2 (βOpp = −1.7 rather than rather than T rue βOpp = −3).13 The large upward bias is able to rationalize why many studies nd mixed evidence of the opportunity cost mechanism using commodity prices (Blattman and Bazzi 2013) or even a positive impact for other types of shocks, such as those driven by development programs. In contrast, a regression run on seasonal shocks delivers almost exactly the true M eas. T rue. opportunity cost estimate of βOpp = βOpp = −3 (green circle), because the value of the prize of victory is kept constant. This nding motivates the empirical work in the next section. 2.1.5. Anticipated Shocks. The second advantage of seasonal shocks is that they are not only temporary but also anticipated . The unobserved value of the prize of winning in Equation 2 is the expected discounted value of future prots, much like a share price. Just as for shares, anticipated changes in future prots have less of an eect on the current prizeeven if those changes are persistentbecause they are already priced in. Hence the estimated size of the opportunity cost mechanism is close to unbiased using anticipated shocks, even when those shocks are persistent and the value of the prize is unobserved (see Appendix 1.1). 12To be clear, the results in Dube and Vargas (2013) for coee prices are correct but are possibly even stronger than estimated if there is ghting to control the revenues from coee plantations. Dube and Vargas (2013) argue that this channel is relatively weak because coee production is more labor intensive, and the earnings are dicult to appropriate. 13Moreover, our estimates of the persistence of commodity prices are conservativewith a dif- ferent sample and the Andrews (1993) median-unbiased estimator, Cashin et al. (2000) nd that M eas. commodity price persistence is often ρ = 1, suggesting βOpp = 0. 12 2.2. Grievance Model. Although the motivation for ghting is very dierent, a grievance model generates very similar results for seasonal and persistent shocks as the greed model presented above. The model's details are presented in Appendix 2, but we sketch the argument here. In the grievance model, there is no monetary benet from winning, but rather the insurgent holds a grievance which means she gains utility from time spent allocated to violence (V). The insurgent also needs to work at wage W to consume (C). We make the standard assumption that she likes both consumption and time allo- cated to violence but has diminishing marginal utility in both. Higher wages have two eects in the model: (i) they increase the opportunity cost of ghting (a substitution eect) which implies a reduction in V and (ii) they make the agent wealthier (higher consumption), which makes her want to increase V, as the grievance is now relatively more subjectively important (an income eect). As consumption is typically not controlled for in conict regressions, the income eect enters the error term, and univariate opportunity cost regressions with persis- tent shocks will be biased upward (towards zero). More specically, we can write the agent's rst-order condition in the grievance model in the same form as in Equation 4 but where φlnEt ΠP rize t+1 is replaced by φlnCt .14 Households smooth consumption over time by the PIH. Hence, persistent unanticipated increases in wages lead to an increase in φlnCt , which increases violence and leads univariate estimates of the opportunity cost mechanism to be biased upward toward zero. In contrast, consumption remains constant under temporary and anticipated sea- sonal shocks (households save temporary income shocks under the PIH), keeping φlnCt constant and yielding unbiased univariate estimates of the opportunity cost mechanism. Appendix 2.3 also includes a counterinsurgency information model of vi- 15 olence, which generates similar results through the same income eect mechanism. 14Here we assume log preferences for consumption, so σ=1 drops out. See Appendix 2 for the complete form. 15In that counterinsurgency information model, the level of violence depends negatively on the counterinsurgency information provided to the government. Households get paid to provide that information, which they view as snitching and intrinsically dislike providing. An increase in wages that increases the opportunity cost of ghting also means that the richer households want to pro- vide less counterinsurgency information, increasing violence and biasing univariate opportunity cost estimates. 13 3. Harvest and Conflict: Empirical Methodology The theoretical framework above suggests that seasonal shocks are well suited to gauge the true importance of the opportunity cost mechanism in conict settings. We now investigate this empirically in Iraq, Afghanistan and Pakistan. These three countries were chosen because they have long-lasting conicts, experience a large seasonal shock to labor demand due to the harvesting of wheat, and have ne-grained data on both harvest and conict intensity. This section describes our methodology and data, and Section 4 presents our main results. 3.1. Empirical Specication. Ideally, we would run a regression of time spent ghting on variation in monthly wages driven by harvest shocks, as suggested by our theoretical model. Unfortunately, in conict settings there are typically no detailed monthly panels of time use or local wages. Hence we estimate the reduced-form eect of the onset of harvest (and the amount harvested) on conict intensity. In this context, a negative coecient on harvest intensity is consistent with increases in local labor demand due to harvest reducing the attractiveness of ghting (the opportunity cost mechanism). Civil war conict intensity is typically measured as the number of conict incidents (attacks) in a particular subnational region over a certain period, as in Dube and Vargas (2013) and Berman et al. (2011). However, a key practical challenge is the wide variation in measures of conict intensity across countries, within countries, over time, and in dierent datasets. For example, descriptive statistics in Appendix Table A.4 suggest the mean number of attacks per district-year in Iraq ranges from 19 to 317, depending on the dataset and denition of attack (e.g., whether it involves casualties). In Pakistan, there are, on average, 24 terrorist/militant attacks per district-year, and in Afghanistan there are 345 attacks per district-year, again depending on the denition and dataset.The intensity of conict also varies substantially across years depending on geopolitical shocks. For these reasons, we seek to produce a normalized measure of attacks that is comparable across datasets and countries with widely varying conict intensity, with the number of attacks per district-year being the normalizing variable. In other words, our outcome of interest, %Attacksimt = Attacks imt Attacksit × 100, is the percentage of attacks in district i in month m in year t (Attacksimt ) relative to the total number of attacks in the same district in year t (Attacksit ). By construction, the mean of %Attacksimt 14 is close to 1/12 = 8.3%, though it can depart slightly from 1/12 when datasets start or nish mid-year. Our key independent variable seeks to capture when a region is in harvest and the intensity of harvest (as a proxy, the size of the shock to labor demand). It is constructed as the fraction of district i in harvest in a particular month m (Harvim ) interacted with the land area harvested (P rodi ). P rodi is based on pre-war wheat production data and is thus time invariant, but Harvim varies within each country due to local climatic conditions (though it is also based on pre-war data). Hence our independent variable Harvim × P rodi is the number of hundred square kilometers of wheat in harvest in district i, month m, and year t. Our main estimated equation is (6) %Attacksimt = αit + γmt + β (Harvim × P rodi ) + δ ximt + eimt , where αit is a district-by-year xed eect to account for district- and year-specic factors aecting conict, γmt is a month-of-the-year xed eect (e.g., June 2005), and ximt is a vector of monthly district characteristics such as temperature and precipi- tation. In all specications we also control for the eect of the timing of planting on 16 conict intensity (not shown). The parameter of interest is β, which captures the eect of harvest on conict intensity. Standard errors are clustered at the district level, which accounts for serial correlation in the error terms for that spatial unit. For Afghanistan, opium production is a key confounding factor but one that appears less common in areas where wheat is rain-fed rather than irrigated (discussed further in Section 3.2.1). Hence, for Afghanistan, we include separate variables for areas where wheat is predominantly rain-fed (our focus) or irrigated. That is, Harvim × P rodi RHarvim × RP rodi , which is the area of in Equation 6 is replaced by rain-fed wheat (100 square kilometers) in harvest in district i, month m, and year t. We add the analogous variable for irrigated wheat IHarvim × IP rodi as a control. 16Planting is also a temporary and anticipated labor demand shock, and indeed we nd some evidence that conict intensity falls at planting (and with a larger area cultivated). However, the results for planting are not as strong or robust as those for harvest, so in the interests of brevity, we do not report them. This might be because planting is a less urgent task than harvest, and possibly it is less labor intensive. Our planting variable is constructed in the same way as the harvest variable, as the interaction between a local planting month dummy variable and the local area under cultivation. 15 3.2. Identication Strategy. There are two challenges to identifying the eect of temporary harvest shocks on conict intensity: reverse causality and omitted vari- able bias. Reverse causality is where changes in conict for other reasons (changes in strategy, geopolitical shocks, and blind luck in the fog of war) drive changes in harvest timing or intensity Harvim × P rodi . We rule out reverse causality because our harvest timing and intensity variables are constructed using pre-war data. More specically, wheat production is driven by a range of time-invariant agricultural suitability indi- cators, such as local climate, rainfall, and soil type. The pre-war harvesting calendar Harvim is also driven by a combination of geographic and climate factors. By using pre-war data on the size of the local harvest rather than contemporaneous data, we rule out (for example) conict aecting the ability of farmers to plant or harvest 17 wheat. Because harvest is very time-sensitivecrops will not be ready or will start to rotthere is a limited ability for conict to move the harvesting month away from that dictated by the pre-war calendar based on climate and geography. Perhaps a more important threat to identication comes from omitted variable bias, where variables that are correlated with harvest timing or intensity can also aect conict. In Section 2, we showed that in many dynamic models of conict, shocks to opportunity cost are correlated with other unobservables (e.g., the value of the prize of ghting), biasing estimates of the strength of the opportunity cost mechanism. We also showed this was not a problem for harvest shocks given that they were temporary and anticipated. Nonetheless, other omitted variables may remain. The rst way we deal with omitted variable bias is to include district-by-year xed eects αit . These xed eects will capture any time-invariant determinants of con- ict (e.g., strategic location) as well as most other trends in conict intensity in each district (e.g., how far away the district is from the front line). All remaining aggre- gate variation in conict intensity is removed by month-of-the-year xed eects γmt (e.g.,the ghting season in the country, aggregate commodity prices, and geopo- litical shocks). These xed eects mean that any remaining confounding variables must vary both at the district level and within the year, greatly reducing the set of potential confounding variables. Two of the variables that might meet these criteria are precipitation and tem- perature. Burke et al. (2009), Hsiang et al. (2013), and many others argue that 17Put another way, the amount harvested in a location in a given month has an endogenous com- ponentdue to conict disrupting agricultural production (for example)as well as an exogenous component due to the local climate, geography, and soil suitability. We only consider variation due to the second exogenous component. 16 temperature tends to increase violence. Miguel et al. (2004) and others argue that 18 rainfall shocks aect conict in Africa. For this reason, all specications include controls for local temperature and precipitation (contemporaneously) in ximt .19 The nal general threat to identication is where the timing and intensity of har- vest does aect conict but not through the opportunity cost mechanism posited. For example, it could be that insurgents try to capture the harvest during transportation, leading to an increase in conict just after the harvest month, which makes conict look lower during harvest by comparison. We address this threat by estimating dy- namic specications for conict intensity in the months around harvest. 3.2.1. Opium Production in Afghanistan. One identication challenge specic to Afghanistan is the presence of opium production, as opium poppy grows in many of the same places as wheat and has a similar harvesting calendar. It is well known that opium produc- tion funds the Taliban (and other insurgent groups): they tax opium harvest and transportation, and are also involved in opium storage, trading and barter (Peters 20 2009). This increases the ability of the Taliban to conduct attacks around har- vest time, and indeed Piazza (2012) nds that Afghan provinces with higher opium production have higher levels of terrorist attacks. Thus opium production is a con- founding factor that would bias our estimates of the strength of the opportunity cost mechanism toward zero. We address this concern in two ways. First, as mentioned above, we take advantage of the fact that poppydue to its high valueis mainly cultivated in irrigated areas, 21 whereas wheat is cultivated in both irrigated and rain-fed ones. As such, we focus on the coecient of rain-fed wheat as the cleanest measure of the opportunity cost 18In our context Appendix Figure A.4 shows how the onset of harvesting (roughly from May to July) is indeed accompanied by an increase in temperature and a decrease in precipitation in Iraq. Without controls, these factors would tend to bias our estimated strength of the opportunity cost. 19Our precipitation data (in millimeters) and average temperature (in degrees Celsius) are from Willmott and Matsuura (2015) for Iraq, Afghanistan, and Pakistan. See Appendix Figure A.7. 20 [Opium] proceeds from ushr [a 10% Islamic tax applied to harvest], as well as commodities collected in barter deals, appear to supply village and district-level Taliban with the bulk of their operational needs, everything from salaries for ghters and transport, fuel, food, weapons, and explosives. (Peters 2009: 19). While the Taliban announced a ban on opium cultivation in 2000-01, this was temporary. By 2002, opium production increased almost 10-fold (relative to 2001), making Afghanistan again the world's largest producer (Peters 2009). According to the UNODC (2010: 5) planting opium poppies is six times more protable than growing wheat. 21See Appendix Figure A.10 portraying the relative prevalence of wheat versus opium in rain-fed versus irrigated areas according to the UNODC (2004) report. 17 mechanism. Second, we limit the sample in Afghanistan to those districts for which the median opium production during our sample period is zero. This does not mean there was no production of poppy in that district but rather that it occurred in less 22 than half of the sample months. 3.3. Conict Data, Samples, and Background. Given the well-known dicul- ties in measuring the intensity of violence, we use dierent datasets on violence for each conict setting to avoid assigning disproportionate weight to a single data col- lection procedure. We use both very precisely geolocated datasets of attacks (e.g., latitude, longitude) and datasets where attacks are aggregated to the district level. However, in the former case we always aggregate data to the district level to reduce measurement error. Additional details on data sources and variables used are de- scribed in Appendix 3.1. All of our datasets indicate that violence is geographically 23 and temporally concentrated in particular district-years. As discussed above, our main dependent variable is the percentage of attacks in Attacksimt district i and month m relative to its total in year t ( Attacksit × 100), which is not dened if the total number of attacks in a district-year is zero, and can be noisy when the number of attacks is very low. As such, we limit the sample to either district- years that are at least above the median of conict intensity or above the mean (if the median is zero). 3.3.1. Iraq. Between 20032011, Iraq experienced a civil conict against coalition forces and along sectarian lines. The intensity of the conict, its large contingent of 24 part-time ghtersat least 200,000 according to some estimates coupled with the strong reliance on wheat cultivation in rural areas makes it an ideal setting to explore the importance of seasonal labor markets for violence intensity. For Iraq, we use four conict datasets over a common sample of 20042009. Coali- tion forces invaded Iraq in March and April 2003. However, the initial invasion was a regular interstate war and not a subnational conict, which is the focus of this paper, so we start our sample in 2004. Our period of study covers the initial insurgency against the coalition forces during 20042006 as well as its evolution into a sectarian 2248% of district-year observations between 20042009 (WITS data) and 70% of observations between 20082014 (SIGACTS data) meet this citeria. 23As noted in Appendix Table A.4, the median number of attacks in a district-year across datasets is zero or one. 24The same source estimates around 40,000 core or full-time ghters. See (last accessed February 2, 2020): https://www.globalsecurity.org/military/ops/iraq_insurgency.htm. 18 conict following the bombing of the Shia al-Askari mosque in February 2006. Our sample includes the surge of US troops in 20072008. We nish our sample at the end of 2009, in anticipation of the withdrawal of US combat forces from Iraq by the middle of 2010. Although the US maintained a large number of troops until the end of 2011, they were used mostly to support the Iraqi military, and 20102011 reported 25 lower conict intensity than earlier years. We use four datasets in Iraq because each one provides an incomplete (and noisy) measure of the conict intensity. Our rst data source is the Iraq Body Count dataset (IBC), which tracks civilian deaths due to sectarian conict (the vast majority), insurgent-coalition conict, and coalition forces directly. The IBC data are collected by a non-prot organization, are based on media and administrative sources and are sourced from the Empirical Studies of Conict (ESOC) Project's conict data reposi- tory at Princeton University (https://esoc.princeton.edu/). Our main results use the total number of deadly events, but we also consider disaggregated data by perpetra- tor. One advantage of the IBC data is that attacks do not have to be witnessed by coalition forces, though the disadvantage is that many serious attacks only involve combatants (not civilians) and so would not be included in the IBC. Our second data source for Iraq is the World Incident Tracking System (WITS), which exclusively focuses on terrorism events based on media reports. The WITS data were collected by the National Counter-terrorism Center for the US State Department until they were discontinued in 2012 (they were also download from ESOC). We focus on total attacks across all categories, though we also consider types of attacks that might be more labor intensive (armed attacks) or less labor intensive (bombings). The WITS dataset captures terrorist attacks against civilians and noncombatant targets that are both premeditated and politically motivated, which may end up excluding many attacks against coalition forces (and is limited to those reported in the media). Our nal two data sources are based on coalition reports of events of signi- cant activity (SIGACTS). The rst one is more general/aggregate from Berman et al. (2011a) (SIGACTS-BFS), which includes all enemy activity (deadly and not deadly). We also use a version of the same underlying dataset disclosed by Wikileaks to The Guardian that focuses on signicant activity that results in deaths on either 26 side of the conict (SIGACTS-WIKI). Berman et al. (2011a) report their results 25Our sample does not cover the Iraqi civil war from 2014 following the rise of ISIL. 26Last accessed February 2, 2020: https://www.theguardian.com/news/datablog/2010/oct/23/wikileaks- iraq-data-journalism#data. 19 weighted by population, so we report both weighted and unweighted results using the SIGACTS-BFS dataset. We also report heterogeneity by the type of signicant attack: more labor intensive (direct re), less labor intensive (explosives/IEDs), and capital intensive (indirect re). The advantage of the SIGACTS datasets is that they mostly record conict incidents that are closest to those in theorycombat between the government and an insurgent groupand the data are ocial. The disadvantages are that they only cap- ture events witnessed by coalition forces, and so they would undercount, for example, sectarian conict. Moreover, the quality of this dataset could greatly vary across the combat units collecting it (Berman et al. 2011a: 790). 3.3.2. Afghanistan and Pakistan. Following the 9/11 terrorist attacks, the US and its allies invaded Afghanistan in late 2001 to overthrow the Taliban, which had been harboring Al-Qaeda. Many of the Taliban and Al-Qaeda ghters were not captured but instead escaped into rural or mountainous areas or moved across the porous Pakistani border. In Afghanistan, the Taliban launched an insurgent movement to regain power. The insurgency waged asymmetric warfare against the US and its allies, known as the International Security Assistance Force (ISAF), as well as against members of the Afghan military and the Afghan government. The ISAF ceased 27 combat operations at the end of 2014. Afghanistan is a good location to test the opportunity cost mechanism through harvest given the Taliban's reliance on rural part-time ghters. Most of the Taliban recruits come from poor madrassas, motivated by local grievances, and participate only on a part-time basis due to their work as farmers or laborers (Qazi 2011: 10). Taliban cells are composed of around 10 to 50 part-time ghters (Afsar et al.2008: 65) who periodically gather to launch attacks but then return to work as laborers or farmers. Ideally, our Afghanistan sample would cover the whole ISAF period from 20022014, but in practice our sample is dictated by sample period of publicly available data. Our rst dataset for Afghanistan is the WITS, which focuses exclusively on terrorism events based on media reports. It is available for 20042009 and was downloaded from the ESOC repository. Like the WITS data for Iraq, our default variable is total attacks, but we also report results for more labor-intensive types of attacks (rearms or attacks excluding bombings) or less labor-intensive attacks (bombings). 27Since 2015, the NATO-led mission Resolute Support has been focused on training and sup- porting the Afghan military. 20 The second dataset for Afghanistan is on signicant activity, reported by the ISAF troops in Afghanistan from 2008 to 2014. This dataset is similar to the SIGACTS dataset for Iraq, though it is a little more detailed. Our main variable is total enemy attacks, but as for Iraq we also use direct re as a more labor-intensive attack type and use bombings as a less labor-intensive type. The data were declassied in 2014 and can 28 be downloaded from Centcom FOIA library les or from Vincent Bauer's website. In the appendix we consider a placebo test using counterinsurgency actions. For Pakistan, our sample starts in 2002 following the arrival of the ISAF in neigh- boring Afghanistan, and the start of a rising trend of violence in a number of provinces in Pakistan (Shapiro and Gulzar 2012). Our sample ends in 2010 due to data avail- ability. As in the other contexts, conict in Pakistan is concentrated geographically, though unlike Iraq, it is of lower intensity, perpetrated by several groups, and concen- trated in urban areas (Blair et al. 2013: 32). Our rst data source for Pakistan is the BFRS Political Violence Dataset (Bueno de Mesquita et al. 2015), which is available at the district level over 20022010 and was collected from local newspaper reports (as opposed to only those in English). This dataset codes all instances of political violence, including those perpetrated by militant and state forces, and classies them into whether they are conventional (more labor intensive) or asymmetric (e.g., less reliant on laborIEDs, suicide bombs). Our main variable from the BFRS dataset is the number of militant attacks, which can be dened by either attack type or attack perpetrator. We also present results for more labor-intensive attacks (conventional attacks), less labor-intensive attacks (asymmetric attacks), and placebo tests using attacks by the state or foreign forces. Our second dataset for Pakistan's terrorist attacks is the Global Terrorism Dataset (GTD) for the same period, which includes events dened as the threatened or actual use of illegal force and violence by a non-state actor to attain a political, economic, religious, or social goal through fear, coercion, or intimidation. It is available over 29 20022010 and was collected by the University of Maryland. Our main variable is total attacks, though we also consider labor-intensive attacks (total attacks excluding bombings) and less labor-intensive attacks (bombings). Both the BFRS and GTD datasets were downloaded from the ESOC data repository. 28https://stanford.edu/~vbauer/data.html. 29See http://start.umd.edu/gtd/. Starting in 2012, GTD became the dataset used by the US State Department to report terrorist incidents, as WITS was discontinued. Because WITS is similar to GTD in sources and motivation, for each country, we either use WITS or GTD. 21 3.4. Harvest Calendar and Harvest Intensity. Based on the intensity of wheat cultivation and the timing for harvesting for each district-month, our main explana- tory variable is the interaction of the size of the area harvested with a dummy variable for whether that district that is in harvested in that month. The right panels of Fig- ures 24 show the distribution of this variable for Iraq, Afghanistan, and Pakistan such that each column represents the total weighted number of square kilometers at har- vest (or planting) in a given month. The intensity of harvesting varies across months within the year. For Iraq (Figure 2), most districts harvest wheat in MayJune, yet some areas also harvest as early as April or as late as July. Wheat Cultivation Intensity. Wheat intensity is measured in hundreds of square kilometers and is calculated by the Food and Agriculture Organization (FAO) as the historical average of the period 19601990, which clearly precedes our sample. The left panels of Figure 24 illustrates the data by showing the raw images provided 30 by the FAO Global Agro-Ecological Zones (GAEZ v3.0) of the intensity of wheat harvesting at the 5x5 arc-minute grid cell level. For Afghanistan, the images are di- vided into areas of rain-fed wheat and irrigated wheat (for an example, see Appendix Figure A.5). This ne-grained information is then aggregated at the district level to calculate the intensity with which a given district is in harvest. Figure 2. Wheat Production (L) and Calendar (R) in Iraq 30Available at: http://www.gaez.iiasa.ac.at/. 22 Figure 3. Wheat Production (L) and Calendar (R) in Pakistan Figure 4. Wheat Production (L) and Calendar (R) in Afghanistan Harvesting and Planting Calendars. To calculate the timing of planting and harvest, we again rely on the FAO-GAEZ data, which provide high resolution maps for the start and length of the wheat growing cycle depending on the level of inputs 31 used (high, medium, or low) and water sources (rain-fed or irrigated by rivers). Specically, we combine the calendar start day of the cycle (e.g., day 30, meaning late January) with how long wheat is expected to grow in that grid cell (e.g., 60, 31For an example and additional details, see Appendix Figure A.6. 23 90, or more days) to identify the month of harvest (and planting). Our planting and harvesting indicator thus generally takes the value of 1 in the month the growing cycle starts and ends, respectively (and zero otherwise). However, in some grid cells there can be more than one harvesting and planting month, depending on the inputs used. For example, while high-input wheat in a grid cell might be harvested in June, low-input wheat in the same grid may be harvested in May. For these very few cases, we created an average harvest indicator by dividing equally across these months (e.g., weighting by 0.5 if there are two months of harvest or 0.33 for three months). As with the wheat intensity data, we then aggregate this variable to the district level, as these encompass numerous 5x5 arc-minute grid cells. 4. Harvest and Conflict: Empirical Results In this section we present empirical estimates of the extent to which harvest reduces conict intensity. We present the results for Iraq rst in more detail (Section 4.1) and then the results for Afghanistan (Section 4.2) and Pakistan (Section 4.3) more concisely, as they are broadly similar. In all three countries we nd that harvest tends to reduce conict intensity. 4.1. Iraq. Table 1 presents our main empirical results of the eect of harvest on conict intensity in Iraq. In sum, four of our ve estimates suggest that the onset of harvest (and greater harvest intensity) leads to a statistically signicant reduction in the monthly share of attacks. In column 1, our dependent variable is the monthly share of deadly events resulting in civilian deaths (relative to the district-year total), taken from the IBC dataset. The coecient of interest on harvest intensity is 1.58, signicant at the 1% level. This can be interpreted as follows: an increase of 100 square kilometers of wheat production at harvest leads to a reduction in the share of deadly episodes by around 1.6 percentage points. Given that the average wheat cultivation intensity per district is 1.26 hundred square kilometers in our sample, and the mean share of attacks in a month is around 9%, this represents a 22% reduction (−1.58 × 1.26/9.0). In column 2, which uses total terrorist attacks from the WITS dataset, the estimated coecient is slightly smaller such that an extra 100 square kilometers of wheat at harvest reduces the share of attacks in the month by around 1ppt (signicant at the 5% level), representing a 15% fall in the number of monthly attacks. Column 3, which uses the monthly share of deadly enemy actions recorded by coalition forces (SIGACTS-WIKI), generates 24 almost identical results to the IBC dataset in column 1 (also signicant at the 1% level). Columns 45 present results using the monthly share of total attacksincluding those that are not deadlyas recorded by coalition forces and used in Berman et al. (2011a). Column 4 presents unweighted estimates, as in the rest of the paper, with 32 column 5 reporting estimates weighed by population, as in Berman et al. (2011a). The weighted estimates indicate that an extra 100 square kilometers of wheat at harvest reduces the monthly share of attacks by around 0.5ppts, signicant at the 5% level (a 6% fall in the number of monthly attacks). The unweighted estimateswhile still negative in signare smaller than absolute value and are insignicant. The fact that both sets of point estimates are smaller in absolute value than those in columns 13 could indicate that terrorist attacks and more serious incidents (resulting in casualties) are more sensitive to labor demand shocks than broad categories of all events used by Berman et al. (2011a). 4.1.1. Dynamics. The estimates in Table 1 indicate that conict is lower at harvest relative to the rest of the year. But this dierence does not tell us whether harvest reduces conict in an absolute sense or whether insurgents conduct the same number of overall attacks but simply delay them until the part-time ghters return from harvest (insurgents could also bring forward attacks, in anticipation of upcoming reductions in ghting strength). Moreover, a relative fall in conict intensity at harvest time can also mask direct eects of harvest on conict intensity through channels other than the opportunity cost mechanism. For example, conict could be lower at harvest time, not because of the opportunity cost mechanism but because violence spikes after harvest, as its 33 proceeds are used to nance insurgent activity. Alternatively, insurgents could try to capture the harvest itself while it is being stored and transported. To test these hypotheses, Table 2 presents the estimates from a full dynamic spec- ication where conict intensity is regressed on harvest with two leads and lags. If conict is simply moved to surrounding months, we would expect to see positive and signicant leads and lags. Alternatively, if the proceeds of harvest were nancing 32Note that all other estimates are similar when weighting by yearly district population (see Appendix Table A.6). But because we lack similar population data for Pakistan and Afghanistan, we prefer the unweighted estimates for consistency. 33Likewise, the government could use taxes on harvest to nance counterinsurgency operations, leading to more violence. 25 Table 1. Seasonal Labor and Conict Intensity in Iraq (20042009) (1) (2) (3) (4) (5) Data Source: IBC WITS SIGACTS-WIKI SIGACTS-BSF SIGACTS-BSF Dep. Variable: Deadly Total Deadly Enemy Enemy Enemy Events Attacks Activity Activity Activity Pop. Weight Harvimt × P rodi -1.583*** -0.975** -1.526*** -0.213 -0.452** (0.584) (0.487) (0.466) (0.169) (0.181) Observations 4,110 4,416 4,680 3,215 3,175 District X Year FEs Yes Yes Yes Yes Yes Time Fixed Eects Yes Yes Yes Yes Yes Clusters 89 88 96 66 66 Mean DV 8.978 8.333 8.333 8.989 8.472 Mean Wheat Intensity 1.262 1.252 1.185 1.168 1.166 Avg. Eect % -22.24 -14.64 -21.69 -2.762 -6.223 Controls Temperature, Precipitation, P lantingimt × P rodi Sample Jan2004- Jan2004- Jan2004- Feb2004- Feb2004- Jun2009 Dec2009 Dec2009 Feb2009 Feb2009 Clustered robust standard errors at the district level are in parentheses. Wheat Intensity P rodi is measured in 100 square kilometers. DV is in monthly %. *** p <0.01, ** p <0.05, * p <0.1 conict (or fought over), only the lags of harvest would be positive and signicant. However the estimates in Table 2 suggest that, in most cases, there is only a signif- 34 icant fall in conict at harvest time. In fact, in no dataset do we see signicantly higher levels of conict immediately before or after the conict, which rules out the 35 possibility that harvest in some ways nances conict or creates it. 4.1.2. Tactics and Perpetrators. In addition to looking at aggregate measures of insurgent activity, we also examine the types of attacks employed by insurgents. Our 34In the IBC dataset, there is some evidence of a reduction in conict the month before and after harvest, but it is only signicant at the 10% level and is half the size of the reduction in the harvest month itself. This may be due to some measurement error in the timing of harvest, which can vary slightly from year to year. 35This pattern also makes other explanations unlikely, for example, that the drop in conict is driven by insurgents letting the harvest take place as a way to curry their favor. If that were the case, we would expect higher levels of conict before and after the harvest, and there would also be no dierence in the type of attack (e.g., labor intensive versus others). 26 Table 2. Dynamic Specication in Iraq (20042009) (1) (2) (3) (4) (5) Data Source: IBC WITS SIGACTS-WIKI SIGACTS-BSF SIGACTS-BSF Dep. Variable: Deadly Total Enemy Enemy Enemy Events Attacks Activity Activity Activity Pop. Weighted Harvim−2 × P rodi -0.54 -0.10 -0.54 0.00 -0.03 (0.80) (0.68) (0.59) (0.16) (0.14) Harvim−1 × P rodi -1.07* -0.61 -0.06 0.24 0.15 (0.55) (0.55) (0.70) (0.15) (0.14) Harvim × P rodi -2.02*** -0.94 -1.76** -0.30* -0.57*** (0.72) (0.60) (0.73) (0.16) (0.16) Harvim+1 × P rodi -1.39* -0.09 -0.53 0.15 0.01 (0.75) (1.10) (0.84) (0.28) (0.34) Harvim+2 × P rodi 0.85 0.05 -0.12 0.21 0.14 (0.87) (0.44) (0.66) (0.24) (0.27) Observations 4,110 4,416 4,680 3,215 3,175 District X Year FEs Yes Yes Yes Yes Yes Time Fixed Eects Yes Yes Yes Yes Yes Clusters 89 88 96 66 66 Mean DV 8.978 8.333 8.333 8.989 8.472 Mean Wheat Intensity 1.262 1.252 1.185 1.168 1.166 Avg. Eect -28.45 -14.07 -24.98 -3.917 -7.859 Controls Temperature, Precipitation, P lantingimt × P rodi Sample Jan2004- Jan2004- Jan2004- Feb2004- Feb2004- Jun2009 Dec2009 Dec2009 Feb2009 Feb2009 Clustered robust standard errors at the district level are in parentheses. Wheat Intensity P rodi is measured in 100 square kilometers. DV is in monthly %. *** p <0.01, ** p <0.05, * p <0.1 idea is that insurgents might change tactics depending on the availability of labor ver- sus other inputs (Bueno de Mesquita 2013). Specically, we focus on labor-intensive attacks: those that require greater manpower to be carried out, e.g., direct/armed attacks involving groups of individuals using small arms or rocket-propelled grenades (Condra et al. 2018: 3208). The rst three columns of Table 3 replace total attacks in our main specication with measures of labor-intensive attacks (which vary by dataset): armed attacks from the WITS dataset (column 1), direct re resulting in casualties from the SIGACTS-WIKI datasets (column 2), and all direct re attacks 27 36 from the SIGACTS-BSF dataset(column 3). In each case, harvest results in a re- duction in the monthly share of labor-intensive attacks, which is consistent with the opportunity cost mechanism. In contrast, asymmetric attacks with no exchange of re generally have lower man- power requirements (such as bombings and IEDs). These type of attacks can act as a placebo test for our mechanism. Appendix Table A.7, columns 24 shows that, indeed, for a range of less labor-intensive attacks, harvest has no eect on conict in- 37 tensity. Combined with the ndings in Table 3 (columns 13), these results suggest that harvest is more likely to reduce labor-intensive attacks and not other attacks, which is what the opportunity cost mechanism would predict. The IBC dataset does not record the type of attack but does record the attack's perpetrators. Our theory suggests that only the capacity of non-state actors like insurgents or sectarian militias are likely to be constrained by a shortage of part- time labor at harvest. In contrast, the US-led coalition forces use only professional full-time soldiers, and so their ghting capacity will be unaected by harvest. The nal two columns in Table 3 show that a reduction in conict intensity at harvest is driven by insurgent groups and sectarian violence, consistent with the opportunity cost mechanism. However, there is also some weak evidence that harvest also reduces the number of attacks by coalition forces (Appendix Table A.7, column 1). This could also be due to the coalition responding to insurgent attacks, and so a reduction in 38 the latter also aects the former. 4.1.3. Corroborating Evidence. In this subsection, we discuss some suggestive ev- idence in favor of the eect of harvest on conict intensity through the opportunity cost mechanism. Employment and Harvest. The mechanism in our theoretical model suggests that harvest aects conict intensity through the demand for labor. However, there 36The attack variable is the percentage of total annual labor-intensive attacks in a district that occur in a given month. 37The results for capital intensive attacksindirect re in the SIGACTS datasetare mixed; see Appendix Table A.7, columns 56. In the SIGACTS-BFS dataset, (population-weighted) harvest signicantly reduces the intensity of indirect re, but this is not robust to indirect re resulting in casualties in the SIGACTS-WIKI dataset. 38For example, the First Battle of Fallujah (Operation Vigilant Resolve) was launched in April 2004 in response to insurgent attacks that killed four Blackwater US security contractors. If the insurgent attacks did not happen due to harvest, then likely the US-led coalition response would not have happened either. 28 Table 3. Heterogeneity by Attack Type and Perpetrator (Iraq 20042009) (1) (2) (3) (4) (5) Data Source: WITS SIGACTS-WIKI SIGACTS-BSF IBC IBC Labor-Intensive Types Perpetrator Dep. Variable: Armed Deadly Direct Direct Insurgent Sectarian Attacks Fire Fire Attacks Attacks Pop. Weighted Harvim × P rodi -1.137* -1.033** -0.727** -0.905* -1.530** (0.640) (0.421) (0.291) (0.521) (0.681) Observations 3,828 4,320 3,229 1,494 3,948 District X Year FEs Yes Yes Yes Yes Yes Time Fixed Eects Yes Yes Yes Yes Yes Clusters 84 93 69 34 86 Mean DV 8.333 8.333 8.486 8.568 8.992 Mean Wheat Intensity 1.294 1.181 1.195 1.264 1.254 Avg. Eect % -17.66 -14.64 -10.24 -13.35 -21.34 Controls Temperature, Precipitation, P lantingimt × P rodi Sample Jan2004- Jan2004- Feb2004- Jan2004- Jan2004- Sample Dec2009 Dec2009 Feb2009 Jun2009 Jun2009 Clustered robust standard errors at the district level are in parentheses. Wheat Intensity P rodi is measured in 100 square kilometers. DV is in monthly %. *** p <0.01, ** p <0.05, * p <0.1 are no monthly panel data on employment or wages in our conict settings to show this directly. Nonetheless, there is suggestive evidence from the 2006 World Bank Living Standards Measurement Survey (LSMS) that rural agricultural employment increases at harvest (and also slightly at planting). Appendix Figure A.8 shows the dierence in the monthly probability of employ- 39 ment for rural agricultural workers relative to rural non-agricultural workers. These dierences roughly follow the harvesting calendar in Iraq even after controlling for a number of individual factors and governorate xed eects. This W-shaped pattern is consistent with the idea that harvest aects conict by boosting demand for labor. Migration. Our test of the opportunity cost mechanism requires that harvest and reduced conict intensity happen in the same district. If instead people migrated to other districts to work on the harvest, the harvest location would be disconnected from the location of any reduction in violence, making it impossible to identify the 39More specically, the coecient is from a regression of monthly employment status on a dummy for agricultural workers (the excluded category are rural non-agricultural workers). The LSMS is a pooled cross-section, so the monthly employment status dummies are retrospectives. 29 mechanism, even it was present. Luckily, the LSMS above asks about migration, and only 3.67% of agricultural workers report an absence from home for an extended 40 period. Harvest Income. Another concern is whether a reduction in conict intensity could be due to the income received from harvest rather than to an increase in labor demand. Theoretically, our grievance model (see Appendix 2) predicts that the pay- ment from harvestwhich is temporary and anticipatedshould be saved and so it 41 will not aect the marginal utility of consumption or the valuation of the grievance. Practically, there is also likely to be some delay between harvest and payment, as most farmers sell their grain to the governmental Iraqi Grain Board, which only is- sues a receipt once all harvest is collected and stored in silos. The farmer then has to cash the receipt at a bank, adding further delays. Religious Calendar. The timing of religious events can aect the intensity of conict in either direction. All of our specications include month-year xed ef- fects, which will remove the aggregate eects of Islamic religious festivities common to all districts, even if their exact dates changes each year. However, month-year xed eects will be less eective if agricultural areas are more religious than others. Nonetheless, the 2008 Iraqi time-use survey suggests there is no signicant dierence in religiosity between agricultural and non-agricultural workers (see Appendix Figure A.9). 4.2. Afghanistan. The Taliban's reliance on rural part-time ghters suggests the opportunity cost mechanism is likely to be important in Afghanistan, and indeed we nd that conict intensity is much more sensitive to harvest than in Iraq (or Pakistan). However, as harvested areas are smaller, the average eect turn out to be broadly similar. Recall that as opium cultivation is a potential confounding factor in Afghanistan (see Section 3.2.1), we focus on rain-fed wheat (where opium cultivation is less common) and restrict the sample to districts with zero median cultivation. Our main results for Afghanistan are shown in the rst two columns of Table 4. Column 1 reports the eect of rain-fed wheat harvest on the monthly share of terrorist attacks over 20042009 from the WITS dataset. We nd that an extra hundred square 40There is little evidence that individuals switch occupations, which would allow harvest demand to be met by non-agricultural workers. In the LSMS, only 4.67% reported more than one occupation throughout the year. 41Even if it was spent, this would reduce the marginal utility of consumption and hence encourage an increase in violence. 30 kilometers of rain-fed wheat at harvest reduces conict intensity by 22 percentage points, more than 20 times the size of the eect in Iraq (or Pakistan). Districts in the WITS sample cultivate around 0.07 hundred square kilometers of rain-fed wheat, so harvest reduces conict intensity by about 18%, on average. Column 2 reports the eect of rain-fed harvest on the share of monthly total enemy attacks over 20082014 from the SIGACTS dataset. An extra hundred square kilometers of rain-fed wheat at harvesting areas reduces conict intensity by 4.4 percentage points, which reduces average conict intensity in the average district by 8%. The size of the coecient (4.4) is still much larger in absolute value than in Iraq or Pakistan but is smaller than using the WITS dataset and sample. Unlike our estimates for Iraq (and Pakistan), the WITS and SIGACTS samples cover dierent years, and the dierence in these estimates may be due to shifts in the intensity and nature of conict occurring between these two periods. We also nd the wheat harvest in irrigated areas has a smaller or insignicant eect on conict intensity, which (as agged above) is unsurprising given that coincident poppy harvesting will confound those estimates. Dynamics. In the rst two columns of Appendix Table A.8, we reestimate with two lags and two leads of harvest to test whether violence might be moved to the months surrounding harvest or if harvest might be funding conict. The results for both the WITS (20042009) and SIGACT (20082014) datasets suggest a signicant fall in conict intensity at harvest (similar in size to that in Table 4), with no eect before or after. Tactics and Perpetrators. As for Iraq (and Pakistan), we also explore the eect of harvest on dierent types of labor-intensive attacks, which are likely more sensitive to the supply of part-time ghters. As one can see in columns 35 of Table 4, harvest leads to reduction in attacks involving rearms (column 3, WITS dataset), attacks excluding bombings (column 4, WITS dataset), and direct re attacks (column 5, SIGACTS dataset). The size and signicance of harvest eects on labor-intensive attacks is similar to that on total attacks. As a placebo test, Appendix Table A.9 estimates the eect of rain-fed harvest on the number of less-labor-intensive attacks, such as those involving bombs (mainly IEDs). For both the WITS and SIGACTS datasets, the harvest variable is insignif- icantly dierent from zero. Column 4 of Appendix Table A.9 investigates the eect of harvest on the intensity of counterinsurgency operations (attacks by the coalition forces between 20082014). Unsurprisingly, the harvest variable is also insignicant, 31 Table 4. Main Results and Heterogeneity in Afghanistan (20042014) (1) (2) (3) (4) (5) Data Source: WITS SIGACTS WITS WITS SIGACTS Main Results Type of Attacks Dep. Variable: Total Enemy Firearm Attacks Direct Attacks Attacks Attack excl. Bombing Fire RHarvim × RainP rodi -22.176*** -4.461*** -29.618*** -28.865*** -5.010*** (5.903) (1.111) (8.895) (7.282) (1.832) IHarvim × IrrigP rodi 3.423 -1.570** 4.613 3.169 -0.682 (3.377) (0.737) (4.003) (3.774) (1.022) Observations 2,961 9,888 3,225 3,546 17,652 District X Year FEs Yes Yes Yes Yes Yes Time Fixed Eects Yes Yes Yes Yes Yes Clusters 88 171 98 100 264 Mean DV 8.815 8.333 8.775 8.799 8.333 Mean Wheat Intensity 0.0718 0.148 0.0727 0.0727 0.157 Avg. Eect % -18.06 -7.915 -24.55 -23.85 -9.443 Controls Temperature, Precipitation, P lantingimt × P rodi Sample Jan2004- Jan2008- Jan2004- Jan2004- Jan2008- Sep2009 Dec2014 Sep2009 Sep2009 Dec2014 Clustered robust standard errors at the district level are in parentheses. Wheat Intensity P rodi is measured in 100 square kilometers. DV is in monthly %. *** p <0.01, ** p <0.05, * p <0.1 given that the coalition uses only full-time professional soldiers and not part-time ghters. 4.3. Pakistan. As for Iraq and Afghanistan, our results for Pakistan suggest that the onset of harvest reduces conict intensity, mostly driven by labor-intensive attacks. The rst three columns of Table 5 present our main empirical results of the eect of harvest on conict intensity in Pakistan. The rst column shows that the onset of harvest is associated with a reduction in the monthly share of attacks when the identity of the militant perpetrator is known (ethnic, Islamist, sectarian, or other militant group) in the BFRS dataset, signicant at the 5% level. The second column shows that the onset of harvest is associated with a reduction in all types of attacks typically used by militants (terrorism, conventional attacks such as direct re or ambushes, or guerrilla attacks) using the BFRS dataset, signicant at the 1% level. The size of the estimated eects are similar in Columns 1 and 2, suggesting that an 32 increase of a 100 square kilometers of wheat production at harvest leads to a reduction in the monthly share of militant attacks in a district by around 0.6 percentage points. Given that the average wheat cultivation intensity per district is 2.53.1 hundred square kilometers (depending on the sample), these estimates imply a 20% fall in the monthly share of attacks at average harvest intensity. The coecient sizes are slightly smaller than those in Iraq though, as there is more wheat being harvested, the average eects is similar. The results in column 3 using a dierent datasetterrorist attacks using the GTD datasetare similar in size, signicance, and average magnitude as those using the BFRS dataset. Table 5. Main Results and Heterogeneity in Pakistan ( 20022010) (1) (2) (3) (4) (5) Data Source: BFRS BFRS GTD BFRS GTD Labor-Intensive Attacks Dep. Variable: Militant Militant Attack Total Conventional Attacks Perpetrator Types Attacks Attacks excl. Bombing Harvim × P rodi -0.652** -0.545*** -0.537** -0.542*** -0.760*** (0.283) (0.177) (0.243) (0.201) (0.218) Observations 2,100 2,652 2,700 2,268 2,400 District X Year FEs Yes Yes Yes Yes Yes Time Fixed Eects Yes Yes Yes Yes Yes Clusters 47 60 60 68 74 Mean DV 8.333 8.333 8.333 8.333 8.333 Mean Prod 2.456 3.149 2.921 3.504 3.701 Avg. Eect % -19.21 -20.61 -18.82 -22.78 -33.76 Controls Temperature, Precipitation, P lantingimt × P rodi Sample Jan 2002-Dec 2010 Clustered robust standard errors at the district level are in parentheses. Wheat Intensity P rodi is measured in 100 square kilometers. DV is in monthly %. *** p <0.01, ** p <0.05, * p <0.1 Dynamics. As for Iraq and Afghanistan, we examine the dynamic changes in conict intensity around harvest. Columns 35 of Appendix Table A.8 re-estimates the main specications in Table 5 with two monthly leads and two monthly lags of harvest intensity. In general, we nd that the largest and most signicant falls in conict intensity occur at harvest time, but the results are more mixed than for the other countries. For militant attacks in the BFRS dataset (classied by perpetrator, 33 column 3), the only signicant change in conict intensity is at harvest, with a coef- cient almost signicant at the 5% level (t-stat = 1.9). For total attacks in the GTD dataset (column 5), we also nd the largest and most signicant fall in conict inten- sity at harvest month. However, for the GTD dataset, we also nd some evidence of a reduction in harvest intensity in the month before harvest, which might be due to 42 harvest occurring on the border of two months. More important for our mechanism, there appears to be an increase of attacks following harvest for BFRS militant attacks classied by type (column 4), raising the possibility that harvest delays conict or could fund it. However, as we do not see any increase in post-harvest conict intensity in either of the other two specications (in columns 3 or 5), we do not put too much weight on this result. Tactics and Perpetrators. The nal two columns in Table 5 show results for labor-intensive attacks in Pakistan. For the BFRS dataset, these are conventional militant attacks in which there is a larger chance of exchanging re: ambushes, direct re, artillery, pitched nettle and troop captures (Bueno de Mesquita et al. 2015: 544). For the GTD dataset, labor-intensive attacks are total attacks excluding bombings (which are thought to be less labor intensive). In both cases the coecients are of similar (or larger) magnitude as the main results using all attacks, and they are negative and signicant at the 1% level. We also consider several placebo tests, organized by attack type and perpetrator. Columns 7 and 8 of Appendix Table A.9 consider attacks that are less labor inten- sivebombings and asymmetric attacks. The results are more mixed than for Iraq and Afghanistian: it appears that these type of attacks are sensitive to the onset of harvest in the BFRS dataset (column 8) but not in the GTD dataset (column 7). Turning to attacks organized by perpetrator, column 4 of Appendix Table A.9 shows that the onset of harvest has no impact on the intensity of attacks from foreign parties (mainly the US, India, Afghanistan, or multilateral), which is not surprising, as these parties are not reliant on part-time ghters. For state-initiated attackinitiated by the military, paramilitary, or police against civilians or militantsresults are mixed. Column 5 of Appendix Table A.9 shows state-initiated attacks increase with harvest onset (and harvest intensity), providing some tentative evidence that the state may be trying to protect the harvest from 42For example, if harvest usually took place in the rst few days of a month, then climate shocks could move it into the previous month in some years. This would explain the negative coecient in the month before harvest in Appendix Table A.8, column 3. 34 militants or are perhaps taking strategic advantage of their reduced ghting capacity. However these results are based on a sample of district-years where the state attacks, which is dierent from the district-year sample of militant attacks used in the main results. If the state were responding to the threat of militants, the eect of harvest on state attacks should be even stronger in the sample with militant attacks. However, when we restrict the sample to district-years based on militant attacks, the eect disappears (column 6). As such, we do not put too much weight on these results. 5. Conclusions and Policy Implications This paper studies variation in the opportunity costs of ghting as a key mechanism to explain the intensity of armed conict across dierent settings. Our theoretical framework suggests that the wide range of estimates in the literature might be caused by variation in shock persistence, with estimates of the opportunity cost driven toward zero for the more persistent shocks. Instead, we propose using seasonal income shocks due to harvestwhich are temporary and anticipatedas a more accurate way to estimate the eect of opportunity cost on conict when the dynamic gains from ghting are hard to observe. Applying our methodology in Iraq, Pakistan, and Afghanistan, we nd that at average cultivation intensity, harvest reduces the average share of monthly attacks by around 6% to 22%, depending on the country and dataset used. Results are robust to a wide array controls, and dynamic specications generally suggest an actual fall in conict at harvest time rather than a shift in conict to adjacent non-harvest months. We also nd some evidence that the reduction in total attacks is driven by a fall in labor-intensive attacks. Our results have three sets of policy implications. First, our empirical evidence in favor of the opportunity cost mechanism suggests that, in principle, employment programs or development aid can reduce the intensity of conict by increasing the opportunity cost of ghting. However, our second implication is that how these policies interact with the dy- namic incentives to ght is crucial, and more permanent policies can have no eect on conict intensity or can even it. For example, permanent transfers (in cash or food) may encourage ghting over the rents from these schemes or mean that households are wealthy enough to devote time to ght for causes they care about. For this reason, we are wary of policy prescriptions based on long-lasting employment programs or on permanent forms of development aid, as their persistence across periods may lead 35 to unintended consequences. This may explain the counterintuitive results of several recent studies (Nunn and Qian 2014; Crost et al.2014; Weintraub 2016). Our nal implication is that it might be possible to design more sophisticated policies that increase the opportunity cost of ghting without changing its dynamic benets. Closest to our empirical results, a temporary employment program at a critical junction in a conict, or in key ghting months, may increase the opportunity cost of ghting without aecting its dynamic benets or subjective valuations of 43 grievances. A thorough investigation of these implications would be an interesting area for future research. References [1] Afsar, S., Samples, C., and Wood, T. (2008). The Taliban: An organizational analysis. Naval Postgraduate School, Monterey, CA. Department of National Security Aairs, Chicago IL. [2] Bazzi, S., and Blattman, C. (2014). Economic shocks and conict: Evidence from commodity prices. American Economic Journal: Macroeconomics, 6(4), 138. [3] Beath, Andrew, Christia, Fotini and Enikolopov, Ruben, (2017). Can Development Programs Counter Insurgencies?: Evidence from a Field Experiment in Afghanistan. MIT Political Science Department Research Paper No. 2011-14, Available at SSRN: https://ssrn.com/abstract=1809677 [4] Becker, G. (1968). Crime and punishment: An economic approach. Journal of Political Econ- omy, 76(2), 169217. [5] Berman, E., Shapiro, J. N., and Felter, J. H. (2011a). Can hearts and minds be bought? The economics of counterinsurgency in Iraq. Journal of Political Economy, 119, 766819. [6] Berman, E., Callen, M., Felter, J. H., and Shapiro, J. N. (2011b). Do working men rebel? Insurgency and unemployment in Afghanistan, Iraq, and the Philippines. Journal of Conict Resolution, 55(4). [7] Blair, G., Christine Fair, C., Malhotra, N., and Shapiro, J. N. (2013). Poverty and support for militant politics: Evidence from Pakistan. American Journal of Political Science, 57(1), 3048. [8] Brückner, M., and Ciccone, A. (2010). "International commodity prices, growth and the out- break of civil war in Sub-Saharan Africa." The Economic Journal, 120(544), 519534. [9] Bueno de Mesquita, E. (2013). Rebel tactics. Journal of Political Economy, 121(2), 323357. 43A wage subsidy funded by local taxation (or a reduction in food/fuel subsidies) would boost the incentive to work without increasing the general level of income (aecting the value of the cause) or creating additional local rents to ght over. Making employment programs conditional on a successful counterinsurgency would mean that the extra income generated is disconnected from the value of winning. 36 [10] Bueno de Mesquita, E., Fair, C. C., Jordan, J., Rais, R. B., and Shapiro, J. N. (2015). Mea- suring political violence in Pakistan: Insights from the BFRS Dataset. Conict Management and Peace Science, 32(5), 536558. [11] Burke, M. B., Miguel, E., Satyanath, S., Dykema, J. A., and Lobell, D. B. (2009). Warming increases the risk of civil war in Africa. Proceedings of the National Academy of Sciences, 106(49), 2067020674. [12] Cashin P, Liang, H., and McDermott, C. J. (2000). How persistent are shocks to world com- modity prices? IMF Sta Papers, 47(2), 177217. [13] Chassang, S., and Padró i Miquel, G. (2009). Economic shocks and civil war. Quarterly Journal of Political Science, 4(3), 211228. [14] Condra, L. N., Long, J. D., Shaver, A. C., & Wright, A. L. (2018). The logic of insurgent electoral violence. American Economic Review, 108(11), 3199-3231. [15] Crost, B., and Felter, J. H. (2016). Export crops and civil conict. Journal of the European Economic Association, 18(3), 14841520 [16] Crost, B., Felter, J., and Johnston, P. (2014). Aid under re: Development projects and civil conict. The American Economic Review, 104(6), 18331856. [17] Dal Bo, E., and Dal Bo, P. (2011). Workers, warriors, and criminals: Social conict in general equilibrium. Journal of the European Economic Association, 9(4), 646677. [18] Dube, O., and Vargas, J. F. (2013). "Commodity price shocks and civil conict: Evidence from Colombia." The Review of Economic Studies, 80(4), 13841421. [19] Fearon, J. (2008). Economic development, insurgency, and civil war. In Institutions and Eco- nomic Performance, ed. E. Helpman, 292328. Harvard University Press: Cambridge MA. [20] Fearon, J., and Laitin, D. (2003). Ethnicity, insurgency, and civil war. American Political Science Review, 97(1), 7590. [21] Figes, O. (1996). A People's Tragedy: The Russian Revolution 18911924. Pimlico: London. [22] Garnkel, M. R. (1990). Arming As a strategic investment in a cooperative equilibrium. American Economic Review, 80(1), 5068. [23] Garnkel, M. R., and Skaperdas, S. (2007) Economics of conict: An overview. in Handbook of Defense Economics, Volume 2, Defense in a Globalized World, ed. T. Sandler, and K. Hartley, 649710. Elsevier, North-Holland: Amsterdam and Oxford. [24] Giure, K. A. (1997). First in ight: Desertion as politics in the North Carolina Confederate Army. Social Science History, 245263. [25] Grossman, H. I. (1991). A general equilibrium model of insurrections. American Economic Review, 81(4), 912921. [26] Guardado, J. (2018). Land tenure, price shocks, and insurgency: Evidence from Peru and Colombia. World Development, 111, 256269. [27] Haavelmo, T. (1954). A Study in the Theory of Economic Evolution. North-Holland: Amster- dam. [28] Hirshleifer, J. (1988). The analytics of continuing conict. Synthese, 76(2), 201233. [29] Hirshleifer, J. (1989). Conict and rent-seeking success functions: Ratio vs. dierence models of relative success. Public Choice, 63(2), 101112. 37 [30] Hodler, R., and Raschky, P. A. (2014). "Economic shocks and civil conict at the regional level." Economics Letters, 124, (3), 530533. [31] Horowitz, D. L. (1985). Ethnic Groups in Conict. University of California Press: Berkeley. [32] Hsiang, S. M., Burke, M., and Miguel, E. (2013). Quantifying the inuence of climate on human conict. Science, 341(6151), 1235367. [33] Iyengar, R., Monten, J., and Hanson, M. (2011). Building peace: The impact of aid on the labor market for insurgents. NBER Working Paper 17297. [34] IIASA/FAO. (2012). Global Agro-Ecological Zones (GAEZ v3.0). [35] McClintock, C. (1998). Revolutionary Movements in Latin America: El Salvador's FMLN and Peru's Shining Path. United States Institute of Peace: Washington, DC. [36] Miguel, E., Satyanath, S., and Sergenti, E. (2004). Economic shocks and civil conict: An instrumental variables approach. Journal of Political Economy, 112, 725753. [37] Nunn, N., and Qian, N. (2014). US food aid and civil conict. The American Economic Review, 104(6), 16301666. [38] Peters, G. (2009). How opium prots the Taliban. Peaceworks No. 62. Washington, DC: United States Institute of Peace. [39] Piazza, J. A. (2012). The opium trade and patterns of terrorism in the provinces of Afghanistan: An empirical analysis. Terrorism and Political Violence, 24(2), 213-234. [40] Qazi, S. H. (2011). The neo-Taliban, counterinsurgency and the American endgame in Afghanistan. Institute for Social Policy and Understanding. [41] Shapiro J. and S. Gulzar (2012). Political Violence in Pakistan: Myths vs. Realities, IGC Policy Brief, https://www.theigc.org/wp-content/uploads/2014/09/Rais-Et-Al-2012-Policy-Brief.pdf [42] Skaperdas, S. (1992). Cooperation, conict, and power in the absence of property rights. American Economic Review, 82(4), 720739. [43] United Nations Oce on Drugs and Crime. (2010). World Drug Report 2010. United Nations Publications: Herndon, VA. [44] UNODC, Counter-Narcotics Directorate. (2004; 2010) "Government of Afghanistan: Afghanistan: Opium Survey 2004/2010." [45] Weintraub, M. (2016). Do all good things go together? Development assistance and insurgent violence in civil war. The Journal of Politics, 78(4), 9891002. [46] Wigle, J. (2010). Introducing the worldwide incidents tracking system (WITS). Perspectives on Terrorism, 4(1). [47] Willmott, C. J., and Matsuura, K. (2015). Terrestrial air tempera- ture and precipitation: monthly and annual time series (19502014). http://climate.geog.udel.edu/~climate/html_pages/Global2014/README.GlobalTsT2014.html [48] World Bank. (2011). World Development Report 2011: Conict, Security, and Development. World Bank: Washington, DC. [49] Wright, A. (2019). Economic shocks and rebel tactics. Unpublished Manuscript. Accessed on Feb 2, 2020: Link. 38