Optimal Coercion

Sandeep Baliga Tomas Sjöström Northwestern University Rutgers University

Abstract A protagonist (hegemon, rebel group) faces an antagonist (hostile country, colonial power) and a third party (representative citizen of the antagonist’s country). The protagonist wants the antagonist to cooperate (give up WMDs, share power). The protagonist can use targeted interventions that directly harm the antagonist, and comprehensive interventions that provoke social unrest by harming the representative citizen. The antagonist’sinterests are imperfectly aligned with his citizen’s. Targeted interventions reduce the misalignment, so the stronger are the targeted measures, the less likely is social unrest — the “rally around the flag”effect. To provoke social unrest, the protagonist may reduce the targeted measures, but this policy is only credible if the protagonist does not know the antagonist’s true motives. An exogenous constraint on comprehensive interventions may make it optimal to reduce the targeted measures even further, making the citizen worse off, in contrast to the common belief that comprehensive and targeted measures are substitutes. But if social unrest is suffi ciently unlikely to trigger cooperation (e.g., in a totalitarian state), then it is optimal to abolish the comprehensive measures and maximize the targeted measures, making the citizen better off.

1 Introduction

A protagonist uses coercion to change the behavior of an antagonist. In international relations, the protagonist may be a hegemonic power who uses economic sanctions to coerce a minor power (the antagonist) into disarming. In a civil conﬂict, the protagonist may be a rebel group who uses terrorism to coerce the government (the antagonist) into power-sharing. Our aim is

1 to study when and how coercion might be effective. In particular, we will contrast comprehensive measures that impose costs on the population as a whole (comprehensive economic sanctions, the bombing of marketplaces) with measures that target a political elite (“smart”sanctions, the assassina- tion of political leaders). Will such interventions unite or divide the population and the elite? What is the optimal mix of targeted and comprehensive interventions? Do these different types of interventions work with or against each other? The United States has frequently imposed comprehensive economic sanctions (e.g., trade embargoes) on countries that are perceived as threats against international peace and stability. More recently, there has been a push towards targeted sanctions (blocking the elite’s financial assets and transactions, restricting their ability to travel etc.). For example, Iran has long been subject to comprehensive sanctions which restrict its fuel imports and access to U.S. technology and investment. The Obama administration added targeted sanctions, including sanctions on financial transactions with the Iranian Revolutionary Guard, and blocked U.S. backed assets of Iranian government- controlled entities.1 The Trump administration went yet further and singled out specific individuals in the Iranian government for targeted sanctions. North Korea has faced comprehensive U.S. sanctions that restrict trade and damage the North Korean economy; targeted sanctions on foreign financial transactions were introduced after nuclear tests.2 By contrast, South Africa, Iraq and Libya faced mainly comprehensive sanctions.3 In a different context, a rebel group faces a similar strategic choice of targeted versus comprehensive measures. It may target a political elite in order to persuade the regime to share power or relinquish control of territory. Alternatively, it may attack the population at large in order to destabilize the regime or subject it to popular pressure to satisfy its demands. For example, the I.R.A. staged “spectaculars”or “massive attacks with maximum impact” such as the bombing of a military procession and a military band in Hyde Park and Regent’sPark in 1982 (Foxwell, 2013). These attacks harmed both

1 See Department of Treasury (2013) for details. 2 See Reynolds and Tan (2012) for an overview of sanctions on Iran, North Korea and others. 3 Similarly, after the collapse of the Soviet Union, Russia used economic coercion against the new independent states which relied on Russia to export their products, import energy and other inputs in production and to transport oil and natural gas to the West (see Drezner (1999)).

2 civilians and soldiers. The I.R.A. also bombed Prime Minister Margaret Thatcher and her Cabinet in Brighton in 1984, killing or injuring members of the ruling Conservative Party. The Taliban and other insurgents have attacked civilians, but also government oﬃ cials.4

In general, targeted measures are meant to incentivize the political elite to change course by directly impacting its cost-benefit analysis. Compre- hensive measures have a more indirect purpose: to cause social and popular unrest which either destabilizes the regime or forces it to change course to avoid being destabilized.5 However, coercive measures may backfire if, instead of creating popular unrest, the population unites with the elite and “rallies around the flag”(Ostrom and Job, 1986). For example, the bombing of Buckingham Palace by the German Luftwaffe during World War II only served to endear the monarchy to the population.6 Margaret Thatcher and the Conservative Party gained popularity after the Brighton attack (Lanoue and Headrick, 1998). The average citizen will often be poorly informed about the nature of external threats. The government will use espionage, expert advisors, diplo- macy and other forms of information gathering to become better informed. In view of this information asymmetry, the citizen’s response to coercive measures will depend on his beliefs about how well his country’sleaders use their superior information to represent his interests. If the average Iranian citizen blames Iran’sleaders for their dispute with the U.S., the result may be social unrest. But if the citizen believes the dispute is caused by the in- herent hostility of the U.S., then he will rally around the flag. In this paper we consider how these beliefs depend on the nature of sanctions, and how optimal sanctions take this into account.

Our model has three players: the protagonist, the antagonist, and a third party referred to as the (representative) citizen. The antagonist may decide to cooperate with the protagonist, who is either a hostile “Hawk”or a

4 See the U.N. report on civilian casualties (2018) and Sediqi (2019). 5 Strategic bombing of population centers in wartime is another “comprehensive”measure which can be justiﬁed in a similar way: it may cause uprisings which will persuade the government to surrender or else be destabilized (Douhet 1921). 6 Queen Elizabeth the Queen Mother said, “I am glad we have been bombed. Now we can look the East End in the eye”(Shew (1954, p. 76). King George VI said, “[W]e have found a new bond with them [the citizens] as Buckingham Palace has been bombed as well as their homes and nobody is immune”(Shawcross (2009)).

3 benevolent “Dove”. The incentives of the antagonist are partially, but not perfectly, aligned with those of the citizen. Both favor non-cooperation if the protagonist is a Hawk. However, the citizen would prefer cooperation with a Dove, while the antagonist may prefer non-cooperation even in this case. The reason for this misalignment is that the antagonist derives a private benefit from the “strength”projected by non-cooperation, while the citizen benefits from the “peace dividend” of cooperation. The private benefit the antagonist derives from non-cooperation, and hence his inclination to cooperate, depends on his (privately known) type. If the citizen does not trust the antagonist to make the right decision, he may stage an uprising which, if successful, will lead to cooperation. A successful uprising can be interpreted as either causing the antagonist to lose offi ce (“regime change”), or forcing the antagonist to cooperate in order to avoid losing offi ce. However, because the citizen does not know the protagonist’strue type, the uprising is a risky gamble which may lead to a bad outcome if the protagonist is a Hawk. In the leading interpretation of the model, the protagonist is the leader of a hegemonic country, the antagonist is the leader of a small country with a program for developing WMD, and the third party is the representative citizen of the small country. To cooperate means to terminate the program (“disarm”). Non-cooperation (i.e., not disarming) means the country misses out on a peace-dividend if the protagonist is a Dove, but it deters the protagonist from taking hostile actions (e.g., invading the small country) if he is a hostile Hawk.7 In addition, the leader of the small country derives a private benefit from the “strength”afforded by developing WMD. For example, the WMD might be used to deter local competitors for regional influence or to control a restive population, or may simply be a source of prestige. In an alternative interpretation of the model, there is only one country. The protagonist is the leader of a rebel group, the antagonist the leader of the government, and the third party is (again) the representative citizen. The cooperative decision is to share power with the rebel group (or concede territory). If the rebel group has truly circumscribed motives (a Dove), then power-sharing will lead to peaceful coexistence (the peace dividend). But if the rebel group has hostile intent (a Hawk), then power sharing may lead to a slippery slope where the rebel group eventually takes over completely and

7 For example, North Korea sees nuclear inspections “as only the ﬁrst wedge in opening the entire North Korean state to penetration...The North would be accepting a sure path to change in regime and society and ultimately, state survival”(Bracken, 1993, p. 150).

4 violates citizens’rights.

The protagonist can use coercive interventions to increase the cost of non-cooperation: targeted interventions hurt the antagonist, while comprehensive interventions hurt the citizen. In the leading interpretation, targeted measures are “smart” sanctions, comprehensive measures are trade embargoes. In the alternative interpretation, targeted measures are attacks on the government, comprehensive measures are attacks on the population at large. Strong targeted measures align the antagonist’s interests with those of the citizen. Specifically, the antagonist becomes more inclined to cooperate when his information suggests that the protagonist is a Dove. This makes the citizen more willing to delegate decision making to the antagonist, which we refer to as the “rally around the flag”effect. The protagonist’s preferences over coercive interventions depend on his type. Directly incentivizing the antagonist to cooperate voluntarily is more likely to work for a Dove than for a Hawk, because the antagonist’sinforma- tion makes him more likely to cooperate with the former. Therefore, a Hawk tends to favor provoking social unrest by comprehensive measures, while a Dove tends to favor incentivizing the antagonist by targeted measures. How- ever, if the Hawk reveals his true type then the antagonist and the citizen will be united in favor of non-cooperation. Therefore, a separating equilibrium, where the two types choose different coercive policies, cannot exist. Although the two types disagree over coercive policies, there is a pooling equilibrium which is best for each type (the expected payoff for each type is greater than in any other pooling equilibrium). Moreover, this pooling equilibrium satisfies standard equilibrium refinements. The coercive policy that is chosen (by both types) in this equilibrium is the optimal coercive policy. Of course, being best for each type, this policy maximizes the protagonist’s ex ante expected payoff. We characterize the optimal coercive policy and show how it depends on the parameters. In essence there are two possibilities. The optimal coercive policy is either a targeted policy, with maximum targeted interventions but no comprehensive interventions, or a comprehensive policy, with maximum comprehensive interventions and limited targeted sanctions. The comprehensive policy aims to provoke an uprising. This is the optimal coercive policy if the antagonist is unlikely to cooperate voluntarily, but an uprising has a good chance of success. The comprehensive policy employs only weak targeted measures, making sure that the antagonist’sand the citizen’s interests are suffi ciently misaligned. This prevents the “rally around

5 the flag”effect from taking hold. If, for exogenous reasons, the protagonist is forced to soften the comprehensive measures, then non-cooperation becomes less costly to the citizen, so the targeted measures must be even weaker to prevent the citizen from rallying around the flag. Thus, comprehensive and targeted interventions are complements. Constraints on comprehensive measures may come from moral concerns. For example, harsh trade embargoes that cause great suffering among civilians may become morally intolerable and politically infeasible. This is one reason why current policy discussions focus on the idea of using targeted sanctions instead of trade embargoes (Drezner, 2011). Since disarmament requires either voluntary cooperation by the antagonist (who is incentivized by targeted measures) or a successful uprising by the citizen (who is incentivized by comprehensive measures), the idea that targeted sanctions can substitute for comprehensive sanctions is quite intuitive. However, if the protagonist’s aim is to provoke an uprising, then comprehensive and targeted interventions are complements. The optimal comprehensive policy compen- sates for an exogenous restriction on comprehensive measures by softening the targeted measures, which is not consistent with the current policy discus- sion. Moreover, because the softening will make the antagonist less inclined to cooperate with a benevolent protagonist, there is less likely to be a lifting of sanctions and a peace dividend. Thus, the constraint on comprehensive measures, even if motivated by concern for the population, may not make them better off. This illustrates that it may be valuable to study how the two kinds of interventions interact. Constraints on comprehensive measures may also be due to technologi- cal factors. During World War II, German bombing of British population centers at first seemed to reduce public support for the war effort. However, Hitler could not effectively block Britain from trading with other countries, and until the development of the V-2 rocket he did not have the military technology that could hurt the British population suffi ciently. The V-2 came too late to have a significant impact, and targeting the Buckingham Palace only made the British rally around the flag. If the optimal policy is a targeted policy, then comprehensive measures should be abolished and targeted measures maximized. The targeted policy aims to elicit voluntary cooperation from the antagonist without attempting to provoke an uprising. The representative citizen prefers the targeted policy, most obviously because he dislikes comprehensive sanctions, but also because targeted sanctions make a peace divided more likely (the antagonist will

6 more inclined to cooperate with a benevolent protagonist, even if there is no uprising). The targeted policy is optimal if an uprising is diffi cult to provoke and unlikely to succeed even if it occurs, and if the antagonist is not too unwilling to cooperate voluntarily. It is also likely to be optimal if the “rally around the flag” effect is very strong, since in this case the targeted measures would have to be severely limited in a comprehensive policy. An uprising is diffi cult to provoke if the protagonist is ex ante likely to be hostile, or if the peace-dividend is small. An uprising is unlikely to succeed in highly authoritarian countries. Thus, the comprehensive policy is more likely to be optimal against democratic countries, to the detriment of the population (who prefer a targeted policy). If the protagonist uses an optimal coercive policy, then the representative citizen may have an unexpected preference for living in a country where it is hard to stage a successful uprising.

The optimal comprehensive policy prescribes removing targeted measures in order to provoke social unrest. This implies that some types of antagonist will not cooperate voluntarily, although they could have been induced to cooperate by targeted measures. The feasibility of this policy relies on our assumption that the protagonist is uninformed about the antagonist’stype. If the protagonist is informed, and the citizen knows that the protagonist is informed (but does not himself know the antagonist’stype), then the optimal comprehensive policy is not implementable. Indeed, if the protagonist knows that the antagonist can be incentivized to cooperate, then he would surely do this. This would align the antagonist’s interests more closely with the citizen’s, making an uprising less likely. If the protagonist tries to provoke an uprising, then the citizen concludes that the protagonist is likely a Hawk (as the antagonist cannot be incentivized to cooperate), so he will rally around the ﬂag. We show that if the protagonist knows the antagonist’stype, then the best equilibrium for the protagonist is separating: the Hawk and the Dove choose diﬀerent policies. If the protagonist is a Dove and the antagonist is a type that can be incentivized to cooperate, then the protagonist implements a targeted policy. Otherwise, the protagonist tries to provoke an uprising, insofar as this is possible. This separating equilibrium is worse for the protagonist than being uninformed and implementing the optimal comprehensive policy.

7 2 Related Literature

Smith [36] and [37] argued that a leader who performs well during an international conflict signals his competence to a domestic audience, so a competent leader has an incentive to trigger a conflict in order to gain popularity.8 Our logic and its implications are quite different from his. In our model, a small country can increase the probability of conflict by disarming. Smith’stheory would suggest that a competent leader of the small country might disarm to signal his type, and his increased popularity might be interpreted as the population rallying around the flag. In our theory, the population may rally around the flag when the leader refuses to disarm, not because they have up- dated their beliefs about his competence, but because his refusal to disarm signals the hostility of the foreign power. Another theory of the interaction between international and domestic politics was proposed by Di Lonardo, Sun and Tyson [11]. If it is commonly known that the interests of the domestic opposition are misaligned with those of a foreign hegemon, then the small country’sleader can economize on domestic security and yet be more likely to stay in power, because the opposition knows that an uprising may trigger a foreign intervention. In our model, the representative citizen of the small country does not know what motivates the hegemon. If the hegemon is benevolent, then the citizen wants to disarm, so his interests are more aligned with the interests of the hegemon than with those of his own leader (who derives a private benefit from the WMD). But if the hegemon is hostile, then the citizen wants to stay armed, so his interests are more aligned with their own leader’sinterests. Our focus is on how the hegemonic power can exploit the different interests in the small country to achieve either regime change or voluntary disarmament. Thus, the models formalize rather different scenarios. While Di Lonardo, Sun and Tyson [11] may describe the current conflict between President Assad of Syria and ISIS (where ISIS is the domestic opposition whose interests are surely misaligned with those of the U.S.), our model may better capture some aspects of the current situation in Iran (where the Iranian citizens don’tknow whether their interests are better aligned with the interests of the U.S. or with those of their own leaders). Whang, McLean and Kuberski (2013) considered a model of sanction

8 For example, Margaret Thatcher’ssuccessful prosecution of the Falklands War against Argentina helped her win popular support.

8 threats. Using our terminology, the hegemon would move ﬁrst by threaten- ing sanctions. If the small country resists (i.e., does not disarm), then the hegemon may proceed to actually impose sanctions. Thus, in their model, actions are strategic complements in the sense that the hegemon may take the hostile action (impose sanctions) only if the small country takes a hostile action (not disarming). In our model, actions are strategic substitutes in the sense that the hegemon may take a hostile action (attack) only if the small country takes the non-hostile action (disarming). The two models may therefore capture quite diﬀerent real-world scenarios (e.g., disputes over territory versus disputes over WMD). More importantly, Whang, McLean and Kuberski (2013) treat the small country as a unitary actor, while we focus on the interaction between the political elite and the population, and how to optimally exploit (and boost) their misalignment of interests. Coe and Kang (2020) study comprehensive sanctions on a representative citizen to destabilize the status quo peace with a targeted government. Such sanctions are deliberately designed to be rejected by the targeted government and hence to induce potential regime change. This attempted regime change may end up with the existing government still in charge or with the opposition in power. Coe and Kang (2020) provide an interesting bargaining perspective on sanctions but do not study the informational issues that are the main focus of our paper. Eaton and Engers (1999) and Spaniel and Smith (2015) allow asymmetric information between a protagonist and an antagonist with private information of their resolve. The sanctions may fail when the antagonist resolve. Our informational setting is where it is the protagonist’smotives that are not known to a representative citizen - a quite natural scenario that deserves attention. Empirical research shows that coercive interventions have a mixed record. Years of debilitating comprehensive sanctions did eventually lead South Africa to end apartheid. Iran did agree to limit its nuclear program, and open it up to monitoring, after comprehensive and targeted sanctions. But severe sanctions against Iraq did not lead to the desired outcome. In a dataset consisting of 115 cases between 1914 and 1990, Hufbauer et.al. (1990) found that sanctions were successful in one third of the cases. But Pape (1997) re-examined these cases and argued that the success rate was much lower, mainly because many of the “successes” could be attributed to the use or threat of force.9 Drezner (2003) argued, however, that sanctions may be

9 For example, the United Nations imposed sanctions twelve times during the 1990’s,

9 more successful than it appears in these datasets. In many cases, the threat of sanctions causes a change in behavior: “If the target prefers conceding to incurring the cost of sanctions, it has an incentive to acquiesce before the imposition of sanctions” (Drezner, 2003, p.644). These successful cases do not appear as such in the datasets because the sanctions are not actually imposed. The empirical literature has also investigated the relative effi cacy of different kinds of sanctions. Several studies concluded that if the target nation is a democracy, then comprehensive sanctions are more likely to succeed than targeted sanctions (Allen, 2005, 2008, Bolks and Al-Sowayel, 2000, Brooks, 2002, Lektzian and Souva, 2007). After reviewing the United Nations sanctions of the 1990’s, Cortright and Lopez (2002, p. 8) concluded that “the obvious conclusion is that comprehensive sanctions are more effective than targeted or selective measures. Where economic and social impact have been greatest, political effects have also been most significant”. Empirical evidence supports the theory that comprehensive sanctions will lead to social unrest (Allen, 2008) and even regime change (Marinov, 2005). Some studies have found that targeted financial sanctions can be more effective than trade sanctions (Shagabutdinova and Berejikian, 2007). But overall, the empirical literature does not seem to justify the recent push towards targeted sanctions. Drezner (2011) concludes, based on a review of the empirical literature, that policymakers and scholars need to look beyond the targeted sanctions framework, with specific attention to the domestic political constraints of non-democratic countries. One interpretation of our model is that the protagonist is a rebel group who can attack government buildings or market places (or both). The group’s chosen tactics drive the interaction between the targeted government and the representative citizen, who have asymmetric information about the group’s motives. As far as we know, this problem was not previously studied in the literature on rebel group tactics — this literature has asked questions that are interesting but quite different from our main focus. Kydd and Walter (2006) provide an excellent taxonomy of the objectives and strategies or terrorist groups. For example, different groups may compete for support from most conspicuously on Iraq after the invasion of Kuwait and on the former Yugoslavia during the war in Bosnia. These two countries did eventually change their behavior: Iraq dismantled its WMD, and the Serbs signed the Dayton agreement to end the fighting in Bosnia. But as military interventions by international forces had also occurred, it is not clear whether these outcomes were actually due to the sanctions or to the use of force.

10 a domestic audience which does not know which group would best represent it against a foreign power. Bueno de Mesquita (2008) studies factors that might cause a terrorist organization to split into extreme and moderate factions. Bueno de Mesquita (2013) allows a rebel organization to choose between conventional tactics (warfare), unconventional tactics (guerilla attacks), or to withdraw from conflict entirely. He studies how the optimal method to mobilize the population varies with the outside option of not joining a rebel group. He finds an interesting non-monotonicity: when outside options are very good, no-one will mobilize so the rebel group withdraws from conflict, but when options are very bad, mobilization is high so the rebel group uses conventional tactics. Only when options are intermediate does the rebel group use unconventional tactics. Empirical work has focused on the impact of outside options on conflict, rather than on the issues we study. Miguel, Satyanath and Sergenti (2004) find that civil conflict becomes more likely when outside options deteriorate — the opportunity cost effect. There is a possible countervailing effect: if the outside option can be expropriated by a rebel group then conflict may increase as there is more to fight over (see Fearon, 2005). Dube and Vargas (2013) find evidence for this “rapacity”effect as well as the opportunity cost effect. The less labor intensive production is, the more likely is the rapacity effect.

3 The Game

3.1 Players and Actions There are three players. Player A is the protagonist, player B the antagonist, and player C the representative citizen. Player A wants player B to make a cooperative decision, or more brieﬂy, to cooperate with player A. We think of non-cooperation as the status quo; switching to cooperation requires a deliberate action. For now, we abstract from player A’s choice of coercive measures —this choice will be considered in Section 6. In the leading interpretation, player A is the leader of hegemonic country A, player B is the leader of small country B and player C is the representative citizen of country B. In the status quo, country B operates a program for developing WMD. The cooperative decision is to terminate the program, i.e., to “disarm”. An alternative interpretation is that there is only one

11 country, player A is the leader of a rebel group, player B is the leader of the government, and the cooperative decision is power-sharing. Player C may attempt to change the status quo by an uprising. If the uprising is successful then player B is removed from oﬃ ce (“regime change”). It will become evident that the only reason for player C to start an uprising is that he wants cooperation with player A. Therefore, we assume without loss of generality that the cooperative decision will be taken after a successful uprising.10 At the end of the game, player A can take a hostile action. In the leading interpretation, the hegemon may attempt to subjugate the small country in order to control its natural resources or to fundamentally change its political system. In the alternative interpretation, the rebel group might renege on a power sharing agreement and bring into play extreme policies that violate citizens’rights.

3.2 Time-line The three players play the following three-stage game. Stage 1: Player C chooses whether or not to attempt an uprising. If an uprising is attempted, it is successful with probability α (0, 1). Stage 2: If there was a successful uprising in stage 1 then∈ the cooperative action is implemented. Otherwise, player B chooses whether or not to take the cooperative action. Stage 3: Player A may decide to take the hostile action. All actions in the three stages are publicly observed.11 It is convenient to think of players B and C as belonging to country B. If the cooperative action was implemented in stage 2, either because there was a successful uprising or player B decided to cooperate, then we say that country B cooperates (or is

10 Alternatively, “successful uprising”may signify social and political unrest which causes player B to cooperate, either to avoid losing offi ce or to reduce intolerable costs of domestic unrest. With this interpretation, player B may not actually lose offi ce, but the successful uprising triggers cooperation, so the conclusions are the same. For simplicity, we refer to a successful uprising as a “regime change”. 11 The assumption that the decision to cooperate is publicly observed may seem strong but it is not, at least in the applications we are interested in. There is no point in disarming unless this is verified so that sanctions can be lifted. Hence, we may assume a country which disarms will allow weapons inspections to verify it.

12 cooperative). Otherwise, country B is said to have chosen non-cooperation (or is non-cooperative).

4 Types, Payoﬀs and Beliefs

4.1 The Protagonist Player A has two possible types, Hawk (H) and Dove (D). His type is denoted θA H,D . A key aspect of our model is that player B knows something about∈ { player} A that player C doesn’tknow. For simplicity, we assume both player A and player B, but not player C, know the true θA. Player C thinks 12 θA = H with probability h, where 0 < h < 1. Both types of player A want country B to cooperate, but for diﬀerent reasons. Roughly speaking, the Dove wants peace and the Hawk wants war. To be more precise, if country B is non-cooperative, then both types of player A get payoﬀ -1 from taking the hostile action, and 0 otherwise. If country B is cooperative, then a hawkish player A (with θA = H) gets +1 if he takes the hostile action and 0 otherwise; a dovish player A (with θA = D) gets 0 if he takes the hostile action and +1 otherwise.13 Thus, dovish player A will never choose the hostile action. Hawkish player A will choose the hostile action if and only if country B cooperates.14 In the leading interpretation of the model, a dovish hegemon wants country B to disarm (i.e., to cooperate) so that the two countries can coexist peacefully — country B’s weapons are considered a threat to stability and world peace. In contrast, a hawkish hegemon wants country B to disarm

12 More generally, we could assume players B and C observe noisy signals that are cor- related with player A’s type. This would complicate the exposition without adding new insights. 13 For simplicity, we assume regime change per se does not have any value to player A. It does have indirect value, as an instrument for ensuring that country B cooperates. In the Conclusion, we discuss the implications of dropping this assumption. 14 Actions are strategic substitutes for player A in the sense that aggressive behavior by player B (e.g., develop WMD) makes player A less likely to aggress (e.g., invade). However, one type (the Dove) has a dominant strategy not to be aggressive. Baliga and Sjöström (2020) study bargaining with commitment and ﬁnd that when the cost of conﬂict is very high, actions are strategic substitutes. Types with a high cost of commitment have a dominant strategy not be hostile, while types with a low cost of commitment have a best-response to be hostile if and only if the other player is not hostile.

13 so that he can invade or otherwise encroach on country B’s sovereignty — country B’sweapons would deter the Hawk from doing this. In the alternative interpretation, a dovish rebel group has limited objectives and is willing to live in peace if these objectives are met. A hawkish rebel group would impose its ideology on the population. In reality, the rebel group’s “type” may be determined by an internal power struggle between a dovish wing and a hawkish wing. For example, the I.R.A. had a moderate political arm Sinn Fein as well as an violent extremist arm. The latter found the former to be lacking in its embrace of Catholic doctrine.15 The Palestinian resistance to Israel has some who are willing to coexist in peace and others, like Hamas, who are committed to continuing resistance.16 Uncertainty about the rebel group’stype is the main concern in negotiations with the Taliban (International Crisis Group, 2020): Do they only seek U.S. withdrawal (Dove), or do they seek imposition of extreme Islamic law with violations of women’s and minority rights (Hawk)? It is plausible that the government (player B) has spies and informants and other sources of information about the rebel group’sinternal power struggles. This information is not necessarily available to the average citizen (player C).

4.2 The Antagonist If there is a successful uprising then player B takes no action and his payoff is normalized to zero. If there is not a successful uprising then player B gets “rents from offi ce”v > 0. In addition, if player B chooses not to cooperate then he gets a private benefit θB (“strength”) and incurs a cost t 0 (“targeted measures”against non-cooperators).17 Finally, if player A chooses≥ the hostile action then player B incurs a cost ΩB > 0. 18 Only player B knows the true value of θB. Players A and C think θB is

15 An I.R.A. member Gerry McGeough said: “You would never get a leader of Sinn Fein condemning abortion, homosexual "marriage" or anything of that nature. I, as an Irish nationalist and Catholic, never want to see the day when there are abortion clinics in every market town in Ireland. But looking around there is no political grouping willing to take a stance against that”(McDonald, 2003). 16 See Kydd and Walter (2002) for many relevant examples and also an analysis of how extremist might deliberately try to sabotage peace negotiations. Bueno de Mesquita (2008) studies factors that feed into increasing extremism. 17 For now, t is considered a ﬁxed parameter. The protagonist’s choice of t will be endogenized in Section 6. 18 Thus, there is an asymmetry: players A and B both know θA but only player B knows

14 distributed on [0, ¯θ] with cumulative distribution function F . We assume F is diﬀerentiable and strictly increasing on [0, ¯θ], with F (0) = 0 and F (¯θ) = 1. The two types, θA and θB, are uncorrelated. To rule out uninteresting corner solutions where cooperation is guaran- teed, we make the reasonable assumption that the cost of non-cooperation is less than both the maximum possible private beneﬁt and the cost of the hostile action.

¯ Assumption 1: t < θ and t < ΩB.

In the leading interpretation, a weapons program confers strength and helps build a position as a local superpower that can dominate weaker neigh- bors —or perhaps helps deter a revolution. This captured by the private ben- eﬁt θB. The cost t represents targeted sanctions, which are lifted if country B disarms but remain in place otherwise. In the alternative interpretation, t represents the cost of attacks against the government, while θB represents the beneﬁt of “looking tough” when dealing with rebels (which might give electoral or other advantages).

4.3 The Representative Citizen If country B does not cooperate, then player C incurs a cost c 0. If country B cooperates and player A chooses the hostile action, then player≥ C incurs a cost ΩC . If country B cooperates and player A does not choose the hostile action, then player C gets a beneﬁt λ > 0. In the leading interpretation, c is the cost of comprehensive sanctions, which are eliminated if country B cooperates.19 The beneﬁt λ is a “peace dividend”, the value of resources transferred from civilian use to to the weapons program, including investment in technology and manpower. To rule out the uninteresting possibility that player C always wants cooperation, we make the reasonable assumption that the cost ΩC of a hostile

θB. For example, the small country (country B) may be a secretive dictatorship, while the hegemon (country A) is an open democratic country with free press (which provides information about the preferences of the political elite). The citizens of the small country have limited resources and their search activities may be censored, so they ﬁnd it hard to obtain information about the hegemon (player C does not know θA). In Section 7 we consider the symmetric case, where players A and B both know both θA and θB. 19 For now, c is taken as given. Player A’schoice of c will be endogenized in Section 6.

15 action is quite large, both compared to the cost c of non-cooperation and the “peace dividend”λ:

Assumption 2: ΩC > c and hΩC > (1 h)λ − 5 Equilibrium

We solve the three stage game by backward induction and ﬁnd the unique Perfect Bayesian Equilibrium. At stage 3, player A will take the hostile action if and only if θA = H and country B cooperates. Now consider stage 2. Suppose there was no successful uprising, so player B makes a decision at stage 2. If θA = H then player B knows that he will get v ΩB if he − cooperates at stage 2 and v + θB t if he does not. (Recall that player B − knows θA.) Since θB 0 and ΩC > t by Assumption 1, player B will surely choose non-cooperation≥ in order to deter the Hawk from taking the hostile action. If θA = D then there is no threat of a hostile action from player A. Player B knows that he will get v if he cooperates, and v + θB t if he doesn’t. − Thus, player B will cooperate if and only if θB t < 0. The probability that − player B will cooperate when θA = D is

Pr θB < t = F (t). (1) { } Since player C thinks θA = D with probability 1 h, he thinks player B will cooperate at stage 2 with probability (1 h)F (t)−. If player B cooperates then player C’s payoff is λ, otherwise it is− c. (Note that player B never − cooperates when θA = H). Thus, if there is no successful uprising at stage 1 then player C’sexpected payoff is (1 h)F (t)λ + [1 (1 h)F (t)] ( c) = c + (1 h)F (t)(λ + c) (2) − − − − − − If there is a successful uprising then country B will cooperate, and player C gets ΩC if θA = H and λ if θA = D. Thus, player C’sexpected payoff is − hΩC + (1 h)λ (3) − − Hence, player C prefers to start an uprising in stage 1 if and only if 20

hΩC + (1 h)λ c + (1 h)F (t)(λ + c) (4) − − ≥ − − 20 We have simpliﬁed by assuming an uprising has no direct cost to player C. Such as cost could be easily added without changing anything important.

16 If c < hΩC (1 h)λ then (4) is violated for all t 0, so there can be no − − ≥ uprising in this case. If c hΩC (1 h)λ then (4) holds if and only if t τ(c), where τ(c) satisﬁes≥ − − ≤

c hΩC + (1 h)λ F (τ(c)) = − − . (5) (1 h)(λ + c) − ¯ Because c < ΩC , the right hand side of (5) is less than 1, so 0 τ(c) < θ. If t > τ(c), there is no incentive to start an uprising as player B’sincentives≤ are already well-aligned with player C’s. Player C wants cooperation with player A if and only if player A is a Dove. If t > τ(c) then the probability h that player A is a Hawk and the probability F (t) that player B cooperates with a Dove are enough to discourage an uprising, given comprehensive sanctions c. This is the manifestation of the rally around the ﬂag eﬀect. To summarize:

Proposition 1 The three-stage game has a unique PBE. If c hΩC (1 h)λ and t τ(c) then there is an uprising; country B cooperates≥ with− proba-− ≤ bility α + (1 α)F (t) if θA = D and with probability α if θA = H. Otherwise − there is no uprising; country B cooperates with probability F (t) if θA = D and with probability 0 if θA = H.

Note that h ΩC + λ 1 τ 0(c) > 0. ≡ 1 h (λ + c)2 F (τ(c)) − 0 If comprehensive sanctions are high, then player C has a big incentive to engage in an uprising even if targeted sanctions are high. But when comprehensive sanctions are low, player C has no such incentive as player B’s incentives to cooperate are already well-aligned with player C’s. Therefore, targeted sanctions have to be lower, so player B will not in eﬀect do player C’sbidding, for player C to have the incentive to start an uprising.

6 Optimal Coercive Policy

We now endogenize the choice of coercive interventions by adding an initial stage, stage 0, to the game. In stage 0, the protagonist makes a publicly observed choice of (t, c) [0, t¯] [0, c¯]. Targeted interventions impose a cost t ∈ ×

17 on player B. Comprehensive interventions impose a cost c on player C.21 The remaining stages 1,2 and 3 are as described above. Assume t¯ and c¯ satisfy Assumptions 1 and 2, so the analysis of Section 5 applies. In a separating (perfect Bayesian) equilibrium, player A reveals his true type by choosing (tH , cH ) when he is Hawk and (tD, cD) = (tH , cH ) when 6 he is a Dove. By Assumption 2, there is no uprising following (tH , cH ), because player C does not want cooperation with a Hawk. But player C would respond to (tD, cD) with an uprising because he strictly prefers cooperation with a Dove, and without an uprising there is always a positive probability that player B refuses to cooperate (see Section 4.2). But then the Hawk prefers to announce (tD, cD) at stage 0, because it makes it more likely that country B will cooperate. This shows that no separating equilibrium exists. From now on we focus on pooling (perfect Bayesian) equilibrium, where player A chooses the same interventions (t, c) regardless of type. Along the equilibrium path, player C’s stage 1 beliefs about player A must equal the prior beliefs, since no information revealed at stage 0. Off the equilibrium path, player C’s beliefs about player A may differ. But player C’s beliefs about player B must always equal the prior, because player A’schoice at stage 0 cannot signal player B’s type, even out-of-equilibrium. It turns out that among all pooling equilibria, there is one that maximizes player A’sinterim payoff regardless of type. In this pooling equilibrium, the coercive policy (t, c) X maximizes the Dove’sexpected payoff over all pooling equilibria. This∈ turns out to be the same (t, c) that maximizes the Dove’s expected payoff over all (t, c) [0, t¯] [0, c¯], when (t, c) is exogenously given as in Section 5. Also, this∈ pooling× equilibrium satisfies reasonable restrictions on out of equilibrium beliefs.

There are two cases that lead to trivial conclusions. First, if c¯ is very small, speciﬁcally c¯ < hΩC (1 h)λ, then Proposition 1 implies that an uprising is impossible, and the− best− policy for both types is (t,¯ 0). Second, 1 if c¯ is very big, speciﬁcally c¯ τ − (t¯) so that τ(¯c) t¯, then player A can get his bliss point: (t,¯ cˆ) such≥ that τ(ˆc) = t¯ will result≥ in an uprising with maximum targeted sanctions t¯. This is clearly the best policy for both of player A’stypes.

21 We assume the interventions do not have any direct cost for player A. In reality, of course, country A may also suﬀer from, for example, trade sanctions imposed on country B. As long as these costs are not too great, they will not change our analysis, so we set them equal to zero for simplicity.

18 From now on we focus on the non-trivial case where c¯ is in the intermediate range: 1 hΩC (1 h)λ c¯ < τ − (t¯) (6) − − ≤ In this case, it is possible to provoke an uprising, but there is a trade-off: for an uprising to occur, the targeted measures must be reduced below t¯ to counteract the rally around the flag effect. The only justification for imposing comprehensive measures is to provoke an uprising. By Proposition 1, an uprising occurs if c hΩC (1 h)λ and t τ(c). If t < τ(c) (which implies c > 0) then reducing≥ c−slightly− makes player≤ C better off without eliminating his incentive to start an uprising. Thus, if (c, t) is Pareto optimal and there is an uprising, then t = τ(c). Moreover, raising both c and t until c has reached the upper bound c¯ will make both types of player A better off. Thus, the best policy which provokes an uprising is (τ(¯c), c¯), where τ(¯c) < t¯ since we assume (6) holds. Since comprehensive measures are maximized, c =c ¯, we refer to (τ(¯c), c¯) as the comprehensive policy. It gives the Dove an expected payoff α + (1 α)F (τ(¯c)) (7) − since his payoff equals the probability of cooperation. Note that this is the sum of two terms: the probability that the uprising is successful (and so brings about cooperation), and the probability that there is no successful uprising times the probability that player B cooperates voluntarily. If (c, t) is such that no uprising occurs, then comprehensive measures do nothing to help player A, and Pareto optimality requires c = 0. To encourage player B to cooperate, targeted measures should be maximized. Thus, the best policy which does not provoke an uprising is (t,¯ 0). Since the comprehensive measures are eliminated, but the targeted measures are maximized, we refer to (t,¯ 0) as the targeted policy. Since there is no uprising, the Dove’sexpected payoff is F (t¯) (8) A Dove prefers the comprehensive policy (τ(¯c), c¯) to the targeted policy (t,¯ 0) if and only if (7) exceeds (8), i.e., α + (1 α)F (τ(¯c)) F (t¯) − ≥ Using (5), this is equivalent to

h (1 α)ΩC (1 h) (1 F (t¯)) λ c¯ − − − − (9) ≥ 1 hα (1 h)F (t¯) − − − 19 A Hawk can only get cooperation if there is an uprising, so he surely prefers the comprehensive policy. Thus, if (9) holds then both types prefer the comprehensive policy. Moreover, this policy can be supported by the (plausible) out-of-equilibrium beliefs that if player A deviates from the comprehensive policy at stage 0, then he is the Hawk (so there is no uprising). Thus, if (9) holds then the optimal coercive policy is the comprehensive policy. If (9) is violated, then there is disagreement: the Hawk prefers the comprehensive over the targeted policy while the Dove have the opposite preference. In this case, there is no pooling equilibrium which implements the comprehensive policy (τ(¯c), c¯). The reason is that the Dove can unilat- erally deviate to the targeted policy (t,¯ 0) and benefit. There is no out-of- equilibrium belief that player C could have about player A’stype that would make the deviation unprofitable. Hence, there is in fact no disagreement over pooling equilibria: (t,¯ 0) is the preferred pooling equilibrium policy for both types when (9) is violated. Again, this is supported by the (reasonable) out-of-equilibrium beliefs that if player A deviates at stage 0, then he is the Hawk. Thus, if (9) is violated then then the optimal coercive policy is the targeted policy. In general, the optimal coercive policy is the Dove’s most preferred policy if payoffs are derived as in Section 5 (with exogenously given policies). Proposition 2 Suppose (6) holds. The optimal coercive policy for (both types of) player A is the comprehensive policy (τ(¯c), c¯) if (9) holds, and the targeted policy (t,¯ 0) otherwise. An analysis of condition (9) reveals that the comprehensive policy (τ(¯c), c¯) is more likely to be optimal if ΩC is small, h is small, λ is big and α is big. A higher α signifies that the population is more able to influence political deci- sions. Countries that are more democratic in this sense would be more likely to suffer comprehensive interventions. The representative citizen is better off if the country becomes more autocratic, so that targeted interventions replace comprehensive measures in the optimal policy. Next consider changes in c¯. Consider a status quo scenario without significant restrictions on comprehensive interventions, so (9) holds with strict in- equality.22 The optimal policy is the comprehensive policy (τ(¯c), c¯), whereby

22 In the leading application, in the status quo there is no signiﬁcant public opinion within the hegemonic power that makes it hard to impose tough comprehensive sanctions. In the case of the rebel group, it feels unconstrained in its use of violence against the general population.

20 the protagonist hopes to trigger a successful uprising. Now suppose the comprehensive interventions are restricted for some exogenous reason: c¯ falls by an amount small enough that (9) still holds.23 In the optimal policy, targeted measures also fall, because τ 0(¯c) > 0. This is necessary to prevent the citizen from rallying around the flag, given that the comprehensive interventions are now smaller. That is, for an uprising to occur when comprehensive measures are restricted, it is necessary that the antagonist’sincentives are suffi ciently misaligned with the citizen’s. This is accomplished by relaxing the targeted measures. Thus, a small exogenous reduction in the comprehensive measures leads to less, not more, targeted interventions in the optimal policy. In this sense, the two measures are complements. Player B is made better off by the restriction on comprehensive measures, because the targeted measures are reduced. However, the citizen is not made better off. If (t, c) = (τ(¯c), c¯) then, using (2), (3) and (5), player C’sexpected payoff is

α (h( ΩC ) + (1 h)λ) + (1 α) ( c¯ + (1 h)F (τ(¯c)) (λ +c ¯)) × − − − × − − = h( ΩC ) + (1 h)λ − − which is independent of c.¯ Thus, restricting the comprehensive measures does not increase player C’sexpected payoff. Holding t constant, the fall in c¯ would indeed make player C better off. However, the protagonist’soptimal policy changes, so t also falls which makes player C worse off. The two effects balance. Now suppose c¯ falls by an amount large enough that (9) is no longer satisfied. Provoking an uprising is no longer optimal for the protagonist: targeted measures would have to be reduced by so much that the antagonist would be very unlikely to cooperate voluntarily. It is better to maximize the targeted measures and forget about the uprising. That is, the optimal policy is the targeted policy (t, c) = (t,¯ 0). Switching to a targeted policy makes player B worse off, but player C is strictly better off. In summary, there is a non-monotonicity: a small reduction in c¯ leads to a relaxation of the targeted measures, but a large fall in c¯ leads to a tightening of the targeted measures.

23 In the leading application, the public opinion may demand an end to tough comprehensive sanctions. In the case of the rebel group, violence against the general population may become less acceptable.

21 7 Symmetric Information

We have studied an asymmetric setting where player B know player A’stype, but player A does not know player B’stype. But perhaps both players have a good an idea of each other’smotives. Iran might monitor media and think tanks and hack the U.S. government to obtain such information. But equally the U.S. and Israeli governments may know the Iranian leadership’smotives via their own hacking and spy operations. In this section, we assume players A and B know each other’stypes (while player C still does not know either type). We show two results. First, for some parameter values the best equilibrium for player A is separating, in the sense that the Hawk and the Dove implement different coercive policies. Second, player A might actually prefer to be uninformed about player B’s type, as the information undercuts his ability to use comprehensive measures to provoke an uprising. If player A is a Dove and knows that player B’s type satisfies θB t¯, then he can make sure country B cooperates by imposing a targeted coercive≤ policy (t,¯ 0). Therefore, in any equilibrium, dovish player A must get a payoff +1 when θB t¯. On the other hand, if θB > t¯ then the only way to get country B to≤ cooperate is to provoke an uprising —if this is possible. The following equilibrium is therefore the best possible for player A. If θA = D and θB t¯, then player A chooses the policy (t,¯ 0). Otherwise, player A chooses (0≤, c˜), where c˜ is the smallest comprehensive measure that provokes an uprising. Feasibility requires c˜ c¯, which we check later. If there is no successful uprising, then player B chooses≤ the cooperative action if and only if θA = D and t θB. If player A chooses (t,¯ 0) then player C concludes that ≥ θA = D and θB t¯ and does not start an uprising (because he thinks player B will cooperate≤ anyway). If player A chooses an off-the-equilibrium-path policy, then player C believes that θA = H, and again he does not start an uprising. Finally, if player C observes that player A chose (0, c˜), then Bayes’ Rule implies that θA = H with probability h h∗ = . (10) h + (1 h) (1 F (t¯)) − − We need to make sure that player C is willing to start an uprising when he observes (0, c˜). If there is a successful uprising then his expected payoff is

h∗ΩC + (1 h∗)λ. − − 22 If there is no successful uprising following (0, c˜), then player C’s expected payoﬀ is c˜, because he expects that player B will choose non-cooperation − (since either θA = H or θA = D and θB > t¯). Thus, the smallest comprehensive intervention such that player C is willing to start an uprising satisﬁes

h∗ΩC + (1 h∗)λ = c.˜ (11) − − − Feasibility requires c˜ c¯. Using (5) and (10), we see that c˜ given by (11) satisfies the feasibility≤ condition if and only if τ(¯c) t¯. Thus, if τ(¯c) t¯, then we have found player≥ A’s optimal policy (i.e., the best equilibrium≥ for both Hawk and Dove). Player A imposes targeted measures if he is a Dove and player B can be incentivized to cooperate, and comprehensive measures otherwise. Player C does not learn player A’stype when player A imposes comprehensive measures. However, he does know that a dovish player A only uses comprehensive measures when it is impossible to get player B to cooperate voluntarily, so unless there is a successful uprising the peace dividend is not realized even though it might exist. If τ(¯c) t¯, then it is possible to implement suffi ciently strong comprehensive measures,≥ ensuring that player C prefers to start an uprising to potentially release the peace dividend. If τ(¯c) < t¯, then the fact that player A knows θB makes it impossible for player A to provoke an uprising. If player A did not know θB, then the Dove could prevent player C from rallying around the flag by setting c =c ¯ and t = τ(¯c) < t¯. Player B would not cooperate voluntarily when θA = D and θB (τ(¯c), t¯], making it more likely that a successful uprising would release a peace∈ dividend. This would give player C an incentive to start an uprising. But when player A does know θB, player A will always make sure that player B cooperates voluntarily when θB t¯. Since player C knows this, it is no longer possible to avoid the rally-around-the-flag≤ effect by reducing the targeted measures below t¯. Therefore, if player A knows player B’stype and τ(¯c) < t¯, uprisings can never occur. Moreover, player B will not cooperate voluntarily if either θA = H or θB > t¯, so player A might as well set c = t = 0 in this case. To summarize:

Proposition 3 If player A knows θB then the optimal coercive policy is as follows. When θA = D and θB t¯ set (t, c) = (t,¯ 0). In all other situations, set (t, c) = (0, c˜) if τ(¯c) t¯ and≤(t, c) = (0, 0) if τ(¯c) < t¯ (where c˜ is deﬁned by (11)). ≥

23 Knowing θB can make player A worse off. Suppose 0 τ(¯c) < t¯, and the comprehensive coercive policy is optimal when player A is≤ uninformed about θB (as in Section 6). This policy is no longer implementable when player A knows θB (and player C knows that player A knows θB). Hence, both types of player A are worse off knowing θB. Player A cannot commit to ignore information he has about player B’stype. Therefore, if player A is a Dove, and knows that player B can be incentivized to cooperate (θB t¯), then he will surely exploit this fact by imposing targeted measures. Intuitively,≤ this means that the antagonist’sinterest’swill be more closely aligned with those of the representative citizen. If player A tries to provoke an uprising by comprehensive measures, the citizen concludes that player A is likely to be be a Hawk, causing him to rally around the flag. To provoke an uprising, player A must make it publicly known that he is not gathering information about the antagonist —this creates a de facto commitment not to exploit any knowledge of θB. He should not use his scarce information- gathering resources in situations where a comprehensive coercive policy is optimal. Given the intuition from (9), he should not gather information about antagonists who are less likely to be hostile, where the peace dividend is likely to be big and where an uprising is more likely to be successful if it occurs.

8 Conclusion

We analyzed a model where a protagonist uses targeted interventions to encourage an antagonist to cooperate, and comprehensive interventions to encourage uprisings against the antagonist. The policy-oriented literature implicitly treats these two types of interventions as substitutes, but we found a sense in which they may be complements. Targeted interventions make the antagonist’s interests more closely aligned with those of the representative citizen. If the antagonist refuses to cooperate even in the face of severe targeted measures, then the citizen concludes that the protagonist is probably a Hawk, and in response to this the citizen is likely to rally around the flag. If comprehensive measures must be limited due to e.g. political constraints, then the protagonist faces a choice. He can try to provoke social unrest by reducing the targeted measures, thus avoiding the rally-around- the-flag effect (this is the sense in which the two measures are complements). Or he can switch to a targeted policy which abolishes all comprehensive mea-

24 sures (and hence does not cause social unrest). The optimal choice depends on parameters such as the likelihood that an uprising will be successful. Against more authoritarian countries, the targeted policy is more likely to be optimal. In the U.S., there is currently a push to replace comprehensive sanctions by targeted (“smart”) sanctions. Before choosing a policy, however, it is worthwhile to consider whether the policy is meant to provoke social unrest, cooperation by the political elite, or both. We made the stark assumption that the protagonist’sonly objective is to get cooperation. What if the Dove is actually a “Neoconservative ”who de- sires regime change for its own sake? (Regime change may signify not only a new leader but a new political system). A Neoconservative Dove would have a greater proclivity to encourage uprisings via comprehensive sanctions. In this respect he is similar to a Hawk. With a comprehensive policy, targeted and comprehensive interventions would still be complements, as this prop- erty does not depend on player A’s preferences. The optimal policy would be more likely to be a comprehensive policy with zero targeted interventions and maximum comprehensive interventions, aiming to encourage uprisings without triggering the rally-around-the-flag effect. At the other end of the spectrum, the Dove might be a “Realist” who fears the unintended conse- quences of regime change (preferring the “devil you know to the devil you don’tknow”). This would make the optimal policy more likely to be a targeted policy without comprehensive interventions. In our alternative interpretation, we assumed that a dovish rebel group would share power even if there is a regime change. But it is possible that power-sharing would become impossible in case of a regime change. A successful uprising will then always lead to the hostile action being taken by the rebel group. This is never beneficial for the citizens so comprehensive interventions never work. A rebel group which cannot credibly show that it will not run rampant after regime change must use targeted interventions to achieve its ends.

References

[1] Allen, S. (2005): “The Determinants of Economic Sanctions Success and failure,”International Interactions 31 (2): 117-138

25 [2] Allen, S. (2008): “The Domestic Political Costs of Economic Sanctions,” Journal of Conflict Resolution 52 (6): 916-944 [3] Baliga, Sandeep and Tomas Sjöström (2020): “The Strategy and Tech- nology of Conflict, Journal of Conflict Resolution, 128(8):3186-3219 [4] Bolks, S. and D. Al-Sowayel (2000) “How Long do Economic Sanctions Last?,”Political Research Quarterly 53 (June): 241-265 [5] Bracken, Paul (1993): “Nuclear Weapons and State Survival in North Korea,”Survival, vol. 35, no, 3, Autumn: 137-153. [6] Brooks, R. (2002): “Sanctions and Regime Type: What Works, and When?,”Security Studies 11 (4): 1-50 [7] Bueno de Mesquita, Ethan (2008): “Terrorist Factions”, Quarterly Jour- nal of Political Science, 3: 399—418. [8] Bueno de Mesquita, Ethan (2013): “Rebel Tactics”, Journal of Political Economy, 121(2):323-357. [9] Cho, In-Koo and David M. Kreps (1987): “Signaling Games and Stable Equilibria”, Quarterly Journal of Economics, 102(2): 179-221. [10] Cortright, D. and G. Lopez (2002): Sanctions and the Search for Secu- rity. Boulder, CO: Lynne Reiner [11] Di Lonardo, Livio, Jessica S. Shun and Scott A. Tyson, “Autocratic Stability in the Shadow of Foreign Threats,”forthcoming American Po- litical Science Review. [12] Drezner, D. (2003): The Sanctions Paradox, Cambridge University Press. [13] Drezner, D. (2003): “The Hidden Hand of Economic Coercion,”Inter- national Organization 57 (Summer): 643-659 [14] Drezner, D. (2011): “Sanctions Sometimes Smart: Targeted Sanctions in Theory and Practice,”International Studies Review 13: 96-108. [15] Oeindrila Dube and Juan F. Vargas (2013): “Commodity Price Shocks and Civil Conflict: Evidence from Colombia,”Review of Economic Stud- ies, Volume 80, Issue 4, October 2013, Pages 1384—1421.

26 [16] Eaton, Jonathan and Maxim Engers (1999): “Sanctions: Some Simple Analytics,”American Economic Review, 89(2), pp. 409—414

[17] Fearon James D. (2005): “Primary Commodities Exports and Civil War”, Journal of Conﬂict Resolution, 2005, vol. 49 (pg. 483-507).

[18] Foxwell, Jeremiah (2013): Case Study: The IRA’s Hyde and Regent’s Park Attacks in London, July 20, 1982,”Small Wars Journal, April 17.

[19] Hovi, Jon, Robert Huseby, and Detlef F. Sprinz (2005) "When Do (Im- posed) Economic Sanctions Work?" World Politics 57, no. 4: 479-99.

[20] Hufbauer, G., J. Schott and K. Elliott (1990): Economic Sanctions Re- considered, Washington: Institute for International Economics.

[21] David J. Lanoue and Barbara Headrick (1998): “Short-Term Politi- cal Events and British Government Popularity: Direct and Indirect Ef- fects,”Polity, Spring, 1998, Vol. 30, No. 3, pp. 417-433.

[22] International Crisis Group (2020): Taking Stock of the Taliban’s Per- spectives on Peace.

[23] Kydd, Andrew H., and Barbara F. Walter (2006): “The Strategies of Terrorism.” Quarterly Journal: International Security, vol. 31. no. 1. (Summer 2006): 49-80 .

[24] Lektzian, D. and M. Souva (2007): “An Institutional Theory of Sanc- tions Onset and Success,”Journal of Conﬂict Resolution 31 (6): 848—871

[25] Marinov, N. (2005): “Do Economic Sanctions Destabilize Country Lead- ers?”American Journal of Political Science 49 (3): 564-576.

[26] McDonald, Henry (2003): “IRA bomber attacks Sinn Fein on abortion”, Dec. 28, Guardian.

[27] Edward Miguel, Shanker Satyanath and Ernest Sergenti (2004): “Eco- nomic Shocks and Civil Conﬂict: An Instrumental Variables Approach,” Journal of Political Economy, Volume 112, Number 4 (August), 725-753.

[28] Miyagawa, M. (1992): Do Economic Sanctions Work? New York: St. Martin’s

27 [29] Ostrom, C. and B. Job (1986): “The President and the Political Use of Force,”American Political Science Review 80:541-63

[30] Pape, R. (1997): “Why Economic Sanctions Do Not Work,” Interna- tional Security 22 (1): 90-136.

[31] Reynolds, Cela Il. and Wilfred T. Tan (2012): Empirical Trends in Sanc- tions and Positive Inducements in Nonproliferation, in Sanctions, State- craft and Nuclear Proliferation, ed. Etel Solingen, Cambridge University Press.

[32] Abdul Qadir Sediqi (2019): “Death toll from attack on Afghan VP candidate’soﬃ ce rises to 20,”Reuters, July 28

[33] Shagabutdinova, E. and J. Berejikian (2007): “Deploying Sanctions while Protecting Human Rights: Are Humanitarian “Smart”Sanctions Eﬀective?”Journal of Human Rights 6: 59-74.

[34] Shawcross, William (2009): The Queen Mother, Vintage.

[35] Shew, Betty Spencer (1954): Queen Elizabeth The Queen Mother, Hod- der and Stoughton.

[36] Smith, Alastair (1996): “Diversionary Foreign Policy in Democratic Sys- tems,”International Studies Quarterly, March, Vol. 40, No. 1, pp. 133- 153

[37] Smith, Alastair (1998): “International Crises and Domestic Politics,” American Political Science Review, Sept., 1998, Vol. 92, No. 3, pp. 623- 638.

[38] Spaniel, William and Bradley C Smith (2015): “Sanctions, Uncertainty, and Leader Tenure,” International Studies Quarterly, 59(4), pp. 735— 749.

[39] Tsebelis, George (1990) “Are Sanctions Eﬀective? A Game-Theoretic Analysis,”The Journal of Conﬂict Resolution, Vol. 34, No. 1, pp. 3-28

[40] Department of Treasury (2013): Treasury Targets Assets of Iranian Leadership, June 4.

28 [41] United Nations (2018): Afghanistan: Protection of Civilians in Armed Conﬂict (Annual Report 2017), February 2018.

[42] Whang, Taehee, Elena V. McLean, and Douglas W. Kuberski (2013): “Coercion, Information, and the Success of Sanction Threats”American Journal of Political Science 57, no. 1: 65-81.