European Economic Review

European Economic Review 56 (2012) 54 71 Contents lists available at ScienceDirect European Economic Review journal homepage: www.elsevier.com/locate/eer Trade, conflict, and political integration: Explaining the heterogeneity of regional trade agreements Vincent Vicard Banque de France, France article info Article history: Received 7 December 2009 Accepted 29 June 2011 Available online 21 July 2011 JEL classification: D74 F15 F51 F52 Keywords: International relations Regionalism Trade War abstract Many historians argue that the main goal of European trade integration was the preservation of peace. This paper investigates whether this reasoning is relevant for the EU and other regional trade agreements (RTAs). I provide empirical evidence that customs unions and common markets (deep RTAs) do reduce the probability of war between members. Partial scope and free trade agreements (shallow RTAs) however have no effect on war probabilities. Accordingly, international insecurity has a differential impact on incentives to create RTAs. Deep RTAs are signed between countries that are involved in many interstate disputes and that have low trade costs with the rest of the world, whereas the opposite is true for shallow RTAs. & 2011 Elsevier B.V. All rights reserved. 1. Introduction The European Union (EU) is unquestionably the most integrated regional trade agreement (RTA) in the world, and a yardstick for other regions of the world. Many historians argue that the main goal of the European integration process was the preservation of peace after three increasingly destructive wars in Europe in less than a century. This view is illustrated by Robert Schuman s proposal for the creation of the European Coal and Steel Community, the forerunner to EU: by pooling basic production and by instituting a new High Authority, whose decisions will bind France, Germany and other member countries, this proposal will lead to the realization of the first concrete foundation of a European federation indispensable to the preservation of peace. 1 This paper investigates whether the reasoning linking security and regional trade integration is relevant for Europe and other regions of the world by asking two questions: do RTAs prevent the outbreak of war and is international security a motive for RTA creation? The literature in international relations identifies two channels through which regional trade integration is likely to affect international insecurity (Bearce, 2003). 2 First, since war disrupts bilateral trade (Martin et al., 2008; Glick and Taylor, 2010), an RTA increases the opportunity cost of war by increasing intra-regional trade (Martin et al., 2008; Polachek, 1980; Oneal and Russett, 1999). Second, supranational institutions created in relation to regional integration promote the exchange of information on military capabilities and resolve and patience in conflicts, through formal security/military substructures, joint military exercises and forums of defence ministers. Moreover, regular meetings of head of states and high level officials or the E-mail address: vincent.vicard@banque-france.fr 1 Declaration of 9 May 1950, http://europa.eu/abc/symbols/9-may/decl_en.htm. 2 Schiff and Winter (2003) identify a third channel related to access to raw materials. 0014-2921/$ - see front matter & 2011 Elsevier B.V. All rights reserved. doi:10.1016/j.euroecorev.2011.06.003

V. Vicard / European Economic Review 56 (2012) 54 71 55 existence of an executive secretariat create habits of negotiation and build trust between political leaders. 3 International institutions are thus likely to reduce asymmetries of information in conflicts and to mitigate problems of credible commitment in interstate negotiations, which reduces the probability that a dispute escalates into war (Fearon, 1995; Grossman, 2004). 4 Supranational institutional frameworks however differ greatly depending on the form of regional trade integration. Creating a customs union requires agreement on a common external tariff and revenue distribution between member states. A common market requires more comprehensive political institutions to agree on a broader set of issues (harmonization of regulation and standards, free movement of goods and factorsy), 5 whereas a free trade agreement or partial scope arrangement involves little or no political or institutional integration. 6 According to the political integration criterion, two categories of RTAs can be distinguished: deep (customs unions and common markets) and shallow (partial scope and free trade agreements) RTAs. Only deep RTAs require a significant common institutional framework likely to promote the negotiated settlement of conflicts and support peace between members. If countries can design regional trade agreements to pacify interstate relations, international security should affect decisions ontrade policy.in apurely economic framework inwhich RTAshave no effect on war probabilities, two partners that have more issues of dispute would have fewer incentives to create an RTA since disputes may escalate into war and disrupt bilateral trade. 7 Conversely, if an RTA reduces the probability of war between members, then having more disputes may increase the likelihood of an RTA being set up. To the extent that only deep RTAs pacify relations between members, international insecurity should have a differential effect on incentives to create RTAs (and should affect both the choice of form of trade integration and the choice of partner). In a nutshell, a history of conflicts should enhance the creation of deep RTAs, but not shallow ones. In this paper, I set out an empirical analysis of the relevance of international security in the creation of shallow and deep RTAs. The motives for choosing different strategies of integration are based on the premise that a deep RTA reduces the probability of war between members while a shallow RTA does not. I first test this proposition using data on militarized interstate disputes from the Correlates of War (COW) project covering the 1950 1991 period. I address the selection issue due to the heterogeneity of dispute occurrence across country pairs using a bivariate probit model accounting for selection and event data from Kinsella and Russett (2002) to measure interstate dispute occurrence. I find a sizeable impact of membership in a deep RTA: it reduces the probability of a dispute escalating into war by two-thirds. Membership of a partial scope or free trade agreement has no significant effect on war probabilities. Second, I test the determinants of the likelihood of deep and shallow RTAs on a cross-section of 2814 unique country pairs in 2005. The endogeneity bias related to past membership of RTAs and omitted variables is addressed by controlling for several co-determinants of regionalism and conflict and by implementing an instrumental variable strategy. I find that deep RTAs are signed between countries that have many interstate disputes and that have low trade costs with the rest of the world, whereas the opposite is true with respect to shallow RTAs. These empirical results provide strong support for the differential effect of international insecurity on incentives to create deep and shallow RTAs. Besides the reduction of tariffs, this paper explicitly emphasizes the role of RTAs as a regulating mechanism for interstate relations. By offering empirical evidence on the choice of RTA partners as well as the form of regional integration, this paper complements Baier and Bergstrand s (2004) analysis of the economic determinants of RTAs. This paper is related to the theoretical literature on the endogenous formation of RTAs that emphasizes non-traditional gains from regional integration. 8 This strand of the literature explicitly recognizes various motives for regional integration and identifies distinct problems that a trade agreement may solve (Maggi and Rodriguez-Clare, 1998; Mitra, 2002; Limao, 2007). Regional trade integration may indeed provide non-traditional gains and help solving problems of time inconsistency, signaling, insurance, cooperation and security (Fernandez and Portes, 1998; Whalley, 1998; Schiff and Winters, 1998). In such a framework, RTAs are not only regarded as engines of preferential liberalization but also as international institutions providing public goods to their members (Limao, 2007; Alesina et al., 2005). These papers however only consider the case of free trade agreements or customs unions, or do not distinguish between RTAs according to the form they take. 9 The usual classification of RTAs, derived from Balassa (1961), considers regional trade integration as a step-by-step process leading to economic union, through free trade area, customs union and common market. The underlying assumption is that more integrated arrangements provide deeper trade integration. 10 Vicard (2009) however 3 For instance, Manzetti (1993/94) reports that discussions of sensitive policy issues such as nuclear proliferation concerns have taken place within the MERCOSUR institutions. 4 Jackson and Massimo (2007) also show, in a setting where countries are at war because of the political biases of their leaders that when state leaders lack the ability to credibly commit to a negotiated deal, the scope for negotiated settlement of disputes is reduced. 5 See, for instance, Alesina and Wacziarg (1999) for a detailed mapping of policy areas carried out at the EU level, and Bouzas and Soltz (2001) concerning the institutional framework of MERCOSUR. 6 The ASEAN free trade agreement is an illustrative example, with weak regional institutions in order to limit any supranationalism (Best, 2005). Pomfret (1997) also emphasizes how the will to limit political integration has been fundamental to the creation of NAFTA. 7 This reasoning assumes either that negotiating or implementing an RTA involves costs or that increasing trade integration increases the number of disputes and thus the probability of trade disruption in the future. 8 The optimal tariff strand of the literature focuses on traditional trade gains (Riezman, 1985; Yi, 1996; Ornelas, 2005). 9 From an empirical point of view, Mansfield and Pevehouse (2000) and Martin et al. (2010) also investigate the impact of RTAs on security without differentiating between RTAs. 10 In his seminal paper, Balassa (1961) also mentions social integration, but he dismisses this second criterion.

56 V. Vicard / European Economic Review 56 (2012) 54 71 shows that different kinds of RTAs have, on average, a similar effect on intra-regional trade. Empirical evidence of gradual regional integration processes is also lacking: out of the 18 customs unions created worldwide since 1948, 14 have been created directly as such. This paper offers an alternative explanation for the choice of different strategies of regional integration. The form of RTAs reflects different institutional arrangements that provide different non-traditional gains to their members. The paper proceeds as follows. The next section presents data on interstate dispute and war. Section 3 presents the analysis of the impact of RTAs on war probabilities. Section 4 investigates the determinants of the creation of deep and shallow RTAs, and Section 5 concludes. 2. Data on interstate dispute and war The probability of war between two countries i and j at date t is the probability of occurrence of a dispute between i and j multiplied by the conditional probability that the dispute escalates into war: Prðwar ijt Þ¼Prðdispute ijt ÞPrðescalation ijt 9dispute ijt Þ: ð1þ The probability of dispute varies by country pair and over time. Disagreements over territorial issues or religious and ethnic minorities are more likely to occur between neighboring countries. Policies fostering international trade integration are also likely to increase dispute occurrence. In the same manner, many factors affect the probability of a dispute escalating into war. For instance, one of the few accepted regularities in international relations is that democracies are less prone to war with each other because their state leaders are accountable to citizens (Oneal and Russett, 1997; Conconi et al., 2008). Accordingly, the history of wars does not accurately reflect the extent of dispute issues at stake between two countries. The question addressed in this paper is to understand what motivates the choice to settle disputes through negotiation rather than through war and how international institutions affect this choice. Distinguishing the presence of dispute issues between two countries from their most severe outcome, i.e. war, is therefore particularly relevant to this study. In this paper, I use two different sources of data to measure the occurrence of interstate disputes and wars. The Correlates of War (COW) project provides detailed data on Militarized Interstate Disputes (MIDs) over the period 1816 2001 (Faten et al., 2004). War is restrictively defined as a MID involving at least 1000 deaths of military personnel, which reduces the number of events considered as war to less than 100 cases since 1815. Adopting such a restrictive definition prevents robust empirical analysis. I follow the literature and use a broader definition of war that includes armed conflicts involving the display or the use of military force, i.e. a MID of hostility level 3 (display of force), 4 (use of force) or 5 (war) in the COW database. 11 Robustness analysis is conducted on a stricter definition of war, i.e. MID of hostility levels 4 and 5. Such qualitative data on military conflicts require that the actors, duration, geographical location and intensity of each conflict have been coded by researchers. Such a process prevents similar exercises on conflicts of lower intensity. In order to measure dispute occurrence, I use an alternative type of data: event data. Event data are reported, by trained students or automatically by computers, on a day-by-day basis from newspapers or wire services and coded by actor, target, form of action and date. Data on daily events have the great advantage of providing information on interstate interactions whatever the intensity of the underlying event. Event data do not allow us to assess the evolution of a given conflict over time. They nevertheless make it possible to measure the occurrence of disputes that clear a minimum threshold between two countries in a given year and that are resolved through negotiation or using military force. I use the event data compiled by Kinsella and Russett (2002) to measure disputes exceeding a threshold defined as strong verbal hostility. 12 Kinsella and Russett (2002) overlap data from three event databases, the Conflict and Peace Data Bank (COPDAB), the World Event/Interaction Survey (WEIS) and the Protocol for the Assessment of Nonviolent Direct Action (PANDA), to construct a dummy variable coded 1 if a dispute occurs for any dyad-year over the 1950 1991 period. 13 Table 1 provides event categories coded as disputes and their level of severity according to the Goldstein (1992) scale. Only events classified at least as conflictual as categories Cancel or postpone planned events and Charge; criticize; blame; disapprove are coded as a dispute. Table 2 shows that the proportion of MIDs and RTA members remains similar when the sample is restricted due to the availability of event and trade data. MIDs are nevertheless slightly biased towards less severe MIDs. Out of the 128 368 dyad-years of our restricted sample, 7937 experience a dispute, of which 586 spillover into MIDs. Appendix A provides details on the source of other data. The dataset includes 103 RTAs, comprising 17 partial scope agreements, 70 free trade agreements, 14 customs unions and two common markets. 11 MID level 2 (threat to use force) is not deemed to be a military conflict. See the COW website (http://www.correlatesofwar.org/) for more information and records of MIDs. 12 See Kinsella and Russett (2002, pp.1054 1055) for more details on databases used and the operationalizing of the minimum conflict intensity threshold. Schrodt and Gerner (2000) present limitations related to the use of event data. Using events exceeding a certain intensity reduces the biases they identify. 13 One hundred and eighty nine cases exhibit a MID but no dispute in the restricted sample, of which 170 exhibit a dispute in t 1. I follow Kinsella and Russett (2002) and treat them as measurement errors, due to the fact that event databases rely on major news media and do not cover all regions of the world equally. The dummy variable is thus recoded as if a dispute occurred.

V. Vicard / European Economic Review 56 (2012) 54 71 57 Table 1 Events and Goldstein scale. Event category Request action; call for Explicit decline to comment Urge or suggest action or policy Comment on situation Deny an accusation Deny an attributed policy, action, role or position Grant asylum Make complaint (not formal) Cancel or postpone planned events Charge; criticize; blame; disapprove Issue formal complaint or protest Give warning Denounce; denigrate; abuse Halt negotiation Turn down proposal; reject protest, demand, threat Refuse; oppose; refuse to allow Reduce routine international activity; recall officials Detain or arrest person(s) Threat without specific negative sanction stated Issue order or command, insist, demand compliance Expel organization or group Order person or personnel out of country Nonmilitary demonstration, walk out on Reduce or cut off aid or assistance; act to punish/deprive Threat with specific negative nonmilitary sanction Ultimatum; threat with negative sanction and time limit Threat with force specified Break diplomatic relations Armed force mobilization, exercise, display; military buildup Noninjury destructive action Nonmilitary destruction/injury Seize position or possessions Military attack; clash; assault Goldstein 0.1 0.1 0.1 0.2 0.9 1.1 1.1 1.9 2.2 2.2 2.4 3 3.4 3.8 4 4 4.1 4.4 4.4 4.9 4.9 5 5.2 5.6 5.8 6.9 7 7 7.6 8.3 8.7 9.2 10 Source: Goldstein (1992). Table 2 Descriptive statistics 1950 1991. Sample Full Restricted a Observations 366 546 128 368 MIDs of which: 1969 586 Display of force (3) 14% 20% Use of force (4) 66% 75% War (5) 20% 5% Dispute 7937 Mean Deep RTAs 0.007 0.008 Shallow RTAs 0.028 0.042 a Sample conditioning on the explanatory variables in column 4 of Table 3. 3. The impact of regionalism on the likelihood of war 3.1. Empirical strategy The aim of this section is to test the proposition that RTAs requiring significant political integration do reduce significantly the probability that disputes between members escalate into war, whereas other RTAs do not. As emphasized above, using a simple probit or logit model to estimate the conditional probability of war would severely bias results because heterogeneity in dispute occurrence creates a selection bias. The literature on the determinants of war have dealt with this issue either by limiting the sample to politically relevant dyads (i.e. pairs of countries sharing a common border

58 V. Vicard / European Economic Review 56 (2012) 54 71 Table 3 Impact of RTAs on war: bivariate censored probit model. Dependent variable (1) (2) (3) (4) MID MID MID Dispute MID Dispute Deep RTA membership dum. 0.21 0.48 b 0.67 a 0.26 a 0.59 b 0.11 (0.20) (0.24) (0.23) (0.09) (0.29) (0.10) Shallow RTA membership dum. 0.07 0.09 0.20 c 0.00 0.06 0.03 (0.11) (0.15) (0.11) (0.05) (0.12) (0.06) No. of peaceful years 0.02 a 0.03 a 0.01 a 0.00 a 0.01 a 0.00 a (0.00) (0.01) (0.00) (0.00) (0.00) (0.00) Log distance 0.23 a 0.10 0.04 0.28 a 0.08 0.32 a (0.04) (0.09) (0.06) (0.02) (0.11) (0.02) Contiguity dum. 0.78 a 0.66 a 0.47 a 0.58 a 0.60 a 0.29 a (0.10) (0.16) (0.15) (0.07) (0.17) (0.07) No. of landlocked countries 0.36 a 0.22 a (0.03) (0.03) Bil. trade dependence (t 4) 0.51 1.66 a (0.55) (0.23) Multil. trade dependence (t 4) 0.26 b 0.30 a (0.13) (0.05) Zero trade dum. (t 4) 0.05 0.21 a (0.19) (0.03) Common language dum. 0.28 a 0.16 a (0.10) (0.04) Colonial relationship dum. 0.17 0.55 a (0.24) (0.09) Common colonizer dum. 0.09 0.11 b (0.13) (0.06) Sum log area 0.00 0.13 a (0.05) (0.01) Sum of democracy indexes 0.38 a 0.23 a (0.09) (0.03) Common defense alliance dum. 0.27 0.48 a (0.21) (0.06) UN voting correlation (t 4) 0.25 0.99 a (0.37) (0.05) Observations 201 627 6503 201 627 128 368 Uncensored observations 12 047 7937 Log likelihood 4518.0 924.8 43 452.2 2358.3 Rho (Wald test of independent eqn.) 0.32 0.18 Estimation method Probit Probit Bivariate probit with censoring Sample Full disto1000 km Full Full Full Full Time dummies Yes Yes Yes Yes Yes Yes Note: Robust standard errors in parentheses. Intercept and time dummies not reported. Standard errors clustered by country pair. a Significance at the 1% level. b Significance at the 5% level. c Significance at the 10% level. or involving a major power) or by focusing on one cross-sectional determinant of disputes, i.e. bilateral distance. I go one step further and use event data to directly measure disputes (Kinsella and Russett, 2002). Determinants of dispute occurrence are indeed numerous. Focusing on some as a proxy for dispute probability ignores much heterogeneity in disputes between country pairs and over time. Indeed, the restricted sample for which data on disputes and other explanatory variables are available (column 4 of Table 3) contains 584 MIDs at the country pair level, of which 404 (69%) occur between politically relevant dyads and 199 (34%) between dyads separated by less than 1000 km. 14 Since the escalation into war process can only be observed when a dispute has occurred, using a bivariate probit with censoring is a natural empirical strategy to estimate the conditional probability of war for each dyad-year. The log-likelihood function is based on the unconditional probabilities associated with the three possible outcomes (Greene, 2003, p. 713): no dispute (dispute¼0), a dispute emerges but does not escalate into war (dispute¼1 and war¼0), and the dispute escalates into war (dispute¼1 and war¼1). Two equations are jointly estimated, one explaining the 14 The corresponding figures for the full sample are 1969 MIDs, of which 1358 (69%) are between politically relevant dyads, and 595 (30%) are between countries separated by less than 1000 km.

V. Vicard / European Economic Review 56 (2012) 54 71 59 initiation of the dispute and the second the dispute s escalation into war. Consider y 1 and y 2, two latent (unobserved) variables, representing the difference in utility levels from the initiation of the dispute and the dispute s escalation into war respectively. The model estimated is derived from a standard bivariate probit model: ( y 1 ¼ b 1 X 1 þe 1 and dispute ¼ 1 if y 1 40 0 if y 1 r0, ( y 2 ¼ b 2 X 2 þe 2 and war ¼ 1 if y 2 40 0 if y 2 r0, where X 1,2 are vectors of explanatory variables, b 1,2 vectors of parameters, and error terms e 1 and e 2 are assumed to be independent from X 1,2 and to follow Eðe 1 Þ¼Eðe 2 Þ¼0, Varðe 1 Þ¼Varðe 2 Þ¼1, and Cov½e 1,e 2 Š¼R. Wooldridge (2002, p. 564) emphasizes that, technically, the coefficients can be identified due only to the nonlinearity of the two equations in the bivariate probit. Hence, it is not necessary for X 2 to be a strict subset of X 1 for the outcome equation to be identified. However, the identification of the model s parameters is better achieved when X 1 contains at least one variable that is not in X 2, so that we have an exclusion restriction, i.e. a variable that influences the selection equation but not the outcome equation. The number of landlocked countries in a dyad is a good candidate as an identification variable. Having no access to the ocean indeed reduces a country s geographical scope of interest and reduces the opportunity for contacts with countries other than neighboring countries. Landlocked countries should therefore be less likely to interact with distant countries, through cooperative as well as conflictual actions, and should have less interstate disputes on average. There is however no reason that being landlocked affects the way conflicts are settled. 15 All specifications control for autocorrelation by clustering standard errors by country pair. ð2þ 3.2. Results The results are presented in Table 3. All specifications include the basic determinants of war put forward by the literature in political science: contiguity, distance and the number of peaceful years between the two countries (Beck et al., 1998). I first present the results from a simple probit estimator. Column 1 shows that RTA membership has no significant effect on war probabilities on the full sample of dyad-years. When only countries separated by less than 1000 km are considered, belonging to a deep RTA, but not a shallow one, does significantly reduce war probabilities (specification (2)). These crude results emphasize the need to account for heterogeneity in dispute occurrence. Specification (3) reports estimates using the bivariate probit accounting for selection. The first and second columns present the results of the escalation into war and dispute initiation equations respectively. As expected, landlocked countries face a lower probability of dispute. The coefficient on the exclusion variable is significant at the 1% level in all specifications. The results confirm that membership of a deep RTA does significantly reduce the probability that a dispute escalates into war. In this specification, membership of a shallow RTA is found to have a lower, albeit significant, effect on escalation into war probabilities. It is however not significant when additional controls are included. Specification (4) in Table 3 includes several potential co-determinants of regionalism and war (see Appendix A for data source). First, I include year dummies to control for any global shock affecting war probabilities and regionalism over time. Other controls can be divided into two sets: trade and political variables. Regarding trade, we control for both bilateral (the log of the mean of bilateral imports as a percentage of GDP) and multilateral (the log of the mean of multilateral (excluding bilateral) imports as a percentage of GDP) trade openness. 16 Martin et al. (2008) indeed argue that trade has an ambiguous effect on peace. Bilateral trade does reduce war probabilities but multilateral trade openness, by reducing the dependence on a particular partner, increases war probabilities. The results confirm that multilateral trade openness increases the probability that a dispute escalates into war, but show no significant effect of bilateral trade. 17 A dummy variable for zero trade flows is included as well as dummies indicating countries sharing a common language, countries that have had a colonial relationship or a common colonizer. Countries sharing a colonial history are indeed more likely to have unresolved disputes. They also share historical, cultural and institutional traits that make them more likely to create a RTA. The set of political variables control for political regime, the size of countries, alliances and diplomatic relations. The level of democracy is included as a control since democracies have been found to be less likely to wage wars (see Levy and Razin (2004) and Jackson and Massimo (2007) for a theoretical treatment, and Oneal and Russett (1997) among others for empirical evidence). Democratic status has also been argued to affect the choice to create an RTA (Mansfield et al., 2002). The (sum of the log of the) area of the two countries is included since a bigger territory is exposed to more opponents. Larger countries also depend less on foreign trade, which could affect their incentives to create RTAs. Finally, I control for diplomatic affinity between countries using a dummy variable for common membership of a defense alliance and the correlation of voting in the UN General Assembly (lagged four years). 15 When introduced into a probit model of the second stage equation, the number of landlocked countries is not statistically significant. 16 Following Martin et al. (2008), trade variables are lagged 4 years to remove any contemporaneous reverse effect of war on trade. 17 More precisely, the peaceful effect of bilateral trade vanishes when other controls are added. Martin et al. (2008) find a similar pattern.

60 V. Vicard / European Economic Review 56 (2012) 54 71 Table 4 Impact of RTAs on war: robustness. Dependent variable (1) (2) (3) (4) (5) (6) (7) MID 4 & 5 MID MID MID MID MID MID Deep RTA membership dum. 0.52 c 1.19 b 0.48 c 0.51 0.51 c 0.57 b 1.04 a (0.29) (0.59) (0.28) (0.32) (0.27) (0.29) (0.28) Shallow RTA membership dum. 0.07 0.23 0.02 0.03 0.05 0.06 0.18 (0.11) (0.26) (0.12) (0.12) (0.11) (0.12) (0.39) No. of major powers 0.36 (0.24) One communist country dum. 0.46 b (0.22) Two communist countries dum. 1.45 a (0.40) GATT membership dum. 0.21 a (0.08) Sum log military 0.07 c expenditure (t 4) (0.04) Abs. diff. log military expenditure (t 4) 0.03 (0.02) First stage IV Shallow Deep Sum deep RTAs with third countries (t 5) 0.06 a 0.24 a (0.02) (0.03) Sum shallow RTAs with third countries (t 5) 0.18 a 0.21 a (0.01) (0.05) Observations 128 368 55 409 128 368 128 368 123 405 7937 7678 7678 Uncensored observations 7937 5013 7937 7937 7776 Log likelihood 23 245.2 12 998.7 22 659 23 447.7 21 642.1 1543.4 1462.6 Rho (Wald test of independent eqn.) 0.40 0.18 0.22 0.02 0.42 Estimation method Bivariate probit with censoring Probit IV Probit Sample Full OECD Full Full Full dispute¼1 dispute¼1 Time dummies Yes Yes Yes Yes Yes Yes Yes Note: Robust standard errors in parentheses. Intercept and time dummies not reported. All estimates include the same control variables as in our preferred specification (specification (4) in Table 3). In specifications (6) and (7), only the second step of the bivariate censored probit is reported. Standard errors clustered by country pair. a Significance at the 1% level. b Significance at the 5% level. c Significance at the 10% level. Controlling for all these potential co-determinants of war and regionalism, specification (4) confirms that deep RTAs promote the peaceful settlement of disputes between members. Membership of shallow RTAs has no effect on war probabilities. Moreover, the insignificant impact of bilateral trade on the likelihood of a dispute escalating into war confirms that the institutional channel dominates in regional integration. It is worth noting that when all these controls are added, the Wald test of independent equations is no longer significant, meaning that the two equations estimated simultaneously are independent. 3.3. Robustness Tables 4 and 5 present a number of robustness tests. 18 The first column tests the sensitivity of the results to a more restrictive definition of war including only MIDs of hostility levels 4 and 5, i.e. implying the actual use of military force and wars resulting in the death of at least 1000 military personnel. This more restrictive definition of war reduces slightly the significance of the coefficient on deep RTAs but the results remain qualitatively similar. I also restrict the sample to country pairs including at least one OECD country in 1991. Event data are indeed coded from international newspapers which do not cover all the world s countries evenly. Disputes involving richer countries may be more likely to be reported by international newspaper than disputes involving poor countries. On this restricted sample, deep RTAs but not shallow ones are still negatively and strongly significantly associated with the likelihood of disputes escalating into war. The coefficient on deep RTAs is however twice as large as in the benchmark estimation, suggesting measurement errors. Finally, columns 3 5 control for additional potential co-determinants of war and regionalism. Controlling for the number of major powers, communist regimes, membership of GATT and the level and difference of (log of) military expenditure 18 To save space, only the coefficients of the variables of interest and of the second stage equation are reported. All specifications include the control variables included in specification (4) of Table 3.

V. Vicard / European Economic Review 56 (2012) 54 71 61 Table 5 Impact of RTAs on war: robustness, continued. Dependent variable (8) (9) (10) (11) (12) (13) (14) MID MID MID MID MID MID MID Deep RTA membership dum. 0.54 c 0.53 c 0.14 a 0.12 a 0.14 a 0.10 a 0.58 b (0.31) (0.29) ( 0.03) (0.03) (0.04) (0.03) (0.29) Shallow RTA membership dum. 0.04 0.02 0.05 0.03 0.00 0.06 0.06 (0.13) (0.12) ( 0.04) (0.04) (0.04) (0.04) (0.12) One oil exporter dum. 0.17 c 0.18 b (0.09) (0.09) Two oil exporters dum. 0.05 0.03 (0.22) (0.21) Abs. diff in log GDP per capita (t 4) 0.03 (0.05) log sum GDP per capita (t 4) 0.11 c (0.06) Interaction deep RTA 0.87 a 0.35 0.76 0.43 (0.2) (0.53) (0.61) (0.92) Interaction shallow RTA 0.57 c 0.34 0.05 0.55 (0.3) (0.29) (0.29) (0.42) Abs. diff in share agriculture in value added 0.06 (0.06) Avg share of agriculture in value added 0.19 (0.12) Abs. diff in share industry in value added 0.04 (0.06) Avg share of industry in value added 0.19 b (0.09) Abs. diff in share manufacturing in value added 0.17 b (0.07) Avg share of manufacturing in value added 0.08 (0.12) Abs. Diff in of share services in value added 0.14 (0.15) Avg share of services in value added 0.13 (0.26) Lagged deep RTA membership dum. 0.56 (0.50) Observations 128 368 128 368 4204 4190 4190 1867 128 368 Uncensored observations 7937 7937 7937 Log likelihood 23 456.3 23 318.9 23457.7 Rho (Wald test of independent eqn.) 0.01 0.15 0.19 Estimation method Heckprob LPM Heckprob Sample Full Full dispute¼1 Full Time dummies Yes Yes Yes Yes Yes Yes Yes Note: Robust standard errors in parentheses. Intercept and time dummies not reported. All estimates include the same control variables as in our preferred specification (specification (4) in Table 3). Standard errors clustered by country pair. a Significance at the 1% level. b Significance at the 5% level. c Significance at the 10% level. does not affect the results. Nonetheless, omitted (unobservable) variables could bias the results. The next section implements an instrumental variable (IV) strategy to deal with this issue. 3.3.1. Instrumental variables The domino theory of regionalism suggests that the creation or enlargement of an RTA increases incentives of nonmembers to apply for membership (Baldwin, 1997). Using spatial econometrics, Egger and Larch (2008) show that the probability that two countries create an RTA increases with the creation of RTAs by third countries. The number of RTAs signed by each country with third countries could therefore qualify as a strong instrument correlated with the existence of an RTA between two countries in a given year. Since we control for bilateral and multilateral trade dependence in our regression, there is no reason to believe that the number of RTAs signed with third countries is correlated with war probabilities. The number of deep and shallow RTAs signed with third countries in t 5 by the two countries are used separately as instrumental variables for RTA membership. The enlargement of an existing RTA to include a new country indeed extends the same form of RTA; a country that is already a member of a deep RTA would be more likely to create a deep RTA with a new country and less likely to create a shallow RTA.

62 V. Vicard / European Economic Review 56 (2012) 54 71 Since our endogenous variables are dummy variables, using the traditional two stage methodology would yield inconsistent estimates unless the first-stage model is exactly right. The fitted value from a probit model may nevertheless be used as an instrument, together with other exogenous covariates, to generate first-stage estimates using an ordinary least square (OLS) model (Angrist and Krueger, 2001). We thus apply a three stage methodology. In the first stage, fitted values for deep and shallow RTA membership are generated using a probit model with our two instruments. The fitted values are used, together with other exogenous covariates, to predict the two endogenous RTA membership dummies using OLS. These predictions are then used in the estimation of war probabilities. In the third stage, we estimate directly the probability of war between countries involved in a dispute, i.e. the second stage of the bivariate probit with censoring. The Wald test of independent equations reported in Table 3 indeed shows that the second equation, i.e. dispute escalating into war, can be estimated independently when we control for enough covariates. For the sake of comparison, specification (6) in Table 4 presents the estimate of the dispute escalating into war equation using a probit model without IV. Results on deep and shallow RTA membership are similar to specification (4) in Table 3. The coefficient on the IVs in the first stage equation are reported at the bottom of column (7) in Table 4. Both are significant at the 1% level, confirming that they are strong instruments. As expected from the domino theory of regionalism, countries that have already signed RTAs with third countries are more likely to conclude similar agreements with new partners. The results of the final stage of the IV empirical strategy confirm that members of deep - but not shallow - RTAs are less likely to escalate a dispute into war. The coefficient is strongly significant and is almost twice as large as in the benchmark regression. It suggests that unobservable variables that make country pairs more likely to belong to the same deep RTA simultaneously increase the probability that a dispute escalates into war. 3.3.2. Asymmetric members Schiff and Winter (2003) argue that a deep RTA may have negative security effects when countries are very different, because tariff preferences under a common external tariff may lead to large income transfers and a concentration of industry in a given country/region. Asymmetric gains from regional integration may therefore increase the risk of conflicts between unbalanced members. Table 5 deals with this issue. First, columns 8 and 9 include additional control variables for differences in export specialization: dummy variables indicating whether one or two countries in the pair are oil exporters, as defined by the IMF in the DoTS, and the difference and average level of GDP per capita. The results show that poor countries and country pairs involving one oil exporter are more likely to escalate disputes into war, but the deep RTA variable remains significant. Differences in GDP per capita are found to have no impact on war probabilities. A direct test of Schiff and Winter s argument requires to account for the interaction between asymmetries between countries and RTA membership. The non-linearity of our estimator however makes the interpretation of coefficients on interaction terms potentially misleading (Ai and Norton, 2003). I therefore estimate a linear probability model (LPM) on the dispute escalating into war stage. This alternative methodology yields the marginal effects on the interaction terms (Angrist and Pischke, 2009). Schiff and Winter s argument involves a positive coefficient on the interaction between countries asymmetries and deep RTA membership. Asymmetries between countries structure of production are measured by the difference in the share of agriculture, industry, manufacturing and services in value added. Columns 10 13 of Table 5 present the results using, in turn, each measure and its interaction with deep and shallow RTA membership dummies. The average share of the sector considered (agriculture, industry, manufacturing and services) in value added is also controlled for. We find weak support for the argument of Schiff and Winter (2003). As expected, the coefficient on the interaction term with the deep RTA variable is positive, but significant only when using the difference in the share of agriculture. Note that the impact of deep RTA membership is negative for the median level of differences in production structure. Finally, column 14 shows that deep RTAs that fall apart do not increase conflicts in the following 5 years. 3.4. Quantification This section aims to quantify the impact of RTA membership on war probabilities. Since the estimator used is nonlinear, coefficients cannot be interpreted immediately. Fig. 1 presents marginal effects (or discrete change for dummy variables) on the conditional (on dispute occurrence) predicted probability @Prðwar ¼ 19dispute ¼ 1Þ=@X, computed from specification (4) in Table 3. The marginal effects are computed at two different levels: at the mean value of all variables and for contiguous countries separated by 1000 km and exhibiting positive trade flows (all other variables being held at their mean). For contiguous countries separated by 1000 km, being members of a customs union or common market reduces the probability that a dispute escalates into war by 9.1 percentage points. This effect is sizeable since the mean predicted conditional probability is 13.8%. The impact of belonging to a deep RTA is comparatively similar when all variables are held at their mean: it reduces the probability of a dispute escalating into war by 1.9 percentage points, while the predicted conditional probability is 2.5%. In comparison with other determinants of war, this effect is also sizeable. Belonging to a deep RTA is equivalent to increasing the number of peaceful years between two countries by 28 years from its mean value. Moreover, the peaceful effect of sharing a common language or a common defence alliance is two times smaller. Finally, when compared to the democratic peace channel, the impact of deep RTA membership is shown to be of the same order of magnitude as an increase from the mean to the top level of democracy indexes for the two countries.

V. Vicard / European Economic Review 56 (2012) 54 71 63 0.16 0.12 Predicted probability Deep RTA membership dum. 10 additional peaceful year Common language Sum democracy indexes (1 std.dev.) Common defence alliance dum. 0.08 0.04 0 Contiguous and trading countries, separated by 1000km Mean country pair Fig. 1. Impact of RTAs on war: quantification. 4. Determinants of deep and shallow regional trade agreements Having established that deep and shallow RTAs have differential effects on the probability that a dispute escalates into war, this section tests whether security issues are a motive for creating RTAs. 4.1. Empirical specification In line with Baier and Bergstrand (2004), I estimate the following specification on a cross-section of country pairs in 2005: PrðRTA h ij ¼ 1Þ¼b 0 þb 1 DP ij þb 2 OPEN ij þb 2 Controls ij þe ij, h ¼fdeep,shallowg ð3þ where Controls ij includes economic and political determinants of regionalism, DP ij is the propensity to dispute and OPEN ij is a proxy for multilateral trade openness. Baier and Bergstrand (2004) show that, in a model of RTA formation in a secure world (i.e. without interstate conflict), two countries facing low trade costs with the rest of the world should have lower welfare gains from joining an RTA compared to country pairs facing high trade costs. Martin et al. (2008) however show that countries more open to multilateral trade are more likely to escalate disputes into war. In an insecure world, a country facing low trade costs should therefore need more mechanisms to prevent dispute from escalating into war and secure gains from trade with their partners. They are more likely to create deep rather than shallow RTAs. Trade costs with the rest of the world should unambiguously decrease the likelihood of shallow RTAs but are likely to have an ambiguous impact on the likelihood of deep RTAs. Baier and Bergstrand (2004) measure the remoteness of countries by their mean distance to their trading partners. Anderson and van Wincoop(2003)nevertheless show that remoteness indexes are a bad proxy for trade openness and propose a strategy for estimating structurally what they call multilateral resistance indexes from a gravity model of trade. The methodology used to compute the multilateral resistance indexes is presented in Appendix B. Other economic determinants of RTAs are those put forward by Baier and Bergstrand (2004). Welfare gains from creating an RTA should increase with the size and similarity of GDPs of the two countries. Pairs of countries with dissimilar relative factor endowment and whose relative factor endowments are similar to the rest of the world should also receive higher welfare gains from creating an RTA. Relative factor endowment is measured by GDP per capita. The set of political controls includes a number of potential co-determinants of disputes and regionalism. First, the fact of sharing a common border, a common language or a common colonial history is likely to affect the probability of bilateral disputes as well as the incentives to create a RTA. Landlocked countries experience fewer disputes on average and face, at the same time, higher trade costs. The level of democracy of the two countries is also included as a control. Finally, diplomatic affinity between states is likely to influence decisions to create RTAs as well as the set of disputes between countries. I use two variables to control for diplomatic affinity: a dummy variable indicating countries sharing a common defense alliance and the correlation of voting in the UN General Assembly (see Appendix A for details on data). Finally, dispute probability is measured using the event data presented above, as the propensity to bilateral disputes between 1950 and 1969. I use lagged dispute data because past membership of RTAs is likely to affect variables in 2005.

64 V. Vicard / European Economic Review 56 (2012) 54 71 An important issue when estimating Eq. (3) is the endogeneity related to past membership of RTAs (Baier and Bergstrand, 2004). Lagged data may suffer from an endogeneity bias as well since some regional integration processes began shortly after World War II (the Benelux, the EU,y) and several current RTAs have been preceded by earlier attempts at integration in the 1950s and 1960s. 19 Moreover, endogeneity may also arise because of omitted variables affecting the likelihood of RTAs and disputes. To deal with endogeneity, I implement an IV strategy. 20 Since the dispute propensity variable is not a continuous variable, there is no straightforward nonlinear IV model available. As suggested by Angrist and Krueger (2001) and Wooldridge (2002), I use a three-stage least square (3SLS) model to estimate Eq. (3) for deep and shallow RTAs simultaneously taking dispute propensity as endogenous. The two exogenous instruments for dispute propensity are derived from international relations theory; they are the number of major powers in the country pair and the difference in national military expenditure. Major powers are countries that have military, diplomatic, economic and cultural power that allow them to exert a global influence (Waltz, 1979). Unlike other countries whose geographical scope of interest is regional or local, major powers have a global geographical scope of interest. They are thus likely to experience more interstate disputes with any country around the globe. Major power status is however not likely to be correlated with incentives to create RTAs since it is not defined solely by economic power Japan and Germany are not considered to be major powers, even if they are major economic powers. Moreover, controls for economic size are included in the main specification, which should remove the economic component of major power status. The second instrument is the difference in (log) military expenditure. Countries that have unbalanced military capabilities are indeed more likely to be involved in disputes (Kinsella and Russett, 2002). While the level of military expenditure is likely to be affected by decisions on RTAs, the difference in military expenditure is not, since both countries should decrease their military spending proportionally in the face of decreasing military threats. Since no exogenous instruments for multilateral resistance terms and GDP are available, these variables are lagged in 1960. This strategy reduces the sample of countries, because several countries were not independent in 1960. We are left with a sample of 2814 unique country pairs. 4.2. Results Table 6 presents estimates of Eq. (3). I start with a simple univariate regression. 21 Dispute propensity is positively and significantly associated with the likelihood of deep RTAs, but has no significant effect on shallow RTAs. These crude results may suffer from an endogeneity bias. Specification (2) of Table 6 controls for endogeneity using an IV model. Regarding the first stage equation, the number of major powers and the difference in military expenditure strongly and significantly increase dispute propensity. The partial R 2 of IVs is 0.29 and the F-statistic of weak identification on IVs greatly exceeds the threshold of ten recommended by Staiger and Stock (1997), confirming the relevance of the two instruments. In this specification, the Hansen Sargan overidentification test however rejects the exogeneity of the instruments. This simple univariate specification does not control for the economic size of countries, which should affect welfare gains from joining a RTA. Since economic power is also one of the attributes of major powers, the number of major powers is not exogenous when economic size is not controlled for. In specification (3), when we control for economic and political determinants of regionalism, the Hansen Sargan overidentification test does not reject the exogeneity of IVs, thus confirming their validity. As expected, the partial R 2 and F-statistic on IVs are greatly reduced, but still confirm that the instruments can be regarded as strong. Controlling for endogeneity confirms that dispute propensity has a differential impact on incentives to create deep and shallow RTAs. The coefficients on dispute propensity are significant at the 1% level. Country pairs that have many interstate disputes are more likely to create a deep RTA and less likely to create a shallow one. The relevance of security motives for decisions on trade policy is further illustrated by the differential impact of trade openness on the likelihood of deep and shallow RTAs. Pairs of countries naturally more open to trade (having larger multilateral resistance indexes) tend to create deep RTAs. Countries naturally more open to trade depend less on one partner in particular, which increases the probability that disputes spillover into war (Martin et al., 2008) and increases accordingly the need for international institutions supporting the peaceful resolution of disputes. By contrast, shallow RTAs are driven by the economic motives put forward by Baier and Bergstrand (2004) and are created between country pairs facing large trade costs with the rest of the world. Geographically close countries and countries sharing a common border are more likely to join all kinds of RTAs. Other economic determinants of regionalism are in line with Baier amd Bergstrand s (2004) results concerning shallow RTAs. Country pairs with similar endowments and with large and similar markets tend to create shallow RTAs. As in Egger and Larch (2008), countries that have endowments similar to the rest of the world are more likely to create shallow RTAs. Economic determinants are however less relevant to the creation of deep RTAs. Differences in endowments have no significant impact on the creation of 19 For instance, the East African Community was disbanded in 1977 and relaunched in 2000 as a partial scope agreement and the Central American Common Market collapsed in 1975 and relaunched as a customs union in 1993. 20 Another advantage of using an IV model is that it also deals with measurement error in the endogenous explanatory variable, which is, as explained above, also valuable in our case. For instance, institutions in deep RTAs are likely to publicize disputes, creating a downward bias on the coefficient of dispute propensity in the case of deep RTAs. 21 Using a bivariate probit estimator yields similar results.