Political Stability and Trade Agreements

Transcription

1 Political Stability and Trade Agreements Lorenzo Rotunno This draft: May, 2012 VERY PRELIMINARY AND INCOMPLETE PLEASE DO NOT CITE OR CIRCULATE Abstract This paper examines how the (in)stability of an incumbent government affects its attitudes towards trade agreements. Previous theoretical work has emphasized the possibility that governments, facing a high probability of losing office, may want to sign Free Trade Agreements (FTAs) strategically to reduce distortions from future protectionist policies. I test empirically this proposition on a panel of country-pairs during the period The average tenure of previous heads of government and the share of seats in parliament held by the incumbent are used to identify and predict the probability of government turnover. The estimates suggest that higher expectations of replacement of the ruling party increase the likelihood of forming an FTA. The effect is statistically and economically significant and robust to alternative specifications. Party turnover, causing important shifts in political power, is found to be an important determinant of FTAs. Keywords: Trade agreements, Political Rents, Incumbent Stability. JEL Classification: F13, F53. The Graduate Institute of International and Development Studies (HEID), Geneva. lorenzo.rotunno@graduateinstitute.ch. 1

2 1 Introduction Political influence from special interest groups (SIGs) is a major reason why governments adopt welfare-reducing, protectionist policies. Yet, countries around the world have progressively liberalized trade policies and have done so increasingly through Free Trade Agreements (FTAs). Several positive theories have been proposed to explain why and under which conditions countries dismantle trade barriers that were politically optimal. A partly overlooked motivation is the possible strategic role of FTAs. Incumbent governments may want to sign trade agreements to constraint the trade policy space of their successors, analogously to the use they may want to make of other binding policies (e.g., fiscal policies; see Alesina and Tabellini 1990). This paper tests empirically the existence of these strategic incentives to sign trade agreements. More specifically, I examine whether the likelihood of political turnover, or, alternatively, incumbent stability, is a determinant of trade agreements. The work of Ornelas (2005, 2007) describes the theoretical mechanism that justifies this analysis. In a fairly general political economy model of trade policy, feasible Free Trade Agreements (FTAs) reduce available rents in the policy-making process and increase welfare. Incumbents, perceiving a high risk of being replaced, might thus adopt FTAs to reduce the possibility that their successors will raise distorting tariffs. In an extension of his 2005 piece, Ornelas (2007) derives formally this prediction by introducing the probability that the incumbent loses power in the next period. Alternative arguments may apparently suggest the opposite, namely that incumbents with short time horizons may be less likely to reduce distorting trade barriers. For instance, highly unstable governments, knowing that they will be replaced soon, may want to seize the opportunity and extract as much office-related rents as they can, for instance engaging in outright corruption (Campante et al., 2009). These governments may find it damaging to liberalize trade policy, which can be an important source of political rents. However, realizing that they will soon be out of office without any policy-related rents, the same governments may start caring more about the social welfare of their constituencies and thus turn in favor of (welfare-enhancing) FTAs. In this paper, I show that indeed more stable governments are less likely to sign FTAs. In particular, countries where the ruling party faces a high probability of losing power are significantly more likely to enter trade agreements. The yearly probability of government turnover is predicted exploiting plausibly exogenous variation in the share of seats in the parliament held by the incumbent and the average tenure of previous governments. Both variables are strong predictors of changes of the largest government party and of changes of the government chief executive (i.e., head of government in parliamentary systems and head of state in presidential systems). Evidence on the positive effect of the probability of leadership turnover on FTA formation is rather weak. This indirectly corroborates the theoretical mechanism insomuch as changes of the chief executive do not entail the kind of shift in political power that may trigger strategic considerations and that is more likely to arise with turnover of the ruling party. To preview the empirical 2

3 results, a 10% increase in the probability of party turnover increases the odds of adopting an FTA 15%. The findings are robust to two alternative empirical setups, one that relates the country-specific probability of political turnover to the country-pair event of an FTA and the other that combines the predicted probability of change of government at the country-pair level. Furthermore, different specifications and estimators broadly support the findings. In particular, while the level of democracy inevitably affects political competition and hence the incumbent s time horizon, I find that expectations of political turnover has a positive effect on the formation of trade agreements that is independent of the (significant) effect of democracy. The remainder of the paper is organized as follows. Section 2 briefly summarizes the relevant literature, focusing on research dealing with the interactions between trade policies and features of the political system. Section 3 reviews the theoretical framework of Ornelas (2005, 2007) to set the stage for the empirical analysis. Section 4 describes the empirical strategy and the data, while Section 5 discusses the empirical evidence. The last Section concludes. 2 Related Literature This paper contributes to the large empirical literature on the determinants of trade agreements. It identifies a new non-economic factor, the likelihood of political turnover, that affects significantly the probability of FTA formation. The objective and the empirical analysis of this paper is closely related to the work of Ornelas and Liu (2011), where a political economy model of FTA formation in the spirit of Ornelas (2007) is used to show that governments in unstable democracies might want to sign FTAs to avoid excessive rent-seeking behavior by likely future autocrats. My analysis qualifies their results: given the rent destruction effect of trade agreements, incumbent governments fearing to be replaced would sign an FTA to reduce rents available to their successors, in both democratic and autocratic regimes. A recent paper by Hollyer and Rosendorff (2011) shows that political leaders who have signed more trade agreements stay in office longer. The argument is that trade agreements, by reducing policy uncertainty, creates welfare gains that are reflected in higher support for the leader. My empirical analysis is meant to test a different prediction, namely that incumbent stability affects negatively the odds of FTA formation. The empirical specification employed in the analysis (see Section 4) minimizes the possibility of reverse causality bias by focusing on FTA onset and thus mimicking a survival analysis setup. Moreover, if the mechanism proposed by Hollyer and Rosendorff (2011) is still at work, it should go against the prediction that more stable governments are less likely to sign FTAs and hence weaken the effect towards zero. Previous research in both international economics and political science has investigated the relation between regime type and FTAs. The basic argument is that political leaders in democratic systems have less room for rent-extracting policies because of more effective 3

4 and widespread public scrutiny. Mansfield et al. (2002) finds empirical support for this prediction and shows that democratic country-pairs are more likely to sign a trade agreement than autocratic or mixed pairs. In another study, Mansfield and Milner (2010) finds that the number of institutional actors whose preferences are not aligned with those of the chief executive (veto points) affects negatively the likelihood that the government signs and ratifies a trade agreement. However, none of these papers investigate explicitly the role of political competition on the incentives the incumbent government has to sign an FTA 1. My empirical evidence speaks to this strand of the literature by showing that the likelihood of political turnover is an important channel, though not the only nor the most important one, whereby democratic regimes are more inclined to sign FTAs than autocracies. Another strand of the literature has looked at the use of trade agreements as commitment devices to solve time-inconsistency problems stemming from misallocations of resources (Maggi and Rodriguez-Clare, 1998, 2007). When the sector-specific factor is mobile only in the long-run, protectionist policies lock-in resources in inefficient sectors and attract political contributions from those sectors. If the government has limited capacity to extract rents from the policy-making process, it might prefer to commit to free trade and remove the distortions. Arcand et al. (2011) confirms empirically that credibility is a significant driver or trade agreements and that those agreements signed because of credibility motivations are welfare-enhancing. This paper tests another type of time-inconsistency, due to the fact that the government might not benefit from political rents in the next period. More generally, other recent papers have explored political economy motivations behind the adoption of trade agreements. Among those, Baldwin and Jaimovich (2010) find supportive evidence of the domino theory of FTAs, whereby countries join existing trade agreements or form new ones to offset detrimental trade diversion effects. Finally, Martin et al. (2010) find that country pairs with a high probability of conflicts are more likely to sign FTAs to avoid the trade disruption costs of wars. 3 Theoretical Framework In this Section, I briefly review the model of Ornelas (2005, 2007) and discuss the empirical prediction that is brought to the data in the next Section. Consider a world of three (large) competitive economies, H, F and Z, producing and consuming N goods (and sectors). Consumers preferences are quasi-linear and give rise to the linear demand function D g = A p g for each non-numeraire good g = 1,..., N 1. The numeraire good is freely traded and produced one-to-one with labour. An inelastic supply of labour in all countries pins down the wage rate to one. Country H is assumed to be the natural importer of a subset of goods m = 1,..., M, F is the natural importer of another subset E and and Z would import the remaining 1 In the democracy literature, a political system is democratic if competitive elections are fair and have regular occurrence (see Mansfield et al., 2002). Democratic countries have higher political competition than autocratic ones, all other factors being equal. I will control for this confounding factor in the empirical analysis. 4

5 goods under free trade. The subset x = 1,..., X includes the goods that are exported by H. Optimal supply in each sector is assumed to be linear and equal to y x = d x p x for each export good and to y m = d m p m for the generic import good (with d x > d m > 0). Non-numeraire goods are produced by labour and a sector-specific input, call it capital, under constant returns to scale. Factor owners receive therefore rents π g π g [p g ] from selling the non-numeraire good. Governments can impose specific import tariffs. Given the symmetry of the model, the analysis is done from the perspective of country H (H stands for Home ). Country F is assumed to be the potential FTA partner, while country Z is the third country that would be excluded from the agreement. Home imposes tariff τ i on imports coming from F and τ e on imports coming from Z. Under a trade agreement between H and F, τ i is constrained to 0, while without trade agreement the MFN principle ensues and τ i = τ e τ. Owners of the factors specific to the import goods can organize and lobby the incumbent for protection. They are assumed to represent a negligible fraction of the population so that their payoff is given by V π m [p m ] C, without any role for consumer surplus and rebated tariff revenues 2. As in the standard Protection for Sale (Grossman and Helpman, 1994) framework, the incumbent government cares about social welfare and contributions: G 1 b ( W M + W X) + C where W M is the sum of consumers surplus, tariff revenue and producers surplus across all import sectors, W X is the sum of consumers surplus and producers surplus in the export sectors, and b is the weight the incumbent gives to political contributions relative to social welfare. Alternatively, 1 /b can be seen as the importance that politicians attach to the welfare of their constituencies (Ornelas, 2007). At the beginning of each period, the incumbent and the domestic lobby Nash bargain over the level of import tariffs and contributions. Let α and 1 α denote the bargaining powers of the incumbent and the domestic lobby, respectively. The political tariff maximizes the joint surplus of the government and the lobby in the generic import sector (see also Maggi and Rodriguez-Clare 1998): { W τ p m } [τ] = arg max + π m [τ] b where I made explicit the fact that welfare and industry rents in each import sector are a function of the tariff τ. If the government has no bargaining power, i.e. α = 0, contributions are such that it receives its reservation payoff, W m [τ ]/b, where τ, is the optimal (non-political) tariff 3. If the incumbent has all the bargaining power, i.e. α = 1, it extracts all the rents and the domestic industry is left indifferent between lobbying or not. Therefore, the equilibrium contributions in each import sector are: Ĉ = α (π m [τ p ] π m [τ ]) + (1 α) (W m [τ ] W m [τ p ]) b 2 See Baldwin and Robert-Nicoud (2006) for a discussion of the implications of this assumption. 3 Because of the large country assumption, τ > 0. 5

6 Following Ornelas (2005), political rents are defined as the difference between the joint government-lobby surplus under political pressure and the same joint surplus without political pressure: P R m 1 b ((W m [τ p ] + bπ m [τ p ]) (W m [τ ] + π m [τ ])) Using this definition, the government total payoff equals: G = 1 b ( W M [τ ] + W X) + αp R M [τ, τ p ] (1) 3.1 Free Trade Agreements and Political Turnover I now introduce the possibility of an FTA between country H and country F that eliminates completely all trade barriers. The exact timing of the FTA is immaterial to the analysis, as long as the MFN political tariff τ p is applied before the agreement is signed. The potential agreement comes as a surprise to the domestic industries and to the export industries in particular. Owners of the factors specific to export sectors are potential lobbysts in favor of the trade agreement, since this would warrant them duty-free access to the market of country F. For ease of exposition, I assume that the opportunity of an FTA comes exogenously or, alternatively, that the incumbent is insulated from political pressure at the international negotiations table. Ornelas (2005) shows that the main rent destruction effect of the agreement is preserved if this assumption is relaxed. Once in force, both parties honour the agreement with no possibility of defection. If, instead, the incumbent or the successor could freely renege on the agreement s obligations, import-competing sectors could lobby the incumbent to undo the agreement, creating new sources of post-fta rents that would undermine the main prediction of the ensuing analysis. The assumption of exogenously enforced FTAs can be seen as an endogenous outcome in the presence of (high) cost of negotiations and withdrawal from the agreement (Ornelas, 2007) 4. Moreover, the assumption is empirically plausible given the very few cases of countries that have withdrawn from FTAs or Customs Unions (CUs) 5. The incumbent of country H signs the agreement with country F if it increases its payoff, that is, if: G > 0 1 b ( W M [τ ] + W X ) + αp R M [τ, τ p ] > 0 (2) where the subscript indicates that variables are in differences between the FTA and the no-fta scenario (e.g., G = G F T A G). The first two terms in condition 2 gives the welfare effects of the trade agreement should politics be absent. The sign of these effects is generally ambiguous. While deviations from the unilaterally optimal trade policy 4 In a setting where trade agreements are used by governments as commitment devices against domestic lobbies, Arcand et al. (2011) show that withdrawal costs increases with the market size of the trading partner. 5 These cases have all occurred in Latin America. The Central American Common Market (CACM), formed in the 60s, suspended its activities in the mid-1980s due to mounting violence and protectionist pressures in member countries. The countries formed a new FTA about ten years later. More recently, Venezuela withdrew from the Andean Pact because of diplomatic tensions with some of its partners. 6

7 decreases welfare in the import sector, it creates terms-of-trade gains for the export sector in the market of the partner country 6. One of the key results of Ornelas (2005) is that the FTA unambiguously reduces political rents. The import-sector has less incentives to lobby under the agreement since now any marginal increase of the external tariff creates additional rents also to the exporters from the partner country. This framework is extended so as to allow two periods and two parties (loosely defined as the incumbent and the opposition parties). The incumbent government receives a payoff given by equation 1 in each period if it stays in office. If out of office, politicians don t receive any rents and their payoff coincides with the one of the average voter, weighted by 1 /b 7. Ornelas (2007) discusses at length how this formulation can be rationalized by a situation where politicians are able to extract rents primarily through enactment of policies rather than campaign financing. In this simple formalization, I assume that parties attach the same weight to social welfare and have the same bargaining power vis-à-vis domestic lobbies. In an endogenous characterization of the probability of government turnover, Ornelas (2007) assumes a oneto-one mapping between constituencies size and the parameter b, which leads to different b s for the incumbent and opposition parties. However, I decide to discard this distinction since predictions about the role of parameters b and α are not used in the empirical analysis. The incumbent faces an exogenously given probability of losing office between period 1 and period 2, denoted by σ [0, 1]. The government signs the agreement if it increases its present value payoff: Γ > 0 with Γ = G + δ ((1 σ) G + σo ) (3) where δ [0, 1] is the politicians discount factor and O 1 ( b W M [τ p ] + W X ) is the change in the politicians payoff due to the FTA if they are out of power. Differentiating this equation with respect to σ, it follows that: dγ dσ = δ b ( W M [τ p ] W M [τ ] bαp R M ) (4) The last term in the round brackets, P R, is negative as discussed above. The difference between the first two terms denotes how the FTA affects welfare differently with and without politics. The trade agreement destructs political rents creating an extra source of welfare gains. These extra gains are absent in an initial situation where politics had no role, that is, when the import tariff is the optimal τ. Consequently, W M [τ p ] > W M [τ ] and equation 4 is unambiguously positive. A country is more likely to sign an FTA as the probability that the incumbent government loses office increases. Facing the probability of not being able to extract political rents in the next period, may want to sign an FTA so as to limit welfare distortions. A sufficiently high expectation of government turnover can thus make politically viable an agreement that wouldn t be viable with a more stable 6 The number of sectors is a key determinant of the net welfare effects of FTAs in this simple setting. It is thus more likely that the trade agreement is welfare-enhancing with more export than import industries. 7 The size of the population is normalized to one. 7

8 government. Few points about this comparative statics prediction are worth emphasizing. First, the theoretical framework is rather illustrative of the possible mechanism whereby the time horizon of politicians can affect the likelihood of FTA formation. Other theoretical arguments may lead to a similar prediction. Aside from any political pressure from domestic groups, an ideologically pro-trade incumbent government might be more likely to sign an FTA when faced with high chances of losing office if the opponent has a more protectionist ideal policy 8. This line of reasoning hinges on the (different) political preferences of the incumbent and of the opposition, while the framework presented above does not require specific political preferences (the parameter b is assumed to be equal across parties). Furthermore, the prediction is confirmed for any value of the incumbent s bargaining power. In particular, the likelihood of political turnover increases the odds of FTA formation even if the incumbent is not able to extract any rents from the policy-making process (see equation 4 when the parameter α = 0). In this scenario, the incumbent would still like to tie the hands of its successor since, when out of office, there is no compensation for the welfare distortions caused by trade protection. Finally, the probability of government turnover has been assumed exogenous to the adoption of an FTA. Ornelas (2007) extends a version of the theoretical framework to a probabilistic voting model and finds that an FTA enhances the incumbent s electoral prospects. The finding relies nevertheless on a particular relationship between the political preference parameter b and constituency size. In other settings, the majority of voters tend to oppose trade liberalization (see, for instance, Fernandez and Rodrik 1991), thus suggesting that FTA adoption would lower the probability of staying in office. In short, reverse causality from FTA formation to government stability is ultimately an empirical issue and the empirical strategy tries to minimize this potential source of bias. 4 Empirical Strategy and Data The empirical analysis tests the theoretical relation between the probability of political turnover and the odds of adopting a trade agreement. It consists in a two-step procedure. First, I estimate the probability that the government of country i changes in year t and then use this predicted probability in a dyadic regression to explain FTA adoption. A key empirical challenge is to characterize the event of government turnover and then predict its likelihood. While different concepts define a government, its chief executive or leader and its main party are two elements that usually define it. I use therefore distinctively changes of the head of government and of the largest government party as the events characterizing political turnover. In line with the theoretical argument, the change of government should be such that the capacity of the losing side to participate in the policy-making process is, at least, substantially reduced. In general, a change of the governing party tend to create more effective political turnover than a change of the leader 8 The ideal trade policy position of each party may stem from other aspects than purely ideological reasons, such as constituency interests. 8

9 (e.g., this is especially the case of lame ducks that will be automatically replaced because of binding term limits, which do not apply to parties). Yet, the argument may not apply generally. For instance, in cases of highly personalized political systems and weak parties, political power is very much related to the leader and not to the parties that support her. To measure the probability of government change, a parametric approach is adopted. In particular, two main variables are proposed to explain the likelihood that the incumbent government is replaced. One is the average tenure of past heads of government (AveT en), the logic being that an incumbent observing more stable previous governments should have a longer time horizon than an incumbent serving a country with an history of easy government turnover. I compute the average duration (in years) of government leaders in the preceding 20 years. The other measure is the share of parliament seats held by the government s parties (Seat). This should reflect the party structure that supports the incumbent. Although each of these measures might identify only partially the probability of government turnover, their significant explanatory power makes me confident about the robustness of the mechanism at work. Furthermore, previous work by Campante et al. (2009) and Persson et al. (2003) has used similar measures to proxy for the stability of the incumbent government. Consequently, the following discrete time model for the hazard that the incumbent government loses office is estimated for each type of political turnover event: P r(newgov i,t = 1) = F [γ 0 AveT en i,t 1 + γ 1 Seat i,t 1 + γ 2 X i,t 1 + D i [T ]] (5) where NewGov i,t is the change in government indicator, namely NewP arty for the largest government party and NewHead for the leader. The function F [.] is either the cumulative logistic distribution function in the Logit model or the identity function in the linear probability model. As discussed above, the coefficients γ 0 and γ 1 are expected to be negative. The matrix X includes the same type of variables that are related to the likelihood of FTA formation so as to be able to identify changes in the likelihood of political turnover due only to variation in the two main covariates, AveT en and Seat, and D i [T ]. The duration dependence term D i [T ] is a flexible functional form of the baseline hazard of political turnover. Following Beck et al. (1998), D i [T ] includes a natural cubic spline of the time since the last change of government (T ) occurred and a set of dummies for the number of previous events in each country, to take into account the fact that multiple events may occur in each country 9. The estimates of 5 are then used to predict the probability of political turnover, σ i,t P r(newgov i,t = 1). In the second step of the empirical analysis, the main estimating equation has the following 9 Six cubic splines of time since the last change in the head of government and four of the time since the last change in the largest government party are selected on the basis of model fit (i.e., likelihood-ratio tests between logit models with one to seven splines). For the baseline hazard of change in the largest government party, seven dummies are included to control for the number of previous events while twelve dummies one for each previous event are added to the model for the probability of change in the head of government. 9

10 dyadic structure: P r(newf T A ij,t = 1) = F [β 0 σ i,t 1 + β 1 Z i,t 1 + β 2 W ij,t 1 ] (6) where NewF T A is an indicator variable taking value 1 if an FTA between country i and country j enters into force in year t. The coefficient of interest, β 0 is expected to be positive; more unstable governments are more likely to sign FTAs. The matrix Z includes country-specific variables that are included in X (see equation 5) and that are not strictly bilateral characteristics. These are mainly features of political institutions (see below). The term W ij,t 1 includes other bilateral determinants of FTA formation. Equation 6 relates a directed, country-specific variable, the predicted probability of government turnover, to an undirected event, the signing of an FTA between two countries. While this provides a straightforward interpretation of the marginal effect of σ i,t 1, it also enlarges artificially the sample, since both (ij, t) and (ji, t) observations are included. To validate the empirical results, I therefore run also a fully bilateral specification: P r(newf T A ij,t = 1) = F [γ 0 σ ij,t 1 + γ 1 B ij,t 1 ] (7) where σ ij,t, is a measure of the odds of government turnover in countries i and j. Although the theoretical framework in Section 3 takes the perspective of a single country, the same analysis applies to any other country. Two countries are more likely to form an FTA if in both of them the incumbent governments face a high probability of being replaced. Empirically, this suggests taking the minimum of the turnover probabilities in the two countries, σ ij,t min( σ i,t, σ j,t ). The matrix B includes all the bilateral determinants of FTA that are in W and the variables in Z, after having made those appropriately specific to each country-pair. Identification of the coefficient of interest becomes problematic if the likelihood of FTA formation affects the probability of government turnover. Two features of the specifications 6 and 7 makes this condition unlikely to occur. First, country-pairs are followed until they sign an FTA or otherwise until the end of the sample and all covariates are lagged one year 10. Moreover, the probability that the incumbent government is replaced is predicted without taking into account participation in trade agreements. Estimation of the main regression of interest (6) requires data on FTAs. Data on trade agreements are from Baier and Bergstrand (2004, 2007). They record the year when a trade agreement between two countries entered into force during the period A characteristic of this data source that makes it particularly suitable for my analysis is that it classifies also the type of trade agreement on a 6-point scale, from no trade agreement to Economic Union, the strongest form of economic integration. The 10 All country-pairs that signed a trade agreement before the sample period are dropped because they do not contribute to the likelihood of FTA adoption as well as country-pairs that switched their FTA status more than once in the sample. Baldwin and Jaimovich (2010) and Martin et al. (2010) adopt a similar specification. 10

11 theoretical arguments should apply to trade agreements that entail destruction of rents related to trade policies. Shallow forms of integration whereby only one country grants trade preferences ( Non-reciprocal preferential trade arrangements, such as the AGOA initiative) or trade is only partially liberalized ( Preferential trade arrangements ) should not lead to an important reduction of rents available from protectionist policies and hence are not effective means of binding future protectionist policies. Therefore, I focus only on forms of economic integration consisting of Free Trade Agreements or deeper ( Customs Unions, Common Market and Economic Unions ). Consequently, I refer to FTAs as Free Trade Agreements or deeper types of economic integration. The variables for the change in government, NewHead and NewP arty are from the Cheibub et al. (2010) database and the Database on Political Institutions (DPI, ), respectively 11. Likewise, the two main identifying covariates for the probability of political turnover, AveT en and Seat are constructed from the Cheibub et al. (2010) database and the DPI 12. Identification of the effect incumbent instability has on the likelihood of FTA formation requires careful consideration in determining the set of control variables so as to avoid omitted variable bias. First, institutional features that may affect both the probability that the government is replaced and the propensity of entering trade agreements are added as independent variables to 5 and 6, 7. Democratic governments tend to have shorter time horizons than dictatorships and democracies are generally more open to free trade (Mansfield et al., 2002). Furthermore, most definitions of regime type already incorporate some measure of political turnover, which makes it difficult to isolate its effect from other features of a regime type 13. The classification of Cheibub et al. (2010) permits to identify autocracies that have been classified as such only because alternation in power has not been observed. I therefore modify their regime type indicator so that a country is classified as a democracy if it holds multiparty open elections 14. As the authors rightly argue, this is not sufficient to identify democracies, but it serves the empirical purpose of isolating the effect of incumbent instability on the propensity to sign FTAs. In the empirical analysis, I thus use the modified democracy indicator (Democ) and an indicator for democratic country-pairs (BothDemo) in 6 and 7. In their theoretical model, Grossman and Helpman (2005) show that majoritarian systems are more protectionist (in terms of average tariffs) than proportional systems, a 11 The government s chief executive is generally classified as the President (or Head of State) in presidential systems, the Prime Minister (or Head of Government) in parliamentary systems and king, presidents or de facto rulers for dictatorship (Cheibub et al., 2010). The largest government party is identified by the gov1me variable in the DPI. Since the DPI records variable as of January, 1st, of each year, while all other time-variant variables are recorded as of December, 31st, I treat the DPI variables as if they are lagged one year. 12 A key advantage of using the Cheibub et al. (2010) database to construct the AveT en variable is that it records instances of multiple changes in government per year. 13 This is especially true for the polity2 scores that are computed on the basis of three institutional features: (i) Competitiveness of political participation; (ii) openness and competitiveness of executive recruitment; and (iii) Constraints on the executive. All three components are highly correlated to long-run political competition and turnover. 14 This amounts to re-classify the regime type dummy as equal to one regardless of the value taken by the type2 variable in their database 11

12 prediction that Evans (2009) confirms empirically in a cross-section of countries. At the same time, the electoral system may affect the political horizon of the incumbent. I thus construct from the DPI database a categorical variables that classifies electoral systems as fully majoritarian (M aj) when legislators are elected solely on the basis of a first-past-the post rule, fully proportional (P rop) if candidates are elected on the basis of the votes received by their parties, and mixed when both elements are present (see also Evans 2009). In the fully bilateral specification 6, I include dummies for whether both countries in the pair have a majoritarian system (BothM aj) and for whether only one country adopts such a system (OneM aj). The type of the government, namely presidential or parliamentary, is another possible confounding factor, as it may affect the incumbent stability and, at the same time, economic policies (Persson and Tabellini, 2003). For instance, presidential systems tend to confer more powers and greater autonomy to the president, thereby insulating her from the need of constantly obtaining parliament support. A dummy for purely presidential systems (P res), taken from the DPI database, is thus added to the battery of control variables. Likewise, indicators for pairs with one presidential system (OneP res) and for pairs with both countries having presidential regimes (BothP res) are included in the FTA regressions. As discussed in the previous Section (3), the ideological preferences of the incumbent government may shape its attitudes towards trade agreements. To control for this aspect, I use the ideology classification, Right, Center, Left or No ideology, with respect to economic policies proposed by the DPI database. The survival-type structure of the sample in the FTA regressions should avoid any reverse causality bias. In the fully bilateral specification (7), I control for ideological similarity in the country-pair. Other country-year variables are added to the auxiliary regression to predict the hazard rate of government change. These are the growth rate of GDP per capita (GDP pcgr) and the inflation rate (Inf l), which should capture the importance of the economic voting argument (see, e.g., Fair 1978) and measures of cultural polarization, namely the ethnic fractionalization index (Etf rac) of Alesina et al. (2003) and the share of population affiliated to major religions (Religion), which have been found to be important determinants of political institutions (Aghion et al., 2004) and hence could also affect politicians time horizons. The rationale for including these variables along with other controls is to make sure that the partial effect of the two key explanatory variables AveT en and Seat is not picking up any other reasonable determinant of the likelihood of political turnover. In the fully bilateral specification 7, the country-pair average of each variable is included. Following the existing literature on the economic determinants of FTAs (e.g., Baier and Bergstrand 2004 and Egger and Larch 2008), the set of bilateral controls W includes the log sum of the real GDP of country i and j in year t (GDP sum) to capture market size effects, their difference in absolute value (GDP dif) to measure dissimilarity in country size, and the absolute difference in log real GDP per capita (GDP pcdif) between the countries in the pair as a proxy for differences in relative factor endowments. Furthermore, bilateral measures of trade costs, namely distance (Ldist), remoteness (intra- and inter- continental) and the 12

13 time zone difference in hours (T imediff) are also added to W. Other bilateral controls are dummies for WTO membership (BothW T O), past colonial relationship (Colony), common legal origin (Comleg) and common official language (Comlang). The same type of bilateral determinants are included in the set of country-year controls X for the auxiliary regression 5. For instance, I add dummies for having ever been a colony of France, UK and Spain to control for colonial status and the average distance to other countries to control for geographical remoteness. Membership in the WTO is not included to avoid any indirect effect of participation in trade agreements 15. The final sample period used in the estimation covers 32 years, from 1975 to 2005 and includes about 120 countries, depending on the set of explanatory variables employed 16. Table 1 report some summary statistics for the variables used in the baseline FTA regression 6. [Table 1 here] 5 Results 5.1 Auxiliary Regression Table 2 reports the Logit estimates of the auxiliary regression 5 that is used to predict the probability of a change of government. As expected, the average tenure of previous incumbents and the share of parliament seats held by the government have a negative and significant effect on the hazard of government turnover, both in terms of change of the largest government party and replacement of the chief executive. Their effect is robust to the inclusion of the type of control variables that are also likely to affect the country-pairs likelihood of signing FTAs (Columns(2) and (6)). Of these covariates, only the growth rate of GDP per capita has a significant coefficient across the different specifications, suggesting that better growth performances enhance incumbent stability. As a robustness, I then add two leader-specific covariates, the leader s Age (in log) and an indicator variable for an irregular mode of Entry into office (Columns (3) and (7)) from the Archigos database (see Goemans et al for details). For years with multiple leaders, I take the characteristics of the executive that served the highest number of days. Despite the resulting measurement error, the two variables have a significant effect on the probability of having a change in leadership. Older leaders are intuitively more likely to be replaced, while leaders that gain power irregularly (e.g., through a coup d état) are less likely to lose office. Not surprisingly, the two variables play no role in explaining the likelihood of change of the largest government party. The Logit coefficients on the covariates of interest, AveT en and Seat, remain virtually unchanged. Finally, a Conditional fixed-effects Logit estimator is employed to see whether results are confirmed exploiting only the within-country variation 15 A dummy for membership in the WTO is nevertheless not significant in the auxiliary regressions (results not reported). 16 I follow Campante et al. (2009) and exclude Switzerland from the analysis given its unique rotating system of presidency for which the two main variables identifying political turnover, AveT en and Seat, would have no meaningful interpretation. 13

14 (Columns (4) and (8)). Countries that did not experience any change of government, in terms of ruling party and head of government, in the estimating sample do not contribute to the respective likelihoods and are dropped from the estimation. The effect of the Seat variable stays negative and significant, while the coefficient on the AveT en variable loses significance. This is probably due to the overlapping role of the dummies for the number of previous changes of government, which become jointly significant also in the regression explaining the hazard of change of the largest government party. In general, the duration dependence terms play an important role in determining the probability of government turnover. The cubic splines of times since the last event (change of the largest government party or of the government leader) are jointly significant in all specifications. The number of previous events within each country also contributes to the baseline hazard of government turnover, though not significantly in the case of changes of the largest government party. In sum, the data exhibit duration dependence that must be taken into account to predict the probability of government change. The Logit estimates from Columns (2) and (6) are used to obtain country-year predicted probabilities of change of the largest government party and change of leaders 17. Likelihoodratio tests of the the models in Columns (2) and (6) against their counterparts without the AveT en and Seat variables soundly rejects the hypothesis that the latter versions have better fit than the proposed models 18. To further assess the predictive power of the two main covariates excluded from the FTA regression 6, I rely on the expected Percentage Correctly Predicted (epcp) measure proposed by Herron (1999). Unlike the more common PCP measure, the epcp does not require to set a threshold in the predicted probabilities to classify cases as correctly predicted or not. This is particularly relevant to the current analysis, since the events of interests are relatively rare (changes in the largest government party occur about 12% of the observations, whereas changes in the head of government arise in 19% of all the country-year cases) and using the traditional threshold of 0.5 would inflate the model predictive power. On average, the model in Column (2) predicts correctly 81.2% of the cases, while, after exclusion of AveT en and Seat, the epcp equals 80.7% and goes down to 80.2% if the duration dependence terms are also excluded. Similarly, the estimates in Column (6) give an epcp of 71.9% and excluding AveT en and Seat decreases the predictive power by 1%. Therefore, given the relatively rare occurrence of the events of interest, the variables AveT en and Seat, together with the duration dependence terms, play an important role in predicting government turnover. [Table 2 here] 5.2 FTA Regression Before turning to the results of the main regression (6), I explore some cross-sectional and purely descriptive relations between the predicted probabilities of government turnover and 17 The variable Infl is not used in the prediction in order to maximize sample size. Results are not altered. 18 For the Logit model explaining the probability of party change, the test statistic equal χ 2 party = For the model explaining the hazard of leader turnover, the test statistic is χ 2 head =

15 the propensity to sign FTAs. The theoretical framework suggests that countries where the probability of government turnover was high during the sample period should have engaged more in FTAs. I thus take the within-country average predicted probabilities of party change and of leader change and relate each of them with the share of imports coming from FTA partners at the end of the sample period. Adopted also in the related work of Ornelas and Liu (2011), the latter measure aims to capture the extent to which countries engaged in FTAs. However, it incorporates also the trade effects of FTAs, especially for the older ones, about which the theory does not have specific predictions. Figure 1 plots the import share from FTA partners against the predicted probabilities of government turnover. Despite the large amount of dispersion, there is a significant positive relationship between the likelihood of government turnover and the share of imports coming from FTA partners. European countries, with their high level of economic integration and relatively unstable governments seem to drive an important portion of the upward trend. This piece of evidence is nonetheless only suggestive of the possible relation between government instability and FTA formation. [Figure 1 here] To assess the validity of the suggested prediction, I estimate regression 6 using the predicted probabilities from 5. Standard errors are clustered at the undirected dyad level to take into account the fact the the event of interest, FTA formation, is bilateral (e.g., the indicator NewF T A equals one in 1994 for the observation (USA, MEX) and for the observation (M EX, U SA)). Table 3 reports the estimates. As expected, having more unstable incumbents increases the likelihood of FTA formation. Before controlling for other determinants of FTAs, the coefficient on the predicted probability of turnover of the main government party is positive and significant (Column (1)) as well as the one on the probability of having a new chief executive (Column (6)). Columns (2) and (7) show the baseline estimates with the full set of controls. While missing data on the explanatory variables reduce the sample size, the predicted probabilities of government turnover keep the expected sign and statistical significance. As for the other covariates, the results broadly confirm existing evidence on the determinants of FTAs. Countries with larger markets are more likely to sign FTAs, while FTAs are less likely to be observed between countries of different sizes. Similarly to Egger and Larch (2008) and Baier and Bergstrand (2004), I find that country-pairs with dissimilar factor endowments, as proxied by real GDP per capita, are less likely to adopt FTAs. Being located in the same continent and far from the rest of the world facilitates the formation of trade agreements, while geographic and time distances are impediments to the creation of FTAs. Democratic pairs are significantly more likely to sign FTAs than pairs with one or no democratic countries. Therefore, other features than increased government turnover that comes with democracy have a significant impact on the likelihood of FTA formation. WTO membership has also a positive and significant effect on the likelihood of discriminatory liberalization through FTAs. Besides affecting significantly the probability of government 15

16 turnover, the growth rate of GDP per capita contributes positively also to the likelihood of FTA formation. Interestingly, other features of political institutions have a significant effect on the creation of FTAs. Countries with presidential regimes are significantly less likely to form FTAs. Given that the type of political regime does not affect political horizons on the basis of the auxiliary regressions (see Table 2), future research might be needed to identify alternative mechanisms through which the type of political regime affects the propensity to sign trade agreements. Majoritarian electoral systems, which have been found to set higher tariffs on average (Evans, 2009), seem to be less likely to form FTAs, but the evidence is not conclusive. Among other controls, ethnic fractionalization affects negatively and significantly the likelihood of FTA formation, probably reflecting, at least partly, the difficulty of adopting trade reforms in heterogeneous and polarized societies. Finally, the ideological position of the incumbent does not seem to have a clear and significant effect on FTA formation, although inference in this case should be particularly cautious due to inevitable measurement error. In the other Logit specification, I add two control variables. First, the economic gains from FTAs are closely related to the existing amount of trade between the signatory countries. At the same time, trade agreements are, at least theoretically, meant to foster trade relationships. I therefore include the (log of) bilateral trade (Biltrade), lagged three years to limit simultaneity bias. This solution is by no means perfect as some trade agreements may be anticipated more than three years before the actual entry into force, while others may come as a relative surprise to exporters and importers. I further add a measure of regime similarity, equal to the absolute difference of the polity score (P olitydiff), which has been recently found to be a negative predictor of FTAs (Baldwin and Jaimovich 2010, and Martin et al. 2010). A problem with this variable is that the polity scores already embed some institutional aspects of political competition that inevitably affect the incumbent s political horizon (see footnote 13). As shown in Columns (3) and (8), the new regressor enters the equation negatively and significantly, confirming previous evidence. Bilateral trade is a positive and significant determinant of trade agreements, as expected. Not surprisingly, the most important change in the estimates is that now market size has no effect on FTAs. More importantly, the sign and significance of the probability of party turnover does not change, while the coefficient on the probability of leader change is no longer significant, a shift that is not caused by the changing sample size 19. Therefore, it seems that the polity similarity measures is indeed picking up, at least partly, some differences also in incumbent stability between the two countries in the pair. In the last regressions, I tackle the issue of time-invariant unobserved heterogeneity. In this case, a Conditional fixed-effect Logit estimator would reduce the sample drastically given that only 6% of the country-pairs signed an FTA in the sample period. A linear probability model with (directed) country-pair fixed-effects is therefore applied to the augmented specification. The results show a positive and significant effect of the probability of party turnover. When the governing party of a country is more likely to lose power, 19 Running the baseline specification on the sample of Column (7) gives a positive and significant Logit coefficient on the probability of change of leader, but smaller than the one reported in Column (6). 16