Ally provocateur

Abstract

The primary purpose of many alliances is to deter attacks on members of the alliance by potentially antagonistic states. Yet some alliances can increase the probability of conflict that may be initiated by alliance members. Cognizant of that possibility, states that wish to sustain peace may nevertheless intentionally form alliance commitments with revisionist leaders of other states. Faced with the partially conflicting goals of deterring antagonistic states while at the same time restraining allies, leaders often include in alliance treaties conditions that oblige allies to provide military assistance only if a member of the alliance is attacked by a state outside the alliance. However, other treaties may contain unconditional obligations to come to the defense of members of the alliance. Such alliances tend to arise from situations where some members of the alliance feel that their alliance partners need to have the flexibility even to engage in provocative behavior in order to deter the target of the alliance. Our analysis of alliance formation processes in the context of priorities that compete with each other provides a basis for two hypotheses. The first is that revisionist states with unconditional commitments from members of their alliance to come to their defense are more likely to initiate militarized conflict than states without such unconditional commitments. The second hypothesis is that revisionist states in alliances whose treaties stipulate that commitments to defend are conditional will be less likely to initiate militarized conflict than such states with allies who are committed to come to their defense without conditions. Statistical analyses of data generated with a view toward evaluations of both hypotheses (some of which provide new, more detailed categorizations of alliance treaties) suggest that they are valid.

Keywords

alliances conflict deterrence

Do formal alliance agreements cause alliance members to behave aggressively? Military alliances may deter external threats, but they may also create incentives for members of the alliance to behave aggressively (e.g. Smith, 1995; Snyder, 1984, 1997; Snyder & Diesing, 1977; Yuen, 2009). Alliance theory and empirical evidence suggest that governments often design alliance treaties, at least in part, to restrain aggressive behavior by alliance partners (Benson, 2011; Gelpi, 1999; Pressman, 2008; Schroeder, 1976; Snyder, 1984, 1997). Such treaties are designed to deter challenges to alliance members, but they also often contain provisions that oblige alliance members to provide military assistance to each other only if the target of the alliance initiates aggression against an alliance member.¹ Other deterrent types of treaties do not contain such conditions. We might therefore expect the unconditional types of commitments to be more likely to lead to aggressive behavior by alliance members. In this article, we build on extant studies of moral hazard in alliances to generate some testable propositions about the expected effects of these two different types of alliance commitments on the behavior of alliance members. To evaluate these conjectures, we estimate an empirical model that examines the likelihood that alliance members holding conditional and unconditional deterrent alliance agreements will initiate a militarized conflict against a targeted non-alliance member. We find that revisionist state leaders holding unconditional deterrent alliance commitments are more likely than those holding no agreement and conditional agreements to initiate conflict.

The effect of alliance commitments on allies’ behavior

The effect of third-party intervention on states’ behavior in crises is a primary focus of alliance and deterrence scholarship. Scholars have long held that having partners in conflicts leads to war expansion because fighting alongside a partner is more appealing than fighting alone (Altfeld & Bueno de Mesquita, 1979; Siverson & King, 1980). The focus of these early studies was on the possibility that alliances would increase the likelihood and spread of war. One way that alliance bonds might increase the probability of aggressive behavior would involve the expectation by one alliance member that it would receive assistance from its alliance partners in any conflict with the target of the alliance. For example, an alliance commitment may cause a state in a dispute over some issue or territory with another state to anticipate assistance from an ally in the event that the dispute escalates to war. While the anticipation of two allies pooling resources may deter adversaries from initiating challenges against those allies, the commitments by allies to each other may also lead them to feel more confident about the outcomes of conflicts with other states, which they will therefore be more willing to initiate. In other words, alliances create a dilemma for states that form them. This dilemma is characterized by the need to have an alliance agreement that is strong enough to deter an adversary but also with obligations sufficiently conditional or limited to restrain allies from provoking a conflict (Fearon, 1997; Jervis, 1994; Snyder, 1997; Snyder & Diesing, 1977).

Researchers have identified two possible ways alliance commitments might foster aggression on the part of an alliance partner. One is by creating an incentive for the ally to demand more in crisis bargaining (Yuen, 2009). The other is by emboldening the ally to entrap its alliance partners in war (Christensen & Snyder, 1990; Snyder, 1984, 1997). Both arise from the moral hazard effect of an alliance commitment, but entrapment involves the ally’s initiating a war and ensnaring its alliance partner in that war. According to Snyder (1984, 1997), entrapment refers to a situation where a government’s commitment of military assistance to an ally emboldens it to drag its alliance partner into a war that the partner would prefer to avoid. For entrapment to occur, then, it is not sufficient for an ally just to behave aggressively in crisis bargaining. The ally must initiate a war and the government that is committed to provide military assistance as a result of the alliance will unwillingly intervene on the ally’s behalf. This perspective of entrapment raises questions about why a country would unwillingly go to war for the ally, and why the ally would respond to the expectation of receiving assistance by initiating a war rather than trying instead to extract more benefits in conflict bargaining.

Some work has addressed the first question. One possible explanation for why a government, which can always renege on its commitment, would become entrapped if it would be unwilling to fight an ally’s war is that the anticipated benefit from the survival of the alliance is greater than the expected costs of joining its ally’s war (Christensen & Snyder, 1990; Snyder, 1994, 1997). A related explanation is that a government might assist an ally to prevent it from realigning with other countries (Crawford, 2003; Zagare & Kilgour, 2006). The accounts of entrapment all suggest that the government’s preferences at the time of alliance formation are such that it is willing to fight some wars for its ally and not others, but at war time it is even willing to fight the undesirable wars. This inconsistency is what emboldens the ally to behave aggressively in crises, though fewer studies explain why the ally would be so aggressive as to respond to an alliance commitment by starting a war. In fact, recent research casts doubt on the war enhancing effects of moral hazard (Yuen, 2009), arguing instead that alliance commitments might lead the ally to engage in aggressive bargaining behavior with little or no effect on the probability of war. In spite of its overemphasis on war, the main point of the early entrapment literature is that an alliance commitment might have a moral hazard effect on the behavior of the ally. This is the theme central to recent studies of moral hazard.

In the past few years, formal studies have taken up the subject of the moral hazard effect of alliance commitments. Rowlands & Carment (1998) formally analyze the impact that moral hazard has in the midst of an ongoing conflict involving human rights violations and find that third-party intervention emboldens the side it is protecting to fight harder rather than settle peacefully. While this model does not incorporate the role of outside assistance leading to conflict initiation, it does show that third-party intervention does not always promote peaceful resolution of conflict. Addressing this issue, Zagare & Kilgour (2003) show that in order to deter an attack, a third party must successfully signal to an adversary that it will intervene should it initiate a conflict with its ally to achieve deterrence, and to its ally that it will only intervene if the target of the alliance initiates the conflict to avoid moral hazard. This model shows that an unconditional alliance can cause an ally to be intransigent in interstate bargaining.

Yuen (2009) shows that a status quo-oriented state’s expectation of assistance from an intervener may cause a range of different behaviors in crisis bargaining. Consistent with other studies, she shows that intervention can create a disincentive for aggressors to initiate challenges against allies, and it can also cause allies to be overly aggressive in conflict bargaining. Her model also yields a novel result: the expectation of intervention by an ally might induce a state leader to make some limited concessions to the aggressor rather than behave aggressively and risk war. This outcome obtains when demands made by the aggressor are so small relative to the targeted ally’s costs of war that it concedes to the aggressor even though the prospective intervener is willing to fight.

The research on entrapment and recent moral hazard studies agree that, in many circumstances, the expectation of assistance from another country in times of war might lead the recipient of the assistance to behave aggressively in crises. Since unqualified alliance agreements signal alliance partners’ commitment to provide assistance to one another (Fearon, 1997; Morrow, 1994; Smith, 1996), then the existence of an alliance agreement might give rise to aggressive behavior. To our knowledge, however, there does not exist a systematic quantitative examination of this conjecture. This is due, in part, to the early lack of data on alliance content. With the introduction of new data on the content of alliance agreements (Benson, 2011; Leeds et al., 2002), it is now possible to determine with a systematic quantitative analysis whether a moral hazard effect exists with certain types of alliances. But it is also because the progression of recent studies about the relationship between alliance agreements and conflict using these new data began by focusing on the deterrent effects of alliances in the hands of prospective targets of aggression (Benson, 2011; Leeds, 2003) and then proceeded to studying the effects of the alliance agreements on the conflict initiation behavior of the alliance members (Johnson & Leeds, 2011).

However, another reason why scholars have not tested the long-held belief that some alliance commitments lead alliance partners to initiate conflict is that early research lacked clear guidance about which types of alliance commitments are more likely to embolden which kinds of alliance partners. Given the moral hazard effect of unqualified commitments of assistance, we might expect forward-looking governments to anticipate these negative wartime incentives and contract around them at the time of alliance formation or through renegotiation. Governments may design alliance agreements to mitigate the effects of moral hazard. There is widespread consensus that alliance agreements are often designed to restrain allies, and commitments lacking such restraining mechanisms can embolden alliance partners (Kim, 2011; Pressman, 2008; Snyder, 1997). Scholars have also made general observations that governments often incorporate conditions and flexibility into the terms of an alliance agreement to restrain the ally’s behavior (Fearon, 1997; Snyder, 1997), but research has not yet established which types of conditions and provisions in alliance agreements are likely to embolden and which will restrain.

Our analysis focuses on the effects of particular types of alliance agreements – conditional and unconditional deterrent commitments (Benson, 2011) – on the likelihood that a recipient of these alliance agreements is more or less likely to initiate a conflict. In our approach, we build on recent studies that have investigated the relationship between types of alliance agreements and conflict. Leeds (2003) uses Alliance and Treaty Obligations and Provisions (ATOP) data (Leeds et al., 2002) to analyze the likelihood that a challenger will initiate a militarized dispute (MID) against a country that holds a defensive alliance. Her results reveal a deterrent effect of alliances but leave open the question of whether a prospective initiator might become more aggressive as a result of an alliance not designed to restrain its allies. To analyze whether alliance commitments embolden alliance members, a model must focus on whether an alliance member is more or less likely to initiate a MID and the data must distinguish between different types of defensive alliances commitments. On the first objective, Johnson & Leeds (2011) study the impact of the broad category of defensive alliances on the probability a state holding such an alliance will initiate a conflict. Unsurprisingly, they do not find a moral hazard effect, because agreements designed to restrain and those that are not are combined into one category. On the second objective, Benson (2011) divides alliance treaties into categories of conditional and unconditional deterrent alliance commitments, but he does not examine the impact of these alliance agreements on the propensity of an alliance member to initiate a conflict. As a result, he finds that conditional alliance agreements deter, and unconditional commitments give rise to some conflict. However, since his model does not test the effects of an alliance on alliance members’ conflict behavior, it is not possible to infer a moral hazard effect from this finding.

To identify a moral hazard effect, a model should be specified to examine the impact of conditional and unconditional deterrent types of alliance on the likelihood a state holding such an alliance will initiate a conflict against a non-alliance member targeted by the alliance. In this article, we provide a test of the proposition that alliances designed to deter that do not include restraint mechanisms in the treaty will lead revisionist alliance members to behave aggressively toward states that are targeted by that alliance. Additionally, those alliances that are designed to restrain do not similarly embolden revisionist alliance partners. Such an investigation can make an important contribution by establishing empirically a long-held view in the alliance literature.

Admittedly, there is a presumption here that the formal language of alliance treaties matters. Altfeld & Bueno de Mesquita (1979) suggest that alliances tend to be general signals of shared preferences and therefore the specific language of the treaties may be of limited importance.² However, Leeds (2003), Leeds et al. (2002), and Benson (2011) all provide evidence that alliance treaties with different formal obligations have different impacts on the behavior of allies and their interactions with targets of the alliances. It is also possible that the formal language of a treaty may become obsolete if the preferences of the allies change over time. But to the extent that such disparities between the language of treaties and the preferences of allies might diverge over time, the data and the evidence we consider here will constitute a conservative test of our hypotheses, which stipulate in effect that treaty language will make a difference, even in the face of changing preferences for the signatories of the alliance treaties.

How alliances intended to deter can provoke conflict

Alliance models show that deterrent commitments of assistance are pacifying because they convey a costly signal to the alliance members and the adversaries of those alliance members that the members are more likely to prevail in a conflict. Knowing it is more likely to win, the ally will respond to an unqualified alliance signal by standing firmer against the adversary’s challenges. Its increased resistance causes the adversary to be less willing to challenge the ally (Smith, 1995; Yuen, 2009).

However, the ally’s knowledge that it is more likely to win with third-party assistance can be problematic, because a signal of unqualified support might also make the ally more aggressive. The ally’s expectation that a third party will provide guaranteed assistance in war increases its war payoff, which leads it to require a larger share of a disputed pie to avoid going to war. If an alliance guarantees support for a wide range of circumstances, then the ally’s aggressive behavior might be extended to crises not limited to the ally’s non-provocation of conflict or the defense of the ally’s status quo holdings. Thus, in addition to the deterrent benefits of alliances, there may also be conflict-enhancing externalities if allies are revisionist and they feel assured that the benefits of the alliance commitment extend to future conflicts of their choosing.

Allies are often revisionist, in the sense that they may not be satisfied with the status quo allocation of a disputed territory or issue. Additionally, some deterrent alliance agreements do condition military assistance on the non-provocation of conflict or even the initiation of conflict by an adversary. On the matter of revisionism in alliances, many studies of alliances and third-party intervention assume that third-party defenders and their allies share a preference for the status quo and are non-revisionist with respect to an adversary’s holdings. Snyder (1997) is an exception. A key feature of entrapment is that defenders are dragged against their will into an ally’s war. For a defender, who has agreed in a defense pact to fight a defensive war on the behalf of its ally, to be dragged into a war that is not required by or congruent with its interests suggests that the ally wishes to risk fighting for something in excess of its own defense. That there exists some disputed pie and the ally wishes to increase its share of that pie is an important aspect of moral hazard, creating the incentive for the ally to respond aggressively to a commitment. In our model specification, therefore, we assess the impact of alliance commitments on the behavior of revisionist states.

In addition to examining the effect of alliance commitments on the behavior of revisionist alliance members, we must also account for differences in the type of alliance commitment. Defenders who worry about moral hazard might be reluctant to extend blanket deterrent commitments to allies if so doing will lead those allies to initiate conflicts. In fact, many deterrent commitments, such as those that specifically stipulate that third parties are obligated to intervene only if their ally is attacked by a non-alliance member, are designed with a view toward restraining the ally. Benson (2011) identifies such alliances as conditional deterrence alliances. However, a non-trivial proportion of alliance treaties do not contain such conditions and some permit allies to engage in active defense or to take pre-emptive military action to deter a threat from states outside the alliance (Benson, 2011). These different types of alliance commitments share deterrent objectives, but they clearly differ from each other. Following Benson (2011), we define conditional deterrent alliances as those that condition third-party military assistance on an external attack against an alliance member and unconditional deterrent alliances as those that do not contain such conditions and, therefore, allow an ally to engage in pre-emptive military action for deterrent purposes. By definition, the unconditional commitments permit more assertive actions by alliance members than the conditional commitments. All previous theoretical and analytical models of alliances as well as the coding rules for prominent alliance datasets (COW and ATOP) are based on the assumption that all deterrent or defensive alliance treaties contain conditional commitments. From this explanation, we derive the following two hypotheses for analysis.

H1: A revisionist ally with an unconditional commitment from its defender is more likely to initiate a conflict than an ally without an unconditional deterrent commitment.

H2: A revisionist ally with a conditional deterrent alliance is less likely to initiate a conflict than an ally with an unconditional deterrent alliance.

It is reasonable to inquire as to the reasons why a government that knows its ally is revisionist would form an unconditional deterrent alliance with the ally. Forward-looking alliance partners likely estimate the level of external threat as well as moral hazard before settling on alliance terms. Prospective defenders may agree to an unconditional alliance with a revisionist ally knowing that moral hazard may arise if a threat is sufficiently imminent and/or the defender is sufficiently sympathetic to the ally’s interests in the outcome of the crisis that it is willing to tolerate the moral hazard costs so as to gain the primary deterrence benefits. Alternatively, the prospective defender may be uncertain about whether the ally is revisionist, and extend an unconditional alliance based on the mistaken belief that the ally is not revisionist. Thus, it is not implausible that governments will form unconditional deterrent alliances even though they are aware of the risk of moral hazard.

Research design and data

We assess our hypotheses using a directed-dyad design with conflict data spanning the years 1816–2000 retrieved from the EUGene dataset (Bennett & Stam, 2008). A directed-dyad design allows us to distinguish between prospective initiators and targets of a conflict. Similar to Senese & Vasquez (2005, 2008), we expect certain alliance patterns will correlate positively with conflict. However, our directed-dyad design enables us to explore the conditions under which certain leaders holding particular types of alliances will initiate conflict. Without directionality, a dyadic design treats the introduction of an alliance or the fact that one or the other state is revisionist as though it were a reagent added to a chemical reaction; there are no ‘sides’ to which an alliance can be added, so it is not possible to know if the addition of an alliance causes a state to initiate or be initiated against but just that an initiation occurred. Since our theory specifically investigates the impact of commitment mechanisms with conditions on initiation, we use a directed-dyadic design with the directed-dyad-year as the unit of observation.

We created a dataset of politically relevant directed dyads that includes pairs of states that are either contiguous or include a major power. We set our contiguity level at separation by open water of at most 400 miles, as done by Reed & Chiba (2010).³ Using politically relevant dyads excludes only cases in which states are substantially less likely to become involved in an interstate conflict.

We estimate a logistic regression with Huber-White robust standard errors and standard errors clustered on the directed dyads to test our hypotheses. The dependent variable for this model indicates the initiation of a militarized interstate dispute (MID). A MID is an event in which a state in the international system uses force, displays force, or threatens the use of force against another state. We code this variable 1 if State A initiated a dispute against State B in that directed-dyad-year and 0 otherwise. We use Zeev Maoz’s dyadic MID dataset (2005: version 2.0) as our source for the dependent variable. We consider only those states in a dispute involved on the originating side on the first day of the dispute as initiators and only those states involved in the dispute on the first day that were not on the originating side as targets. We do not consider so-called ‘joiners’ (which enter a dispute after the first day) to be dispute originators. So if a third party comes to State A’s aid a week after State A initiated a MID against State B, only State A is coded as having initiated the MID and the third-party is coded 0.

Our variables of interest in the model are commitment types. To translate data (Leeds et al., 2002) on commitments into the directed-dyad-year format, we identified the alliances possessed by potential initiators and targets targeting one another for each directed-dyad-year and constructed dichotomous variables for each alliance type, using 1 to indicate the presence of the specified alliance type. In most cases, states have many different alliances of various types at a given point in time. As we have pointed out, we expect that different alliance types will have different effects on the likelihood of conflict, so to ameliorate the confounding effect that a multitude of alliances could have, we try to isolate specific alliance types with our coding scheme.

We use Benson’s (2011) alliance data to specify our agreement categories. In this coding, each agreement is coded to identify its type, which states are party to the agreement, what obligations each ally has with respect to the agreement, and which states are targeted by the agreement. Table I shows the distribution of alliance commitment types in our data.⁴ The most prevalent alliance commitment is the conditional deterrent alliance, and this type is what most people think of when they consider alliances. Less common are those that we are most interested in, the unconditional deterrent agreement, which make up about 25% of all alliance agreements. Finally, the compellent agreement⁵ is the least likely commitment that a state will give to another.

Table I.

Distribution of alliance commitment types

	Frequency of commitments
Alliance commitment type	Number	Column %
Compellent	48	19.3
Unconditional deterrent	63	25.3
Conditional deterrent	138	55.4
Total	249	100.0

Many states are party to several alliance agreements at once, as mentioned above. Since our focus is on the emboldening effects of unconditional deterrent alliances, we need to create an indicator that isolates the effect of these types of agreements distinct from the potential emboldening effects of compellent agreements. To create such a variable, we started by coding whether the target state in each directed-dyad observation is party to a compellent alliance targeting the initiator. If the target state is not the beneficiary of such an agreement, we then assessed as separate variables whether the target state had an unconditional deterrent or conditional deterrent alliance targeting the initiator state. We then followed the same process for agreements that the initiator state possesses against the target state. Thus, both the initiator state and target state in each observation have two deterrent agreement categories both of which are coded 0 if the state either is not party to an agreement of the respective type or has a relevant compellent agreement. The deterrent agreement variables are coded 1 if the state possesses that type of agreement and does not possess a relevant compellent agreement.

Table II shows how these categories are distributed on the initiator side of the directed dyad, since it is the impact on those states that we are most concerned about. In our data, there are a total of 13,238 directed-dyad-years in which the initiator has an unconditional deterrent alliance. In 2,785 of these cases, the initiator also has a compellent alliance. It is not necessarily the case that these commitments are part of the same alliance agreement. Eliminating these instances of the presence of an unconditional deterrent alliance leaves us with 10,453 directed-dyad-year observations where the initiating state has an unconditional deterrent alliance.

As we mentioned above, we proceeded with our coding in this manner so as to avoid mistaking the aggressive effects of compellent types of alliances for the aggressive behavior hypothesized to result from unconditional deterrent types of alliances. Our theory makes no claims regarding the effect of compellent

Table II.

Distribution of deterrent alliance commitments on the initiator side in directed dyads, 1816–2000

	Frequency of deterrent commitments to initiator in directed dyad observations
Deterrent commitment type	Without compellent commitment		With compellent commitment		Total
Deterrent commitment type	Number	Col. %	Number	Col. %	Number	Col. %
Unconditional	10,453	11.2	2,785	62.9	13,238	13.5
Conditional	83,050	88.8	1,641	37.1	84,691	86.5
Total	93,503	100.0	4,426	100.0	97,929	100.0

alliance on conflict, but theoretical and empirical research has shown that initiator states possessing a relevant compellent alliance are much more likely to initiate conflict against the target (Benson, 2011; Leeds, 2003; Smith, 1995). In testing the effects of unconditional deterrent alliances, we wish to isolate as a category prospective initiators holding only those types of alliances. We would risk misattributing conflict to deterrent alliances if we discovered they were associated with conflict when prospective initiators also held relevant compellent alliances. Excluding from the deterrent categories cases where compellent alliances were present (but leaving them together in compellent categories) effectively isolates the deterrent alliances and, therefore, produces the most rigorous possible test of our hypotheses regarding the agreement types of interest here. It should also be noted that some initiators and targets possessed both unconditional and conditional deterrent alliances valid against the target. This is not problematic, as we have clear and distinct expectations for the emboldening effects of these alliances that are conditioned upon the revisionist character of the ally. When a revisionist state holds both a conditional and unconditional deterrent alliance, we expect, all else equal, the impact of the unconditional deterrent alliance will trump the conditional deterrent.

To identify revisionist initiators in dyads, we use the same indicator for a revisionist state in the COW dataset that Senese & Vasquez (2005, 2008) use in their work. A revisionist state has made public statements about its desires to revise the status quo with respect to territory, regime, policy, or other area.⁶ Our variable is coded 1 if the initiator state in the directed dyad has made any public statements regarding these areas in reference to the target state and 0 otherwise.

Our hypotheses make claims regarding the effect of different deterrent alliance types on conflict when these alliances are possessed by states with revisionist preferences. To make inferences about these effects, we construct two interaction variables: the first is the product of the unconditional deterrent alliance variable for the prospective initiator with its value on the revisionism variable, and the second combines the value on revisionism with the conditional deterrent alliance variable on the initiator side. These variables, then, will capture the effect on conflict of providing one of these alliances to a revisionist ally.

We also include several other variables that are conventional controls shown consistently to affect the initiation of conflict or war also expected to impact alliance formation decisions. Jointly democratic pairs of states have been observed to experience fewer conflicts than other pairs of states (Bremer, 1992; Maoz & Abdolali, 1989; Maoz & Russett, 1993; Ray, 1995). Following established practice, we used the POLITY IV dataset (Marshall, Jaggers & Gurr, 2002) to identify jointly democratic dyads. We coded a dyad as jointly democratic if both states in the dyad scored six or higher on the POLITY IV democracy scale in that year.

We also control for the presence of contiguity in the dyad. Contiguous states experience both more conflicts (Bremer, 1992; Vasquez, 1993) and more severe conflict (Moul, 1988; Vasquez, 1993) than other pairs of states. Prior researchers have suggested that this may result because contiguity increases opportunities for disputes over territory, immigration, trade, and other matters. Since we only include politically relevant dyads, each dyad in the dataset is either contiguous or contains at least one major power (or both). We expect that contiguous pairs of states will be more likely to experience conflict than non-contiguous dyads containing a major power. We coded contiguity as a dichotomous variable, taking on the value of 1 if states in the dyad were contiguous on land or if they were separated by at most 400

Table III.

Cross-tabulation of revisionist state preferences, militarized dispute initiation and membership in an unconditional deterrent alliance, 1816–2000

Revisionist initiator?			Initiated a militarized interstate dispute?
Revisionist initiator?			Yes	No	Total
Yes	Initiator with unconditional deterrent commitment (and no compellent commitments)	Yes	109	50	159
Yes		No	1,505	1,562	3,067
		Total	1,614	1,612	3,226
No	Initiator with unconditional deterrent commitment (and no compellent commitments)	Yes	21	10,273	10,294
No		No	564	174,242	174,806
		Total	585	184,515	185,100

miles of water.⁷ Otherwise, this variable takes on a value of 0. This usage is consistent with our definition of political relevance.

We additionally expect that the relative power of prospective disputants will impact the incidence of conflict within a dyad. The theoretical expectation is that states are most likely to experience conflict when the expected outcome of conflict is uncertain (Fearon, 1995; Meirowitz & Sartori, 2008). For instance, Reed (2003) argues that at power parity the informational uncertainty leads to greater variance in the beliefs of the opponents about their counterpart’s ability to win in a conflict. He shows that this increased variance is associated with higher levels of conflict. We measure uncertainty in the dyad on a continuous scale from 0 to 1 and construct this measure using capabilities ratios from the Correlates of War project (Singer, 1988). Our measure increases as uncertainty in the dyad increases; 0 indicates absolute preponderance for one member the dyad, and 1 indicates absolute parity.⁸

States with similar preferences over outcomes in the international system should be less likely to initiate conflicts with one another. The traditional proxy for preference congruence is similarity in the alliance portfolios of states; such similarities are theorized to signal common interests (Bueno de Mesquita, 1981). We measure similarity of preferences by estimating the similarity of states’ alliance portfolios and choose the weighted S-score (Signorino & Ritter, 1999) as the measure of portfolio similarity. This measure is established on a scale of –1 to 1; we recoded this onto a scale of 0 to 1.

Finally, we must address the expectation of temporal dependence (Beck, Katz & Tucker, 1998) in conflict initiation. States that have recently experienced a conflict are more likely to experience another conflict than states that have not experienced a conflict in some time. We use a scaled cubic polynomial of the number of years in which a dyad has experienced peace as our temporal control, as suggested by Carter & Signorino (2010: 24–25).⁹ This measure has been demonstrated to perform at least as well as the cubic splines Beck, Katz & Tucker (1998) recommend, outperforms splines in some cases, and has the virtue of intuitive interpretation.

Data analysis

We begin our data analysis with a cross-tabulation of our variables of interest in Table III: MID initiation, revisionism, and the presence of an unconditional deterrent alliance. That table shows that MID initiations are infrequent. There are only 2,199 initiations out of 188,326 total observations. Of the MID initiations, almost three-fourths of those were perpetrated by revisionist states. In fact, revisionist states initiated a dispute in more than half of the observations where the initiator was revisionist. What about the unconditional deterrent alliances? There are two interesting findings to point out with respect to these. First, when a revisionist initiator had one of these alliance agreements, it initiated a MID more than 68% of the time. Second, when the state that had one of these alliances was non-revisionist, it only

Table IV.

Logit estimate of the effects of deterrent alliances on the initiation of militarized interstate disputes, 1816–2000

	Model 1A	Model 1B	Model 1C
MID initiation	Baseline model	No controls	Full model	δX/δY
Joint democracy (d)	–0.8244**	–	–0.3038**	–0.0009**
Joint democracy (d)	(0.126)		(0.105)	(0.000)
Contiguity (d)	1.4031**	–	0.8031**	0.0029**
Contiguity (d)	(0.100)		(0.104)	(0.000)
Capabilities ratio	1.0268**	–	0.1466	0.0005
Capabilities ratio	(0.132)		(0.160)	(0.001)
S-score	–0.8131**	–	–0.4967**	–0.0015**
S-score	(0.110)		(0.146)	(0.000)
Initiator has unconditional deterrent alliance (d)	0.1448	–0.4714	–0.2379	–0.0007
Initiator has unconditional deterrent alliance (d)	(0.162)	(0.298)	(0.294)	(0.001)
Revisionist initiator X initiator has unconditional deterrent alliance (d)	–	1.2798**	0.8655*	0.0043+
		(0.314)	(0.345)	(0.003)
Initiator has conditional deterrent alliance (d)	0.0544	–	–0.4180**	–0.0013**
Initiator has conditional deterrent alliance (d)	(0.079)		(0.129)	(0.000)
Revisionist initiator X initiator has conditional deterrent alliance (d)	–	–	0.5491**	0.0023*
			(0.174)	(0.001)
Revisionist initiator (d)	–	5.5221**	5.1067**	0.3182**
Revisionist initiator (d)		(0.103)	(0.130)	(0.024)
Target has unconditional deterrent alliance (d)	–0.1692	–	–0.0738	–0.0002
Target has unconditional deterrent alliance (d)	(0.174)		(0.179)	(0.001)
Target has conditional deterrent alliance (d)	–0.1868*	–	–0.0282	–0.0001
Target has conditional deterrent alliance (d)	(0.079)		(0.088)	(0.000)
Constant	–3.7173**	–5.3453**	–5.3238**	–
Constant	(0.103)	(0.117)	(0.142)
Pr(Y=1\|X)	–	–	–	0.0031
Effect of an unconditional deterrent alliance to a revisionist initiator	–	0.8084**	5.7343**	0.3218
Effect of a conditional deterrent alliance to a revisionist initiator	–	–	5.2378**	0.3192
N	188326	188326	188326	188326
Log-Lik intercept only	−11972.0391	−11972.0391	−11972.0391	−11972.0391
Log-Lik full model	−10501.6416	−6139.8811	−6042.5933	−6042.5933
Likelihood ratio Chi²	2940.795**	11664.316**	11858.892**	–
McFadden’s pseudo R²	0.123	0.487	0.495	–
AIC	0.122	0.065	0.064	–
BIC’	−2807.190	−11591.440	−11688.849	–

Marginal effects; Standard errors in parentheses.

(d) for discrete change of dummy variable from 0 to 1.

+ p < 0.10, * p < 0.05, ** p < 0.01.

initiated in 21 of 10,294 observations. These findings certainly suggest that our argument and hypotheses could be correct and, thus, are a nice introduction for our regression models.

Table IV shows the results from Model 1. Model 1A focuses on interstate conflict. In this model, we include only the standard controls and agreement type indicators. Using this model we can establish a baseline for comparing our findings across models and with other conflict theories. The coefficients for the control variables in Table IV all have the expected signs. We also find, like Benson (2011), that conditional deterrent alliances only slightly deter, and unconditional deterrent alliances show no deterrent effect. As expected, the alliance indicators on the initiator side show no effect in the baseline model.

Model 1B contains only variables of special interest. In this model we show that the control variables are not driving our results. The results provide evidence supporting Hypothesis 1. When a prospective initiator has an unconditional deterrent alliance, this has no significant

Table V.

Predicted probabilities of conflict given discrete change from no alliance to only unconditional deterrent alliance and conditional deterrent alliance to initiator in presence and absence of a revisionist initiator

	Non-revisionist initiator	Revisionist initiator
Protégé does not have an alliance	0.0031	0.3213
Protégé has an unconditional deterrent alliance	0.0024	0.3249
Protégé has a conditional deterrent alliance	0.0018	0.3223

effect on the likelihood that it will initiate conflict. However, a prospective revisionist initiator will be more likely to initiate conflict when it has an unconditional deterrent alliance. Finally, Model 1C is the full model with both the control variables and variables of interest included. The discussion that follows will focus on this version of the model.

We use various goodness-of-fit measures presented at the bottom of the table to assess which of our models performs best. Each of the models produces a significant Likelihood Ratio Chi² indicating that the specified model is a significant improvement over a null model. From McFadden’s R² we see that Model 1C provides the most accurate predictions. Although not reported, the McFadden’s Adjusted-R² is only slightly different from the unadjusted version, which indicates that the superior predictions produced by this model do not simply depend on its larger number of variables. As for the two information criteria we report, smaller values for Akaike’s (1973) information criterion (AIC) indicate a better fit, suggesting again that the full model is best. Finally, using the guidelines for assessing the Bayesian Information Criterion (BIC) from Raftery (1995), we determine that the full model is the most preferred since its BIC is more negative than the other two and the absolute difference between it and the BIC of either of the other two is greater than 10.

Our theoretical argument contends that revisionist allies holding certain deterrent types of alliances will behave differently than states without such alliance commitments. Hypothesis H1 states that revisionist allies are more likely to initiate conflict if they have an unconditional deterrent alliance than if they do not. An evaluation of this hypothesis needs to focus on whether or not revisionist allies have unconditional alliance commitments. The variable identifying the existence of an unconditional deterrent alliance held by the prospective initiator is not statistically significant in any of the three versions of our model. However, when the interaction term that corresponds to the combination of unconditional alliance commitments for states with revisionist preferences is included in the model, the unconditional deterrent alliance variable, which in the model with the interactions is an indicator for non-revisionist prospective initiators holding unconditional deterrent alliances, becomes negatively associated with conflict initiation. Thus, we find that extending an unconditional deterrent alliance to a non-revisionist state does not increase the likelihood that that state will initiate conflict. However, if revisionist allies have unconditional alliance commitments, the moral hazard effect of such alliance commitments is dramatic. For revisionist states, having an unconditional deterrent alliance increases the likelihood of conflict initiation exponentially, and this effect is statistically significant. In fact, revisionism combined with the presence of an unconditional deterrent alliance is the best predictor of conflict in the model.

Table V shows the predicted probabilities of conflict in the presence of the alliances types and states with and without revisionist preferences.¹⁰ When the initiator does not have revisionist preferences, an unconditional deterrent alliance leads to a 22.6% decrease in the likelihood of conflict, but a revisionist ally with an unconditional deterrent alliance will initiate conflict 32.49% of the time. Furthermore, this emboldening effect is unique to this type of deterrent alliance; conditional deterrent alliances do not have a significant effect on the initiators' decision to initiate conflict.

Hypothesis H2 addresses the difference in the effects of unconditional and conditional deterrent alliances. Conditional deterrent alliances are given to allies by third parties more frequently, perhaps because they are also more effective at restraining revisionist allies. To test H2, we compare the likelihood that prospective revisionist initiators will initiate conflict when they have an unconditional versus a conditional deterrent alliance. We find that the coefficient of the combined effect of a conditional deterrent alliance to a revisionist initiator is smaller than the combined effect of an unconditional deterrent alliance to a revisionist initiator. In Model 1A without controls for revisionist preferences, initiators with conditional deterrent agreements are no more or less likely to initiate a conflict than states without them. Once we introduce a variable interacting these agreements with revisionist prospective initiators, we find that non-revisionist states are significantly less likely to initiate conflict when they hold one of these alliances. Revisionist protégés, on the other hand, are not less likely to initiate conflict; in fact, the coefficient on this variable is positive. Looking again at Table V, the predicted probability of conflict increases when a revisionist protégé has a conditional deterrent alliance but not to the same degree as an unconditional deterrent alliance. We conclude that our model provides support for H2.

Conclusion

This article argues that deterrent alliances might lead alliance members to initiate conflict, a finding that has not previously been thoroughly analyzed. One viewpoint in the literature is that alliances might lead to conflict when an ally can entrap another alliance member in a war (Snyder, 1984, 1997; Christensen & Snyder, 1990). More recent studies focus less on entrapment and more on the effects of moral hazard on conflict bargaining (Yuen, 2009). Given these incentives for aggression in alliances, signatories often find it in their interests to design alliances to restrain potentially dangerous alliance partners (Crawford, 2003; Pressman, 2008; Snyder, 1984, 1987; Zagare & Kilgour, 2003). Yet, there has been less research on how different mechanisms in alliance agreements actually affect alliance members’ decisions to initiate conflicts. We focus on two types of deterrent alliance agreements in Benson (2011): unconditional deterrent and conditional deterrent alliances. The unconditional types allow pre-emptive violence by an ally and still guarantee third-party military support.

The statistical analysis we present supports our claim that revisionist countries holding unconditional deterrent agreements are more likely to initiate conflict than if they had not been given an alliance or had been given a conditional deterrent alliance instead. By distinguishing between the two types of deterrent alliances to examine the effects of alliance agreements on conflict behavior of alliance members, we offer some insight about how deterrent alliances can give rise to moral hazard if the formal agreement does not explicitly impose conditions on obligations of alliance members to supply assistance to potentially revisionist allies. A government may be willing to form unconditional deterrent alliances, even with the possible moral hazard externality, if it is unaware of how revisionist the ally is or the security benefits of a strong alliance outweigh the costs of moral hazard.

Footnotes

Replication Data

The dataset and do-files for the empirical analysis in this article can be found at .

1

The literature on alliance restraint also makes it clear that alliances may restrain allies in ways including but not limited to conflict initiation. For instance, Weitsman (1997, ) discusses how an alliance can be used to ‘tether’ two adversarial states and promote peace between them. Our main focus on restraint concerns restraining an ally from attacking a state outside of the alliance.

2

More specifically, what Altfeld & Bueno de Mesquita (1979: 98) argue is that both defense pacts and ententes indicate a measure of utility for joining an ally in the event that it is attacked, even though ententes involve no specific obligation to do so. Ententes, they point out, ‘do not preclude war entry, and that is likely to be a suggested course of action during consultations’.

3

We also created datasets at smaller levels of contiguity (24-mile and 150-mile), and our results were unaffected.

4

This table was adapted from Table 3 in . For each row, we summed the number of alliances with that type of provision as found in column ‘No.’ under heading ‘Total’. For instance, for Compellent, we added rows UC, CC, UC & CC, UC & AD, UC & CD, CC & AD, CC & CD, and CC & PD.

5

See for more on compellent alliances. These alliances contain provisions that specifically threaten the target state(s) with offensive action, and in some cases the agreement specifies a particular compellent demand that the parties expect the target to meet.

6

The COW data documentation is unclear on what the ‘other’ category exactly means.

7

When we estimated our models on the datasets with a smaller level of contiguity for political relevance, we adjusted the level of contiguity for this variable accordingly.

8

The measure can be calculated from a measure of preponderance (initiator capabilities divided by the sum of capabilities in the dyad) as follows: subtract ½, take the absolute value of the result, and multiply by 2. This rescales the [0,1] preponderance measure onto a new [0,1] scale such that ½ is at zero and extreme values are at 1. Then subtract 1 and take the absolute value again to reverse the scale.

9

Carter & Signorino (2010) suggest using t, t², and t³ or a scaled version of this variable due to potential instability that can be caused by the large range of values for t, t², and t³. We use t, t²/100, and t³/1000.

10

These values are based on Model 1C.

References

Akaike

Hirotogu

(1973) Information theory and an extension of the maximum likelihood principle. In: Petrov

Csaki

(eds) Second International Symposium on Information Theory. Budapest: Akademiai Kiado, 267–281.

Altfeld

Michael F

Bueno de Mesquita

Bruce

(1979) Choosing sides in wars. International Studies Quarterly 23(1): 87–112.

Beck

Nathaniel

Katz

Jonathan N

Tucker

Richard

(1998) Taking time seriously: Time-series–cross-section analysis with a binary dependent variable. American Journal of Political Science 42(4): 1260–1288.

Bennett

D Scott

Stam

Allen

(2008) EUGene: Expected Utility Generation and Data Management Program. University Park, PA: Department of Political Science, Pennsylvania State University.

Benson

Brett V

(2011) Unpacking alliances: Deterrent and compellent alliances and their relationship with conflict, 1816–2000. Journal of Politics 73(4): 1111–1127.

Bremer

Stuart A

(1992) Dangerous dyads: Conditions affecting the likelihood of interstate war, 1816–1965. Journal of Conflict Resolution 36(2): 309–341.

Bueno de Mesquita

Bruce

(1981) The War Trap. New Haven, CT: Yale University Press.

Carter

David B

Signorino

Curtis S

(2010) Back to the future: Modeling dependence in binary data. Political Analysis 18(3): 271–292.

Christensen

Thomas J

Snyder

Jack

(1990) Chain gangs and passed bucks: Predicting alliance patterns in multipolarity. International Organization 44(2): 137–168.

10.

Crawford

Timothy W

(2003) Pivotal Deterrence: Third-Party Statecraft and the Pursuit of Peace. Ithaca, NY: Cornell University Press.

11.

Fearon

James D

(1995) Rationalist explanations for war. International Organization 49(3): 379–414.

12.

Fearon

James

(1997) Signaling foreign policy interests: Tying hands versus sinking costs. Journal of Conflict Resolution 41(1): 68–90.

13.

Gelpi

Christopher

(1999) Alliances as instruments of intra-allied control. In: Haftendorn

Helga

Keohane

Robert O

Wallander

Celeste A

(eds) Imperfect Unions: Security Institutions Over Time and Space. New York: Oxford University Press, 107–139.

14.

Jervis

Robert

(1994) What do we want to deter and how do we deter it? In: Benjamin Ederington

Mazarr

Michael

(eds) Turning Point: The Gulf War and U.S. Military Strategy. San Francisco, CA: Westview, 117–136.

15.

Johnson

Jesse C

Leeds

Brett Ashley

(2011) Defense pacts: A prescription for peace? Foreign Policy Analysis 7(1): 45–65.

16.

Kim

Tongfi

(2011) Why alliances entangle but seldom entrap states. Security Studies 20(3): 350–377.

17.

Leeds

Brett Ashley

(2003) Do alliances deter aggression? The influence of military alliances on the initiation of militarized interstate disputes. American Journal of Political Science 47(3): 427–439.

18.

Leeds

Brett Ashley

Ritter

Jeffrey M

Mitchell

Sara McLaughlin

Long

Andrew G

(2002) Alliance treaty obligations and provisions, 1815–1944. International Interactions 28(3): 237–260.

19.

Maoz

Zeev

(2005) Dyadic MID dataset (version 2.0) (http://psfaculty.ucdavis.edu/zmaoz/dyadmid.html).

20.

Maoz

Zeev

Abdolali

Nasrin

(1989) Regime types and international conflict, 1816–1976. Journal of Conflict Resolution 33(1): 3–35.

21.

Maoz

Zeev

Russett

Bruce

(1993) Normative and structural causes of democratic peace, 1946–1986. American Political Science Review 87(3): 624–638.

22.

Marshall

Monty G

Jaggers

Keith

Gurr

Ted Robert

(2002) Polity IV Project Political Regime Characteristics and Transitions, 1800–2002. College Park, MD: Center for International Development and Conflict Management, University of Maryland.

23.

Meirowitz

Adam

Sartori

Anne

(2008) Strategic uncertainty as a cause of war. Quarterly Journal of Political Science 3(4): 327–352.

24.

Morrow

James D

(1994) Alliances, credibility, and peacetime costs. Journal of Conflict Resolution 38(2): 270–297.

25.

Moul

William

(1988) Great power nondefense alliances and the escalation to war of conflicts between unequals, 1815–1939. International Interactions 15(1): 25–43.

26.

Pressman

Jeremy

(2008) Warring Friends: Alliance Restraint in International Politics. Ithaca, NY: Cornell University Press.

27.

Raftery

Adrian E

(1995) Bayesian model selection in social research. Sociological Methodology 25(1): 111–163.

28.

Ray

James Lee

(1995) Democracy and International Conflict. Columbia, SC: University of South Carolina Press.

29.

Reed

William

(2003) Information, power, and war. American Political Science Review 97(4): 633–641.

30.

Reed

William

Chiba

Daina

(2010) Decomposing the relationship between contiguity and militarized conflict. American Journal of Political Science 54(1): 61–73.

31.

Rowlands

Dane

Carment

David

(1998) Moral hazard and conflict intervention. In: Wolfson

Murray

(ed.) The Political Economy of War and Peace. Norwell, MA: Kluwer Academic, 267–286.

32.

Schroeder

Paul W

(1976) Alliances, 1815–1945: Weapons of power and tools of management. In: Knorr

Klaus

(ed.) Historical Dimensions of National Security Problems. Beverly Hills, CA: Sage, 227–262.

33.

Senese

Paul D

Vasquez

John A

(2005) Assessing the steps to war. British Journal of Political Science 35(4): 607–633.

34.

Senese

Paul D

Vasquez

John A

(2008) The Steps to War: An Empirical Study. Princeton, NJ: Princeton University Press.

35.

Signorino

Curtis S

Ritter

Jeffrey M

(1999) Tau-b or not Tau-b: Measuring the similarity of foreign policy positions. International Studies Quarterly 43(1): 115–144.

36.

Singer

J David

(1988) Reconstructing the Correlates of War capabilities dataset on material capabilities of states, 1816–1985. International Interactions 14(2): 115–132.

37.

Siverson

Randolph M

King

Joel

(1980) Attributes of national alliance membership and war participation, 1915–1965. American Journal of Political Science 24(1): 1–15.

38.

Smith

Alastair

(1995) Alliance formation and war. International Studies Quarterly 39(4): 405–425.

39.

Smith

Alastair

(1996) To intervene or not to intervene: A biased decision. Journal of Conflict Resolution 40(1): 16–40.

40.

Snyder

Glenn H

(1984) The security dilemma in alliance politics. World Politics 36(4): 461–495.

41.

Snyder

Glenn H

(1997) Alliance Politics. Ithaca, NY: Cornell University Press.

42.

Snyder

Glenn H

Diesing

Paul

(1977) Conflict Among Nations. Princeton, NJ: Princeton University Press.

43.

Vasquez

John A

(1993) The War Puzzle. Cambridge: Cambridge University Press.

44.

Weitsman

Patricia

(1997) Intimate enemies: The politics of peacetime alliances. Security Studies 7(1): 156–193.

45.

Weitsman

Patricia

(2004) Dangerous Alliances: Proponents of Peace, Weapons of War. Stanford, CA: Stanford University Press.

46.

Yuen

Amy

(2009) Target concessions in the shadow of intervention. Journal of Conflict Resolution 53(5): 745–773.

47.

Zagare

Frank

Kilgour

Marc

(2003) Alignment patterns, crisis bargaining, and extended deterrence: A game-theoretic analysis. International Studies Quarterly 47(4): 587–615.

48.

Zagare

Frank C

Kilgour

D Marc

(2006) The deterrence-versus-restraint dilemma in extended deterrence: Explaining British policy in 1914. International Studies Review 8(4): 623–641.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB