Impacts of regional trade agreements on international trade patterns revisited

Transcription

1 . Impacts of regional trade agreements on international trade patterns revisited NGUYEN Duc Bao GREThA - UMR CNRS 5113, Université de Bordeaux, Av Léon Duguit, Pessac Cedex, France Abstract This paper assesses the ex-post effects of seventeen regional trade agreements (RTAs) all over the world based on a gravity model with solid theoretical foundations that involve the Anderson and van Wincoop s multilateral resistance terms. The study covers 160 countries over the period of 1960 through Over a long time span of 55 years, this period covers nearly all waves of regionalism taking place on earth in the wake of the Second World War. We aim to explore whether regional trading blocs around the world have stimulated trade among members as well as trade with non-members or they have increased members trade to the detriment of non-members. The gravity model is estimated by a Poisson pseudo maximum likelihood (PPML) estimator with fixed effects to handle problems of zero trade flows and control for heteroskedasticity. The inclusion of various dummy variables for RTAs in the gravity model allows us to correctly determine Vinerian trade creation and trade diversion effects in the wake of the establishment of trading blocs. Our study is among the first efforts in the literature on RTAs effects that have used the PPML estimation technique to deal with the zero trade flows and has taken into account the Anderson and van Wincoop s multilateral resistance terms at one time. The analysis finds that most RTAs lead to significant trade creation between members, along with trade diversion effects in terms of bloc exports and imports in many cases. RTAs located in America and Europe appear to have more significant effects on intra-bloc trade as well as on extra-bloc trade than Asian and African RTAs. JEL classification: F10, F13, F15. Keywords: gravity model, regional trade agreements, regional integration, trade creation, trade diversion. This is only a preliminary version. Tel.: address: duc-bao.nguyen@u-bordeaux.fr (NGUYEN Duc Bao)

2 1. Introduction Regional trade agreements (RTAs) have rapidly proliferated around the world in recent years. As of July 2016, there are 267 RTAs had been notified to the World Trade Organization (WTO) and are currently in force. This type of trade agreement has become a key component of trade policy for many countries around the globe. Balassa (1961) determined various forms of integration for RTAs, such as free trade areas, customs unions, common markets, economic unions and total economic integration. These forms of RTAs are based on different degrees of suppression of discrimination resulting from trade barriers and national economic programs among member countries. RTAs have always accompanied by the multilateral trading system. However, we have witnessed since the early 1990s a development in the debate on the relationship between regionalism and multilateralism. When the Uruguay Round overcame many challenges in its negotiations before being finally signed in 1994, the number of RTAs entering into force had steadily increased since 1995 following the establishment of the WTO 1. Several studies consider the proliferation of RTAs as a major challenge to the multilateral trade process. Bhagwati (1991) underlined that regionalism embodied a discriminatory characteristic and could induce perverse effects. A steady increase in RTAs could also be detrimental to non-members or the rest of the world (Baldwin, 1993). According to Bhagwati (1995), countries have a greater power will gain from the trade liberalisation while the smaller ones are losing from it within regional groups, and the regionalism could also increase the risk of conflicts between regional trading blocs. Conversely, the regional trading system is perceived by others as a step towards the breakthrough of multilateral trade liberalisation under the umbrella of the WTO. Summers (1991) argued that regional trade liberalisation generates an advance on multilateralism and it leads to more trade creation than trade diversion. Thus, the inclination of regional trade integration did not hinder the achievement of the Uruguay Round negotiations because the countries that drove the multilateral trade system after the Second World War are the same ones that promoted the regional trade liberalisation (Baldwin, 2004). Moreover, RTAs can also encourage foreign direct investment (Lawrence, 1996; Kimura and Ando, 2005; Freund and Ornelas, 2010), and economic growth in member countries through technological transfer. The upsurge in RTAs throughout the world over the past two decades has resulted in the emergence of a dense, complex network of RTAs in which there are several overlapping agreements among the same trading partners. In the context of the multilateral trading system, RTAs are operated under the rules introduced by the WTO. It may seem that RTAs violate the Most-Favoured-Nation (MFN) principle, one of the most important pillars of the WTO, which prohibits countries from discriminating between their trading partners 2. However, RTAs are considered as an exception to the MFN obligation. In fact, the WTO rules lay down a legal framework for RTAs covering trade in goods: Article XXIV (GATT 1994) and paragraph 2(c) of the Enabling Clause (GATT 1979). In this study, we will focus on the term "regional trade agreements" used under the GATT/WTO rules, which takes into account agreements covering 1 Acharya (2016) found that about three RTAs on average were notified per year during the General Agreement on Tariffs and Trade (GATT) period (from 1948 to 1994) compared with the WTO period (since 1995) when on average twenty-five RTAs have been notified per year. 2 The MFN requires that any trade advantages one country grants to another member must also be offered to all other WTO members. 2

3 liberalisation of trade in goods: free trade areas and customs unions 3. In the context of the growing trend towards regionalism, owing to the steady increase in the number of RTAs established all over the world since the early 1990s, this paper will revisit the ex-post effects of RTAs on the multilateral trading system over a long time span from 1960 until This period covers nearly all waves of regionalism taking place on earth in the wake of the Second World War. We aim to study the impacts of several RTAs on their intra-bloc trade and the tendency of member countries to trade with the rest of the world in the wake of their formation. We thus explore whether regional trading blocs around the world have stimulated trade among members as well as trade with non-members or they have increased members trade to the detriment of non-members. The motivation for our study comes firstly from the renewed interest in the application of gravity model to analyse bilateral trade flows, especially after the emergence of its more solid theoretical foundation in the early 2000s. Over the past fifty years, the gravity equation has become the most fruitful and dominant empirical framework for analysing international trade. The basic gravity model introduced by Tinbergen (1962) found that bilateral trade flows between two trading partners depended on their countries incomes positively and bilateral distance negatively. However, this model, which is inspired by Newton s law of gravity, did not have solid underpinnings in economic theory. Several authors have attempted to develop strong theoretical foundations for the gravity model since the late 1970s 4. Much improvement has been achieved; more recently and more notably, Anderson and van Wincoop (2003) laid out and popularised the solid theoretical framework of the gravity equation that takes into account multilateral resistance terms, as also introduced by the two authors. On the other hand, questions have also been asked in the empirical literature about the appropriate formulation of variables in the gravity equation, mostly the dummy variables that can assess impacts of RTAs. There has been a revolution in the choice of dummy variables for a better examination of trade effects associated with RTAs regarding trade creation and trade diversion introduced by Viner (1950). Based on a static and partial equilibrium framework, Viner (1950) argued that an RTA did not necessarily enhance member countries welfare. The author has taught us that RTAs, under the form of free trade areas or customs unions, are likely to produce trade creation if member countries import more from efficient producers located in other members at the expense of less efficient producers in the domestic market. Accordingly, RTAs enhance efficiency from both sides regarding production and consumption and increase welfare for member countries. By contrast, RTAs may lead to trade diversion as well, when members take the imports away from the most efficient suppliers (low-cost producers) in the rest of the world and give them to inefficient suppliers (higher-cost producers) in other member countries. This situation leads to an inefficiency in global production, which is detrimental to the outsiders of RTAs. It can also be harmful to member countries when the consumer surplus is impossible to outweigh the cost of the inefficiency in production. The net effect of trade liberalisation following the formation of RTAs is ambiguous and depends on whether the trade creation effect or the trade diversion effect is dominant 5. Although 3 These two terms are adopted in Article XXIV (GATT 1994). 4 see Anderson (1979); Helpman and Krugman (1985); Bergstrand (1989); Deardorff (1998); Baier and Bergstrand (2001). 5 Viner (1950, p.44) states that: "Where the trade-creating force is predominant, one of the members at least must benefit, both may benefit, two combined must have a net benefit, and the world at large benefits; but the outside world loses, in the short-run at least... Where the trade-diverting effect is predominant, one at least of the member countries is 3

4 Viner s findings only focus on static impacts of RTAs and do not clearly address their net welfare effects, his two principal concepts of trade creation and trade diversion have significantly inspired later theoretical and empirical studies on RTAs effects 6. Since then, results from the empirical literature still have been quite mixed. In this paper, we will adopt the method including three dummy variables for each RTA to adequately capture Vinerian trade effects. These dummy variables will explain the impacts of each RTA on intra-bloc trade, members imports and members exports to the rest of the world. Our motivation also comes from the question about the proper estimation techniques that can handle the presence of zero trade, which arises prominently in trade data. The gravity model has been widely estimated using cross-sectional approach or panel data approach. The latter method emerged in the early 2000s and had been favoured by many researchers because it can alleviate the problem of heterogeneity among the countries for which cross-sectional method is unable to control. Most of the studies using cross-sectional data or panel data to assess the gravity model first have to transform the model to log-linear form by taking logarithms and then estimate the log-linear specification of the gravity model. However, this method is questionable since the loglinear model cannot determine observations concerning zero trade in bilateral trade flows. By excluding the meaningful insight about pairs of countries that do not trade with each other, these studies could generate biased results (Helpman et al., 2008). Particularly in the case of our study, when the proportion of zero trade reaches about 50% of total potential observations, the choice of a proper estimation technique that can deal with the problem of zero trade is quite important. As a result, what is the proper measure to which we can resort to assess the zero trade flows? In this paper, to overcome the zero trade problem, we will estimate the gravity equation by applying the Poisson pseudo maximum likelihood (PPML) estimator proposed by Santos Silva and Tenreyro (2006). Moreover, these authors also find that this approach is consistent in the presence of heteroskedasticity in trade data. In this paper, we will show that PPML estimator with time-varying country fixed effects can provide convincing results of RTAs effects on international trade. This study will contribute to the literature on the ex-post effects of RTAs concerning trade creation and trade diversion using a larger sample of countries and RTAs, and a longer time span than most other empirical studies on this question. Our study is among the first efforts in the literature on RTAs effects that use the PPML estimation technique to deal with the zero trade flows and take into account the Anderson and van Wincoop s multilateral resistance terms at one time. Within the scope of our study, we cover most of the plurilateral RTAs in force in the world and being notified to the WTO with a total of seventeen RTAs. We are able to capture the impacts of RTAs around the world and to observe the distinct trade patterns of RTAs located in different geographic regions on earth and formed by countries with various levels of development. Results from the PPML estimator suggest that in the wake of their entry into force, many RTAs have generated a significant increase in trade flows between member countries, yet many cases, with detriment to the rest of the world. Moreover, RTAs located in America and Europe seem to have more significant impacts on intra-bloc trade and extra-bloc trade than Asian and African RTAs. bound to be injured, both may be injured, the two combined will suffer a net injury, and there will be injury to the outside world and to the world at large". 6 Many authors have attempted to enhance the Vinerian theory and found that parts of Viner s analysis were not complete and acute by introducing the elasticities of demand, the dynamic effects into the model or by taking into account the enlargement of trading bloc over time, based on a partial equilibrium (Johnson, 1960) or a general equilibrium framework (Meade, 1955; Lipsey, 1970; Kemp and Wan, 1976). 4

5 The remainder of the paper is organised as follows. Section 2 summarises the empirical literature on RTAs effects. Section 3 briefly specifies our econometric approach, the gravity model and describes the data set. Section 4 presents our main empirical results involving the average effects of seventeen RTAs over the period of Some robustness analysis is provided in Section 5. Section 6 concludes and indicates some caveats in our paper. 2. Empirical literature reviews Began with only one dummy variable to capture RTAs effects on intra-bloc trade, the works on RTAs impacts had been extended with the addition of a second and third dummy variables to measure RTAs effects on the trade of member countries with non-members. This improvement changes the way researchers interpret the empirical results as they could assess more carefully the trade creation and trade diversion effects following the creation of RTAs, as introduced by Viner (1950). Effects on trade flows between regional blocs and the rest of the world resulting from the formation of RTAs are more clarified with the support of different regional dummy variables. We demonstrate in this section the path of development in RTAs empirical analysis based on the improvement in the set of regional dummy variables. In the interest of evaluating the effects of an RTA on trade flows, many studies first enhanced the basic gravity model by including a regional dummy variable to measure its effects on trade flows between member countries. This dummy represents the sum of trade creation and trade diversion effects generated by the RTA. The results obtained in various studies including just one regional dummy variable have been conflicting. Based on the cross-section gravity model, Aitken (1973), Brada and Méndez (1985) showed that the European Economic Community (EEC) had significantly positive effects on trade flows between participating countries, while the works of Bergstrand (1985), Frankel et al. (1995) found insignificant effects in the same RTA. Meanwhile, Bayoumi and Eichengreen (1997) found significant effects regarding the enlargements of the EEC in the years of 1972, 1981 and In the case of trading blocs in America, Frankel (1997) found that the North American Free Trade Agreement (NAFTA) and the Southern Common Market (MERCOSUR) have positive and significant impact on their intra-bloc trade by means of pooled estimation over the period of 1970 through 1992, while the bloc effect of the ANDEAN Community is insignificant. Cheng and Wall (2005) and Bussière et al. (2005) found that these RTAs all create a positive impact on intra-bloc trade based on panel data method. As regards the European Free Trade Association (EFTA), Frankel et al. (1995) showed that the coefficients for the bloc effect of the EFTA were never significant during studying period, whereas Aitken (1973) found strong evidence that the intra-bloc trade between the EFTA members is above the expected levels predicted from the gravity model following the formation of the bloc, although both studies applied cross-sectional data. Since the studies including only a single regional dummy variable were not capable of capturing the effect of an RTA on trade flows between bloc members and non-members, many empirical studies, in the late 1990s, have added a second regional dummy variable to measure it. This dummy is a binary variable assuming the value of 1 if one of the two countries in a bilateral country-pair participates in a given RTA and the other does not, and 0 otherwise. Frankel (1997) indicates that this variable accounts for the level of openness of RTAs. Studies can identify the trade creation and trade diversion effects separately of an RTA, thanks to the combination of the former regional dummy variable and the second one. In the case when the formation of an RTA leads to an increase in intra-bloc trade and also promotes extra-bloc trade or keeps the latter 5

6 unchanged, this RTA is likely to have trade creation effect. On the other hand, if an RTA increases trade flows between member countries to the detriment of their trade flows with the outsiders, it seems to induce trade diversion effect since the intra-bloc trade can substitute for the trade flows coming from non-members. When the openness term of RTAs is taken into account, Frankel (1997) found significant negative coefficient estimates for trade between members and non-members in the cases of the EFTA and the NAFTA, along with significant and positive coefficient estimates for intra-bloc trade. The author also found that the MERCOSUR and the free trade area by the Association of South East Asian Nations (ASEAN) have an increase in propensity to trade with non-members because the estimated coefficients of both regional dummy variables are positive. For the EEC, Frankel (1997) showed that in 1980 and 1985, the EEC members are more open to trade with the rest of the world than one would predict from the standard gravity variables through the openness coefficient highly significant and positive. By contrast, Bayoumi and Eichengreen (1997) found evidence of negative effects on extra-bloc trade following the formation of the EEC in the 1960s. Lee and Park (2005) showed that the European Union (EU), the NAFTA, the MERCOSUR lead to an increase in extra-bloc trade and greater progress in trade between member countries using panel data, whereas the Central American Common Market (CACM) and the Common Market for Eastern and Southern Africa (COMESA) contribute to a significant decrease in extra-bloc trade. Nonetheless, studies including these two dummy variables seldom precisely identify the trade creation and trade diversion effects for RTAs. Since the dummy variable for the level of openness (extra-bloc trade) covers both of members total exports and imports of goods, it is not capable of separating the impact of the trading bloc on the extra-bloc trade regarding exports from the one regarding imports. As Soloaga and Winters (2001) noted, the import and export flows of member countries may come after different paths. When an RTA improves their trade with non-member countries, the gravity model with two regional dummy variables cannot identify whether this effect comes from the exports towards the rest of the world or the imports from non-members. Similarly, this problem also arises when an RTA has negative effects on extra-bloc trade. The most recent studies since the 2000s have once again extended the model by including a third regional dummy variable to create a set of three dummy variables individualised for each RTA. Among these three variables, one will measure the intra-bloc trade between participating countries, the second one will try to explain the export flows of member countries towards nonmembers, and the third one will capture the import flows from the rest of the world reaching member countries. The last two dummies seek to indicate the level of overall openness for the trading bloc in terms of export and import flows. For the purpose of interpreting the effects of a given RTA, these studies need to compare the value of coefficient estimate for intra-bloc trade and the ones for the extra-bloc trade regarding exports and imports. As a result, when an RTA induces an increase in intra-bloc trade (a positive coefficient) combining with an increase in extra-bloc trade in terms of exports or imports with non-members (a positive coefficient for extra-bloc exports or imports), it can be identified as trade creation regarding export flows or import flows, respectively. By contrast, if an increase in intra-bloc trade combines with a decline in extra-bloc trade regarding exports or imports (a negative coefficient for extra-bloc exports or imports), this situation will be determined by export diversion effect or import diversion effect, respectively. Moreover, as regards the effects of RTAs on welfare, one can identify an RTA as being harmful to non-members when the coefficient for the extra-bloc trade regarding exports to non-members is negative, which leads to a falling inclination of member countries to ship their goods to the rest of the world. As a result, it 6

7 results in welfare losses for the outsiders. On the other hand, supposing the producers within an RTA cost more to produce goods than those in the rest of the world, it means an inefficiency in the allocation of resources in the world, which is also detrimental to the outsiders of RTAs (Trotignon, 2010). Results from the empirical literature involving RTAs impacts on extra-bloc trade in terms of bloc exports and imports have also been mixed. Soloaga and Winters (2001), in a crosssection study, found the presence of export diversion effect in the cases of the EU and the EFTA, whereas Carrère (2006) and Trotignon (2010) found an increase in the propensity of the EU to export to the rest of the world employing the panel data approach. Meanwhile, Endoh (1999) pointed to trade creation effect in the EEC over the period through all three channels: intra-bloc trade, extra-bloc trade in terms of exports and imports as well. For the ANDEAN Community, the MERCOSUR, the NAFTA and the ASEAN, Carrère (2006) showed a falling propensity to import from the rest of the world in the wake of the formation of these RTAs, while Trotignon (2010) found opposite effects as the author demonstrated an increase in extra-bloc trade regarding imports coming from non-members. Although the two authors both use panel gravity model, their studies can lead to very conflicting results due to the discrepancy of specific effects included in their panel gravity model. In sum, many studies have distinctly contributed to the evolution in the empirical analysis of RTAs effects on international trade through the development of the set of regional dummy variables: from single dummy to three dummies. According to our objective, this paper is reasonably in line with those studies including the set of three regional dummy variables individualised for each RTA, which is the most recent development in the set of RTAs dummy variables. As regards the empirical studies that include three regional dummy variables, most of them agree on trade creation effects regarding the intra-bloc trade following the creation of RTAs. Nonetheless, they split over the RTAs impacts on extra-bloc trade. Soloaga and Winters (2001) and Carrère (2006) show trade diversion effects in terms of bloc exports and imports for most RTAs, whereas Trotignon (2010) finds trade creation effects regarding the extra-bloc trade for a majority of RTAs. These mixed results stem mostly from differences between these studies regarding the studying period, the sample of countries as well as the choice of explanatory variables and estimation techniques. 3. Methodology and data Pursuing a study including the set of three RTA dummy variables, we will evaluate the impacts of seventeen RTAs in all over the world on intra-bloc trade and members trade with the rest of the world. Having created a data set involving 160 potential trading partners from , we want to estimate these RTAs effects by applying the PPML estimator both in the traditional gravity equation and in a gravity equation that takes into account the Anderson and van Wincoop s multilateral resistance terms. In Section 3.1, we present our econometric techniques and the fundamental problems in the gravity model. In Section 3.2, we introduce the gravity model. In Section 3.3, we describe our data set Econometric approach Studies on RTAs ex-post effects have lately encountered two main problems in the gravity model. The first one concerns zero trade flows between country pairs. On the one hand, some of the zero trade flows reflect a random rounding error or random missing data. They may also come 7

8 from the systematic rounding of very low reported values of bilateral trade. On the other hand, zero trade flows remaining in the database may naturally originate from the fact that bilateral trade does not exist over a period due to the remoteness of those countries, to the prohibitive transport costs or the small size of the economies, as argued by Frankel (1997), Santos Silva and Tenreyro (2006), and Helpman et al. (2008). Martin and Pham (2015) also found that most of the bilateral trade flows in aggregate trade data display a real absence of trade. The problem of zero trade flows is quite serious since almost 50% of the total observations on bilateral trade are zero in the data set used by Santos Silva and Tenreyro (2006), Helpman et al. (2008), and Burger et al. (2009). As a result, one need to take the problem of zero trade flows more seriously by using proper econometric techniques. The conventional method for estimating gravity model is to keep the model in log-linear form. However, this approach is inappropriate as the log-linearised model is infeasible in the case of observations involving zero trade flow because the natural logarithm of zero is undefined. Hence, several ways have been proposed in the empirical literature to handle the zero trade flows problem. One of the most prevalent ways is simply excluding zero trade from the data set and then estimating the gravity model on a truncated database of country pairs that consists of only positive bilateral trade flows. By omitting observations with zero trade, this method overlooks interesting and useful insight into the real nature of zero trade between countries and induces serious problems and biased results, since these zero trade flows are generally not randomly determined, as shown by Burger et al. (2009) and Martin and Pham (2015). Other studies choose to do not exclude zero trade flows, but use some transformation involving the dependent variable, for instance, adding a small number to the zero trade observation (value of 1 in most cases) before taking logarithms. Another method uses a tobit model and keeps the observations involving zero trade. Santos Silva and Tenreyro (2006) argued that these methods will induce inconsistent estimates in the case when the constant-elasticity model is used. They also pointed out that the standard methods used to estimate gravity models can lead to misleading estimated coefficients in the presence of heteroskedasticity, which appears inherently in trade data. If the problem of heteroskedasticity rises in the multiplicative model, its transformation into log-linear form can lead to a more severe bias in the estimated elasticities. Hence, they do not recommend to estimate the gravity model based on a log-linearised version. According to Santos Silva and Tenreyro (2006), the PPML estimator is a natural method to solve the problem of zero trade flows. Especially, they found that the performance of the PPML estimator is not affected when the proportion of the dependent variable with zero trade is substantial. Since the gravity model is directly estimated from its multiplicative form, where the dependent variable is measured in levels, instead of linearising the model by using logarithms, the zero trade problem is well handled. Moreover, they found that the PPML method seems to yield more robust and consistent results than the other econometric techniques in the presence of heteroskedasticity. Several recent empirical analyses on gravity model have included PPML method and praised the estimator as one of the new workhorses to assess international trade, such as Westerlund and Wilhelmsson (2011), Anderson and Yotov (2012), Martin and Pham (2015). The second problem that many analyses on trade policies have encountered in the gravity model involves the issue of potential endogeneity of RTAs when there is potential reverse causality between RTAs and a higher level of bilateral trade between country pairs. According to the hypothesis of "natural trading partners" or "natural trading blocs" introduced by Krugman (1991), countries show a propensity to form RTAs with other partner countries where there are potentially higher trade volumes between them. Furthermore, there still are many unobserved 8

9 factors between country pairs that may increase bilateral trade and promote the establishment of an RTA concurrently. As a result, the estimated coefficients are likely biased since the RTA dummy variables featuring the existence of the trade agreement are potentially correlated with the error term in the gravity equation. A majority of empirical studies using cross-sectional data and including dummy variables for trade agreements do not take account of the issue of RTA endogeneity. In the past literature, Trefler (1993) and Lee and Swagel (1997) are the first works that attempt to adjust for the endogeneity of trade policies on a cross-section framework by using instrumental variables 7. By contrast, Magee (2003) recently finds that instrumental-variables approach does not appear efficient at adjusting the issue of endogeneity bias of the RTA dummy variable that has binary form, and it is hard to find instruments that are not likely correlated with the error term of the gravity equation. An alternative method of handling the potential endogeneity issue with RTAs is to estimate the gravity model including both bilateral fixed effects for country pairs and time-varying fixed effects for exporter and importer countries 8. According to Baier and Bergstrand (2007), this fixed effects specification can deal with the issue of RTA endogeneity bias because it is able to better deal with the unobserved heterogeneity among pairs of countries, which are one of the most important sources of the endogeneity problem related to RTAs. In addition, Head and Mayer (2013) found that due to lacking adequate instrumental variables, panel data method including country-pair fixed effects can control for part of potential RTA endogeneity bias. Filippini and Molini (2003) likewise used country-pairs fixed effects model and found that long-term data do not have endogeneity problem and produces unbiased results. With the help of the PPML estimator suggested by Santos Silva and Tenreyro (2006), our study will try to tackle these two significant problems in the gravity model Gravity methodology To estimate the effects of RTAs, we use the basic gravity equation that has usually been used in international trade analysis and then, we augment the model with our dummy variables for seventeen RTAs. Our gravity model determines the trade flow between an exporter country and an importer country on a bilateral basis by the following equation: where the variables are defined as: X i jt =β O (GDP it ) β 1 (GDP jt ) β 2 (DIS T i j ) β 3 e β 4(LANG i j ) e β 5(CONT IG i j ) e α intra(rt A_intra i jt ) e α X(RT A_X i jt ) e α M(RT A_M i jt ) ψ i jt (1) X i jt is the value of trade flow in terms of goods in current dollar values from exporter country i to importer country j at time t. GDP it and GDP jt are the current dollar value of the gross domestic product (GDP) in country i and country j, respectively at time t. The impact of these two variables on bilateral trade flows is expected to be positive. 7 Trefler (1993) and Lee and Swagel (1997) concluded that the impacts of trade liberalisation policies tend to be underestimated without considering instrumental variables. 8 Baier and Bergstrand (2007) includes bilateral fixed effects to control for the unobserved time-invariant variables among country pairs and time-varying fixed effects for both exporter and importer countries to control for the Anderson and van Wincoop s multilateral resistance terms. 9

10 DIS T i j is the distance measured in kilometres between country i and trading partner j. The impact of this variable on trade flows is expected to be negative. LANG i j is a binary variable which takes the value of 1 if i and j share a common language, and 0 otherwise. The effect of this dummy variable is expected to be positive, given that a common language between two trading partners could facilitate trade deals and thus, reduce trade costs. CONT IG i j is a binary variable which takes the value of 1 if i and j have a common land border, and 0 otherwise. The effect of sharing a common land border between two countries is likely to be positive on bilateral trade flows. RT A_intra i jt assumes the value of 1 if both trading partners i and j have joined the same RTA at time t, and 0 otherwise. This dummy variable tests intra-bloc trade. RT A_X i jt assumes the value of 1 if exporter country i belongs to an RTA in which importer country j do not participate at time t, and 0 otherwise. This dummy variable captures the impact of the bloc exports to the rest of the world. RT A_M i jt assumes the value of 1 if importer country j belongs to an RTA in which exporter country i do not participate at time t, and 0 otherwise. This dummy variable tests the impact of the bloc imports coming from the rest of the world. e is the natural logarithm base, and ψ i jt denotes error term. The traditional approach to estimating equation (1) in the literature is to transform it to linear model by taking logarithms, leading to the following equation: ln X i jt =β O + β 1 ln (GDP it ) + β 2 ln (GDP jt ) + β 3 ln (DIS T i j ) + β 4 (LANG i j ) + β 5 (CONT IG i j ) + α intra (RT A_intra i jt ) + α X (RT A_X i jt ) + α M (RT A_M i jt ) + ε i jt (2) where ε i jt (= ln ψ i jt ) is the error term of equation (1). The log-linearised model is only feasible whenever X i jt > 0. Thus, the log transformation struggles with observations involving X i jt = 0 because the natural logarithm of zero cannot be determined. As explained in the previous section, our study applies the PPML estimator to deal with the challenges which the log-linear gravity equation has failed to overcome. Thus, the PPML technique is used to estimate the following gravity model: X i jt = exp ( β O + β 1 ln (GDP it ) + β 2 ln (GDP jt ) + β 3 ln (DIS T i j ) + β 4 (LANG i j ) + β 5 (CONT IG i j ) + α intra (RT A_intra i jt ) + α X (RT A_X i jt ) + α M (RT A_M i jt ) + ε i jt ) (3) The equation (1) is only an augmented version of the traditional gravity model introduced by Tinbergen (1962) that lacks solid theoretical foundation. In the early 2000s, Anderson and van Wincoop (2003) concluded their remarkable work by introducing the multilateral resistance terms. According to the two authors, the three trade resistance factors in international trade are, therefore, the bilateral trade barriers, the exporter country s trade resistance towards all other destinations as well as the importer country s trade resistance towards all other trading partners. 10

11 To carry out an easier computational method for taking into account these multilateral resistance terms variables, Anderson and van Wincoop (2003) and Feenstra (2004) suggest the estimation of equation (1) by using time-variant fixed effects for both exporter and importer countries. This type of fixed effects can produce unbiased results concerning coefficients estimates of β 0, β 3, β 4, β 5, α intra, α X and α 9 M. To specify the theoretically-motivated gravity equation including the Anderson and van Wincoop s multilateral resistance terms and adapting to the PPML estimator, equation (3) may be written as follows: X i jt = exp ( β O + β 1 ln (GDP it ) + β 2 ln (GDP jt ) + β 3 ln (DIS T i j ) + β 4 (LANG i j ) + β 5 (CONT IG i j ) + α intra (RT A_intra i jt ) + α X (RT A_X i jt ) + α M (RT A_M i jt ) + λ it + λ jt + ε i jt ) (4) where λ it and λ jt are time-varying exporter and importer fixed effects, respectively. Note that equation (4) may be augmented with bilateral fixed effects λ i j to control for the unobserved factors within country pairs and to overcome the problem of potential endogeneity of RTAs. We now return to our prime variables of interest: the set of three RTA dummy variables. It allows us to assess a precise identification of RTAs trade effects introduced by Viner (1950) on their member countries and multilateral trading system. To capture the trade creation and trade diversion effects of a specific RTA, we need to examine the signs of the coefficients of these RTA variables, which are α intra, α X and α M, respectively. Consider that α intra > 0, which means that the formation of an RTA stimulates intra-bloc trade creation effects between member countries. In this case, there is additional trade induced by both member countries joining in the RTA. Precisely, the domestic production of member countries or the bloc imports coming from the rest of the world can be substituted with the increase in intra-bloc trade resulting from the formation of the RTA. Thus, the coefficients α X and α M will determine the trade creation and trade diversion effects for a specific RTA. We demonstrate our method of analysing the signs of RTA coefficients, inspired by Soloaga and Winters (2001), Carrère (2006) and Trotignon (2010), in Table 1 as follows. Table 1: Trade creation and Trade diversion effects based on sign of RTA coefficients Sign of RTA coefficients α intra α X α M Trade creation and trade diversion effects >0 >0 >0 Intra-bloc trade creation / Export creation / Import creation >0 >0 <0 Intra-bloc trade creation / Export creation / Import diversion (when α intra > α M ) Export creation / Import diversion (when α intra < α M ) >0 <0 >0 Intra-bloc trade creation / Export diversion / Import creation (when α intra > α X ) Export diversion / Import creation (when α intra < α X ) >0 <0 <0 Intra-bloc trade creation / Export diversion / Import diversion (when α intra > α X + α M ) Export diversion / Import diversion (when α intra < α X + α M ) Source: Trotignon (2010). In sum, when α intra > 0 combines with α X > 0 (α M > 0), it indicates trade creation in terms of bloc exports to the rest of the world and bloc imports from the rest of the world. α intra > 0, coupled with α X < 0 or α M < 0, displays trade diversion regarding bloc exports or bloc 9 β 1 and β 2 are no longer determined because exporter and importer fixed effects are time varying. 11

12 imports. The term "export creation/diversion" is used for illustrating higher/lower trade when the exporter country is a member of the RTA and the importer country is not, whereas "import creation/diversion" is used for increased/reduced trade when the importer country belongs to the RTA in which the exporter country do not participate. If α X and α M are both negative, we need to compare the value of α intra with the absolute value of the sum of α X and α M to examine whether the trade diversion regarding bloc exports and bloc imports can completely outweigh the intra-bloc trade creation (in the case when α intra < α X + α M ). Besides, studying the signs of RTA coefficients also helps us to assess welfare for non-members. For instance, when α intra > 0 combines with α X < 0, we could figure out a decrease in welfare for non-members through export diversion effect Data The model is estimated based on a data set including 160 countries over the period of 1960 through Appendix A enumerates the countries used in our study. These countries, on average, accounted for over 95% of total trade in the world over the period of 55 years. Nominal bilateral trade data are collected from the International Monetary Fund s (IMF) Direction of Trade Statistics. Nominal GDPs are from Head et al. (2010) and the World Bank s World Development Indicators. The set of control variables involving geographical and cultural characteristics, such as bilateral distance, contiguity, common language, are sourced from the CEPII gravity database. Dummy variables for RTAs are created from the WTO Regional Trade Agreements Information System (RTA-IS) 10 complemented with the database of Head et al. (2010), and Baier and Bergstrand (2007). In our paper, we include only full (no partial) RTAs covering liberalisation of trade in goods that are notified to the GATT/WTO under GATT Article XXIV or the Enabling Clause for developing countries, which are free trade agreements and customs unions. The date when a given RTA enters into force is used to define whether the dummies for this RTA will take the value of 1 or 0. To capture impacts of the wave of regionalism on the multilateral trading system around the world, we consider seventeen RTAs that exist in different regions. Many of them were either created or revamped during the late 1980 and early 1990s, such as the ASEAN Free Trade Agreement, the NAFTA, the MERCOSUR, the ANDEAN Community, the Central American Common Market (CACM). During the 1990s and 2000s, we also witnessed the great extension of the EU along with the reduction in membership of the EFTA, and the establishment of other RTAs located mainly in Africa, Asia and Central Africa. The lists of all RTAs and countries included in our study are provided in Appendix B. Since this paper takes into account uni-directional trade flows as suggested by Baldwin and Taglioni (2006) rather than the average of the two-way exports, our data set presents a panel structure consisting of a total of 1,399,200 ((160 x 159) x 55) potential annual observations for 25,440 pairs of countries. Once the missing values are taken out, our sample covers thus 1,136,548 observations. Comparing to other empirical studies involving the assessment of RTAs effects on international trade, our work has a fairly large sample. Also based on the IMF s Direction of Trade Statistics, Frankel (1997) pooled data from 1970 through 1992 with five-year intervals and examined a total of 6,102 observations; Baier and Bergstrand (2007) has a sample of 47,081 observations covering 96 countries from 1960 to 2000 at five-year intervals. Carrère (2006) 10 rtais.wto.org/ui/publicmaintainrtahome.aspx 12

13 assesses RTAs impacts with a sample comprising 240,691 observations over the period and trade data sourced from UN COMTRADE. We describe the descriptive statistics of variables in Table 2 as below. Table 2: Summary statistics Variables N Mean Standard Deviation Min Max X i jt 1,377, GDP it 1,262, GDP jt 1,262, lngdp it 1,262, lngdp jt 1,262, DIST i j 1,399,200 7,789 4, ,650 lndist i j 1,399, CONTIG i j 1,399, LANG i j 1,399, * RTA_intra i jt 1,399, * RTA_X i jt 1,399, * RTA_M i jt 1,399, * RTA_intra i jt, RTA_X i jt and RTA_M i jt represent the total observations involving seventeen selected RTAs for intra-bloc trade flows, bloc exports and bloc imports related to the rest of the world. Table 2 suggests that an exporter country trades with an importer country an average of approximately million in current US dollars over 55 years from 1960 to The highest value of export flows in the database corresponds to the exports from China to the United States in 2014 (466.8 billion in current US dollars). As regards the value of exporter and importer GDPs, over 100,000 observations have missing values due to either unreported GDP of some small countries in a period of time or obviously unrecorded GDP of countries prior to their independence. The average distance between pairs of countries is about 7,700 kilometres. The data set also shows that 15.2% country pairs share a common language and only 1.8% country pairs have a common land border. Of all the observations, 31,644 observations (about 2.3%) belong to an RTA included in our sample of seventeen RTAs, that corresponds to 1,640 country pairs (about 6.5%) on the country-pairs level. Among these seventeen RTAs, the EU involves the most member countries with 27 countries included and covers 11,910 observations over a time span of 55 years, whereas the Australia-New Zealand Closer Economic Relations Trade Agreement (ANZCERTA) involves only two country pairs in terms of uni-directional trade flows and covers the least observations (64 observations). As regards the issue of zero trade flows in our data set, approximately 50.5% of the observations are zero (X i jt = 0) 11. This proportion of zero trade is similar to other empirical studies. For instance, 47.6% of the observations in Santos Silva and Tenreyro (2006), and about half of the observations in Helpman et al. (2008) and Burger et al. (2009) involve zero trade flows. Table 3 features the patterns of zero trade flows in the data set based on a set of bilateral distance and sets of exporter and importer GDPs. We find that smaller countries tend to export to a much smaller number of partner countries than others since the percentage of zero trade flows are higher in the set of the 1 st to the 33 th percentile of exporter GDP and importer GDP 11 Note that in the case of missing values in trade flows between exporter and importer countries for over ten consecutive years, we consider them as zero trade flows. 13

14 (72.79% and 66.98%, respectively) than one in the set of the 66 th to the 99 th percentile which corresponds to countries having greater GDP. In addition, countries are more likely to export to partner countries located not far away because the percentage of zero trade flows increases with bilateral distance. The findings from our data set are in line with the literature as bilateral trade is likely absent among small and remote countries due to prohibitive trade costs. Table 3: Percentage of zero trade flows in the total number of possible export flows Distance Exporter GDP Importer GDP 1st to 33th percentile 43.04% 72.79% 66.98% 34th to 66th percentile 51.28% 47.33% 49.06% 67th to 99th percentile 57.32% 36.69% 39.70% Source: Author s calculation. Figure 1a as follows shows a histogram and a kernel density plot for the proportion of zero in the exports of 160 countries included in our study. Among them, 18 countries have total zero trade flows under 15% of their potential export flows with trading partners from 1960 through All of these countries are developed countries. Nonetheless, the majority of countries have zero trade flows in terms of exports with around 40% to 70% of their potential partner countries. Number of countries Percentage of zero trade flows Note: The kernel density estimate is plotted using the Epanechnikov kernel with bandwidth Number of country-pairs Percentage of zero trade flows Note: The kernel density estimate is plotted using the Epanechnikov kernel with bandwidth (a) On country-level export (b) On country-pairs-level export Figure 1: Proportion of zero trade over the period On the country-pairs level, Figure 1b presents a histogram and a kernel density plot for the proportion of zero exports involving 25,400 country pairs. There are 3,975 country pairs completely having no zero trade flows over 55 years, and about 49% of the total country pairs have zero trade flows between 60% to 100%. In particular, we find that 2,399 country pairs, most of which involve small countries and remote from each other, have zero trade entirely during the studying period. Also, Figure 1b shows that zero trade flows are nonrandomly distributed, as expected from trade theory. In sum, the data set used in this paper suggests again that the issue of zero trade flows is quite crucial with 50.5% of the observations having zero trade flows. Thus, it justifies the need to handle the zero trade problem in order to gather valuable information 14

15 contained in zero trade data. 4. Empirical results We present the average impacts of each RTA over the period of 1960 through 2014 based on both the traditional gravity model and the gravity model including strong theoretical foundations, which controls for the multilateral resistance terms. We first carry out some preliminary tests to check for the presence of specific effects and heteroskedasticity in the gravity model. Breusch and Pagan Lagrangian multiplier test rejects the null hypothesis of the absence of specific effects in the log-linearised gravity equation (equation (2)) estimated by Ordinary Least Squares (OLS). Thus, specific effects exist in our gravity model and need to be controlled for. We look for the heteroskedasticity in our data set by using White test. The null hypothesis of the presence of homoskedasticity is rejected. As a result, the problem of heteroskedasticity also should be controlled for in the estimation. As mentioned earlier in Section 3.2, the λ it and λ jt denoting country-year dummies for exporter and importer countries in equation (4) are capable of properly accounting for the possible bias introduced by the potential omission of the multilateral resistance terms for exporter and importer countries. Since the total trade resistance captured by the unobserved exporter and importer resistance terms has a diverging and evolving nature, the inclusion of time-varying country dummies is believed to successfully control for it, especially in our paper when the time span is indeed quite long. However, since our data set includes 160 countries and 55 years, we would require the inclusion of 17,600 dummies in the regressions concerning time-varying dummies for both exporter and importer countries. Hence, this issue makes the estimation computationally unfeasible for econometric software. It cannot deal with this large number of dummies needed for time-varying country fixed effects. To handle the multilateral resistance terms in our study, we consider an alternative specification suggested by Ruiz and Vilarrubia (2007) that includes time-varying country dummies defined as country-three-year period dummies or country-five-year period dummies for exporter and importer countries. These specifications help to reduce the number of dummy variables to less than one-third (with country-three-year period dummies) or one-fifth (with country-five-year period dummies) of that used in the original time-varying country fixed effects. These dummies will remain fixed in three-year or five-year intervals. As a result, instead of including time-varying country dummies for 55 consecutive years, we are taking into account only 18 periods of three years each (except the last one that captures four years from 2010 to 2014) or 11 periods of five years each. Although time-varying country dummies defined as country-three-year period dummies or country-five-year period dummies do not precisely capture the effects of exporter and importer time-varying resistance terms, we believe that the bias yielded from estimations accounting for these dummies will probably be smaller than that in estimations excluding the issue of the multilateral resistance terms. Table 4 provides the empirical findings resulting from both of the estimation of equation (3) that features the traditional gravity model and the estimation of equation (4) that includes the multilateral resistance terms by using the PPML technique. Column (1) lists the estimated coefficients for the traditional gravity equation without any fixed effects. Column (2) and (3) provide coefficient estimates resulting from the PPML method for the Anderson and van Wincoop gravity equation with country-five-year period and country-three-year period fixed 15

16 effects, respectively. Column (4) presents the estimation outcomes resulting from panel data method, which we will discuss later in the next section. Table 4: Average effects of RTAs on international trade in terms of trade creation and trade diversion over the period of Variables The traditional gravity equation by (1) PPML without fixed effects (2) PPML with country-five-year period fixed effects The Anderson and van Wincoop gravity equation (3) PPML with country-three-year period fixed effects (4) Panel data with country-three-year period and bilateral fixed effects lngdp it 0.813** (0.0145) 0.435** (0.023) 0.496** (0.019) 0.384** (0.0155) lngdp jt 0.763** (0.0123) 0.521** (0.038) 0.520** (0.028) 0.552** (0.0150) lndist i j ** (0.0325) ** (0.0321) ** (0.0321) CONTIG i j 0.500** (0.0826) 0.371** (0.0640) 0.370** (0.0641) LANG i j 0.459** (0.0720) 0.276** (0.0547) 0.276** (0.0547) ANDEAN_intra i jt (0.256) 0.593** (0.213) 0.591** (0.213) 1.435** (0.211) ANDEAN_X i jt ** (0.186) (0.0525) (0.0318) (0.0797) ANDEAN_M i jt ** (0.0970) (0.0429) * (0.0459) * (0.0757) ANZCERTA_intra i jt 0.901** (0.252) 1.097** (0.179) 1.097** (0.179) 0.384** (0.110) ANZCERTA_X i jt (0.237) ** (0.0412) ** (0.0433) * (0.0721) ANZCERTA_M i jt (0.103) * (0.0523) * (0.0680) (0.0964) ASEAN_intra i jt (0.174) (0.135) (0.135) ** (0.143) ASEAN_X i jt 0.472** (0.0886) * (0.0389) (0.0321) 0.263** (0.0456) ASEAN_M i jt 0.367** (0.0978) (0.0337) (0.0293) (0.0608) CACM_intra i jt 1.071** (0.248) 1.369** (0.173) 1.370** (0.172) (0.131) CACM_X i jt * (0.160) (0.0541) (0.0440) (0.0759) CACM_M i jt * (0.141) (0.0466) * (0.0448) (0.0686) CAFTA-DR_intra i jt 0.644** (0.198) 0.815** (0.127) 0.824** (0.126) 0.599** (0.134) CAFTA-DR_X i jt ** (0.0889) (0.0321) (0.0217) (0.0528) CAFTA-DR_M i jt (0.116) ** (0.0320) ** (0.0253) * (0.0566) CARICOM_intra i jt 2.429** (0.345) 2.522** (0.258) 2.523** (0.257) 1.119** (0.204) CARICOM_X i jt ** (0.240) * (0.144) * (0.123) (0.0859) CARICOM_M i jt ** (0.116) (0.142) (0.160) (0.0635) CIS_intra i jt 1.722** (0.247) 2.035** (0.177) 2.035** (0.177) ** (0.162) CIS_X i jt ** (0.118) (0.172) 0.228** (0.0562) 0.297** (0.0692) CIS_M i jt ** (0.106) (0.170) 0.170** (0.0624) 0.298** (0.0674) COMESA_intra i jt 1.144** (0.197) 1.187** (0.219) 1.187** (0.220) 0.902** (0.118) COMESA_X i jt ** (0.144) (0.0881) (0.0309) (0.0591) COMESA_M i jt ** (0.0909) (0.0528) (0.0360) 0.110* (0.0525) EAC_intra i jt 1.728** (0.383) 1.874** (0.386) 1.872** (0.387) 0.881** (0.262) EAC_X i jt ** (0.184) * (0.189) (0.0791) (0.0847) EAC_M i jt * (0.146) (0.119) (0.0690) 0.206** (0.0797) EFTA_intra i jt 0.718** (0.213) 0.336* (0.143) 0.337* (0.144) (0.0886) EFTA_X i jt ** (0.122) ** (0.0250) ** (0.0282) ** (0.0457) EFTA_M i jt ** (0.105) ** (0.0269) ** (0.0393) ** (0.0554) EU_intra i jt 0.501** (0.0915) 0.329** (0.0768) 0.329** (0.0768) 0.908** (0.0392) EU_X i jt ** (0.0862) * (0.0467) * (0.0447) ** (0.0236) EU_M i jt ** (0.0701) ** (0.0460) ** (0.0451) ** (0.0294) MERCOSUR_intra i jt 0.973** (0.189) 1.209** (0.152) 1.212** (0.152) 0.503** (0.178) MERCOSUR_X i jt ** (0.124) (0.0453) (0.0544) (0.0669) MERCOSUR_M i jt ** (0.105) 0.206** (0.0541) (0.0625) (0.0824) NAFTA_intra i jt (0.184) 0.740** (0.106) 0.744** (0.107) (0.123) NAFTA_X i jt ** (0.0981) (0.116) (0.102) (0.0457) NAFTA_M i jt (0.111) * (0.0860) * (0.0793) (0.0546) PAFTA_intra i jt ** (0.209) ** (0.226) ** (0.226) 0.741** (0.107) PAFTA_X i jt (0.121) ** (0.0361) ** (0.0335) (0.0466) PAFTA_M i jt ** (0.0865) ** (0.0313) 0.157** (0.0321) (0.0396) SADC_intra i jt 1.545** (0.263) 1.755** (0.236) 1.749** (0.237) 0.788** (0.209) SADC_X i jt (0.192) (0.0437) (0.0630) (0.0577) SADC_M i jt (0.125) (0.0518) (0.0513) * (0.0565) SAFTA_intra i jt (0.468) (0.420) (0.421) ** (0.172) SAFTA_X i jt ** (0.148) (0.0460) (0.0318) (0.0322) SAFTA_M i jt ** (0.123) 0.210** (0.0726) 0.130* (0.0607) (0.0490) WAEMU_intra i jt 2.484** (0.315) 2.695** (0.280) 2.692** (0.280) 0.682** (0.200) WAEMU_X i jt ** (0.154) 2.997** (0.781) * (0.0664) (0.0731) Continued on next page 16

17 Variables Table 4 Continued from previous page The traditional gravity The Anderson and van Wincoop gravity equation equation by (1) PPML without fixed effects (2) PPML with country-five-year period fixed effects (3) PPML with country-three-year period fixed effects (4) Panel data with country-three-year period and bilateral fixed effects WAEMU_M i jt ** (0.137) ** (0.418) (0.0699) (0.0652) Constant ** (0.506) ** (0.823) ** (1.525) ** (0.376) Observations 1,136,548 1,104,295 1,096, ,924 Country-pairs 25,400 25,400 25,400 22,847 R Within R Exporter-Time effects No Yes Yes Yes Importer-Time effects No Yes Yes Yes Country-pair effects No No No Yes ** and * denote statistical significance at the 1% and 5% levels, respectively. Robust standard errors are in parentheses. The dependent variable in PPML regressions with specifications (1), (2) and (3) is the export flows in level; the dependent variable in panel data regression (4) is the natural log of the export flows. R 2 from PPML regressions is defined as the squared of the correlation between the observed and fitted values of the dependent variable. Coefficient estimates for time-varying exporter and importer fixed effects, and bilateral fixed effects are not reported for reasons of brevity. At first glance, the PPML estimates with different kinds of specification in Column (1), (2) and (3) all reveal that the level of GDP of exporter and importer countries are highly statistically significant at 1% and have the expected positive sign as bilateral trade flows increase with the size of exporter and importer countries GDP. The coefficient for distance is negative and statistically significant at 1%; the estimated coefficients for contiguity and common language are also positive and highly significant in all PPML estimates as expected. Our primary interest in this study is to assess the impact of various RTAs on members trade. Thus, we mainly focus on dummy variables for RTAs. We interpret the results based on the framework featured in Table 1. The traditional gravity model without time and country fixed effects estimated by the PPML technique (Column 1) shows significant intra-bloc trade creation along with export diversion and import diversion for most of the RTAs, such as the EU, the EFTA, the MERCOSUR, the CACM, the Caribbean Community and Common Market (CARICOM), the Commonwealth of Independent States (CIS), the COMESA, the East African Community (EAC) and the West African Economic and Monetary Union (WAEMU). There is only the ASEAN that witnesses export creation and import creation. However, the estimated coefficient for the ASEAN intra-bloc trade is insignificantly positive. Meanwhile, only the Pan- Arab Free Trade Area (PAFTA) has a significantly negative coefficient for intra-bloc trade effects. Since this specification of gravity model without fixed effects ignores the recent developments in theoretical foundations of gravity model that involve the multilateral resistance terms, the results in Column (1) may suffer bias. Hence, they are presented for comparison purposes and a connection to past empirical literature only. Column (3) shows coefficient outcomes resulting from our preferred specification with country-three-year period fixed effects because it allows for more time-varying country dummies (18 periods) that could capture the evolution of the multilateral resistance terms over time more precisely than one with five-year intervals (11 periods). The results provided by the PPML methods in Column (2) and (3) present consistent impact of many trading blocs on their intra-bloc trade that is an increase in trade between member countries following the formation of RTAs. The intra-bloc trade of several RTAs is significantly greater above the expected levels that one would predict from the gravity model, with the exception of the ASEAN, the South Asian Free Trade Agreement (SAFTA) and the PAFTA. They are insignificantly positive effects for the ASEAN and the SAFTA intra-bloc trade and 17

18 counter-intuitive significantly negative effects for the PAFTA intra-bloc trade. It may seem surprising that the coefficient for the intra-bloc trade is negative for the PAFTA since the intra-bloc trade tends to increase more than predicted by the gravity model following the formation of an RTA. Soloaga and Winters (2001), Carrère (2006) and Tumbarello (2007) also find a negative sign in the coefficient for intra-bloc trade of some RTAs. In our case, the negative coefficient for the PAFTA intra-bloc trade could be explained by the weak transportation networks between member countries that hinder the effort to promote the intra-bloc trade. In addition, several countries participating in the PAFTA are member countries of the Organisation of the Petroleum Exporting Countries (OPEC) like Saudi Arabia, Kuwait, Qatar and the United Arab Emirates. Consequently, these countries are likely to induce complex impacts on trade patterns of the PAFTA in general through their petroleum export policies over the years. Especially, there are two oil shocks in 1973 and 1979, which hit global economy hard, involve some PAFTA members (OPEC nations). Parra Robles et al. (2012) found an negligible impact of the PAFTA on intrabloc trade based on panel data method. On the other hand, Frankel (1997) indicates that the intra-bloc trade coefficient tends to become smaller when RTA openness variable is added to the model. In our study, we add two extra dummies for bloc exports and imports; consequently, we believe that the inclusion of those dummies could reduce the absolute value of the effects of a particular RTA on its intra-bloc trade like in the case of the PAFTA. Turn to the assessment of trade creation and trade diversion effects of RTAs in terms of the trading bloc exports and imports, the coefficient estimates for extra-bloc trade are quite sensitive to the specification of fixed effects. A useful approach to analysing the coefficients outcomes is to group the RTAs based on whether they are located in same continents, sub-regions or they have same levels of development. As regards the RTAs established by developed countries, both the PPML estimators with country-three-year period fixed effects and with country-five-year period fixed effects produce plausible common patterns of extra-bloc trade for the EU, the NAFTA, the EFTA and the ANZCERTA. All of these trading blocs witness export diversion and import diversion by means of significantly negative coefficients of bloc exports and imports. For instance, as indicated in Column (3), the intra-eu trade is 38.96% (= 100*(e )) above expected levels predicted by the gravity variables along with a propensity to export to non-members and to import from non-members lower by 9.87% and 13.93%, respectively. Our findings are in line with results from Soloaga and Winters (2001), Frankel (1997) and Sapir (1998) as the regional integration process within the EU members impacted negatively both on EU imports from European non-members and the rest of the world in general as well as on EU exports to the outsiders. We conclude that the EU shows trade diversion effects in terms of exports and imports. Accordingly, the regional integration of EU members imposed costs on the outsiders. Especially, our findings cover all enlargement processes of the EU from EU-9 countries to EU-27 countries. Nonetheless, our results of the effects on EU extra-bloc trade is quite different with the findings from Carrère (2006) and Trotignon (2010) that provide export creation and import creation for the EU. In the case of the EFTA, we reach the same conclusion with Soloaga and Winters (2001) concerning export and import diversions for the trading bloc. In addition, the negative propensity to import from the rest of the world of the NAFTA is in line with Carrère (2006) but contradictory with Trotignon (2010). By contrast, we found export and import creation effects for the CIS in the Eastern Europe. In the wake of the dissolution of the Soviet Union in the early 1990s, these 18

19 trade patterns reflect the openness of the CIS members with their neighbours (outsiders of the CIS) that have stimulated strong trade liberalisation in terms of goods since the late 1990s, such as Turkey, members of the EU in the Western Europe, and China, Pakistan, India in South Asia. Turn to the RTAs formed by developing countries in Central and Latin America, the ANDEAN and the CACM witness average import diversion which is significant at the 5% level. Precisely, the ANDEAN and the CACM present a tendency to import from the rest of the world slightly decreasing by 8.75% and 10%, respectively. Carrère (2006) also found import diversion for the ANDEAN. By contrast, the MERCOSUR witnesses, on average, a propensity to import from the rest of the world insignificant in PPML estimate with country-three-year period fixed effects, but superior by 22.87% and highly significant at the 1% level in PPML estimate with country-five-year period fixed effects. The import creation for the MERCOSUR is also found by Soloaga and Winters (2001) and Trotignon (2010) but in contrast with Carrère (2006). Column (2) shows the CARICOM has the same trade patterns with the MERCOSUR in terms of import creation effects, whereas Column (3) confirms only significant export diversion effects at the 5% level. For the Dominican Republic-Central America-United States Free Trade Agreement (CAFTA- DR) that formed by a North partner (the United States) and South partners (smaller developing economies in Central American), there are significant average import diversion effects in a similar way to the impacts of the CACM on extra-bloc trade. We find strong trade ties between countries joining in this RTA-the first free trade agreement between the United States with a group of developing countries, although it entered into force not long ago (in 2006). Turn to the RTAs in Asia, as mentioned above, both of the ASEAN and the SAFTA have insignificantly positive coefficient estimates for intra-bloc trade. It reflects a long implementation period concerning trade liberalisation schedules for the two Asian RTAs 12. The ASEAN has the implementation of tariff concessions over 26 years (from 1992 to 2018), while the SAFTA has implementation period over ten years (from 2006 to 2016). It means that members of these two RTAs have gradually lowered their trade barriers regarding both tariff and non-tariff barriers for goods coming from other members. However, the slow decrease in trade barriers within the ASEAN and the SAFTA does not generate substantial impacts on their intra-bloc trade. Hence, the average effects of the ASEAN and the SAFTA over the period on intra-bloc trade are not significantly different from the expected levels predicted from the gravity model. In addition, the insignificant average effects of the ASEAN and the SAFTA on intra-bloc trade could be explained by the fact that trade patterns of their member countries are actively oriented towards trade with the rest of the world. They have large global markets for their potential exports from different sectors, such as agriculture, textiles and apparel industry, electronics industry, etc. As a result, member countries of the ASEAN and the SAFTA do not show a higher intrabloc trade propensity yet. The PPML estimate with country-three-year period fixed effects show negligible export creation for these two Asian RTAs, and significant import creation for the SAFTA. The PPML estimate with country-five-year period fixed effects presents a propensity to export to non-members increasing by 9.43% in the case of the ASEAN. For the RTAs in Africa, we study the effects of four African RTAs that represent the pillars of the African Economic Community (AEC): the COMESA, the Southern African Development 12 Each RTA is subject to different liberalisation procedure and schedules. In some RTAs, the liberalisation of intrabloc trade takes place upon the date of entry into force of member countries. In our study, this date is used to define whether the dummies for RTAs take the value of 1 or 0. More common for RTAs is a phased implementation of tariff concessions over a period. The WTO s data on RTAs determines the implementation period for a given RTA is the date of final implementation of tariff eliminations undertaken by the slowest liberalising member. 19

20 Community (SADC), the EAC in South East Africa, and the WAEMU in West Africa. Note that there is a complex network of RTAs in Africa with several overlapping trading blocs established by same trading partners, for instance, Tanzania has joined in all of the three RTAs in South East Africa (the COMESA, the SADC and the EAC), Madagascar has participated in the COMESA and the SADC. When bloc impacts of these RTAs are tested at the same time, we do not find any extrabloc trade effects being significantly different from zero in their cases because the outsiders of a given RTA probably participate in another RTA along with member countries of the former RTA. Thus, there are unclear average impacts of RTAs in South East Africa on their trade (exports and imports) with non-members over the period of 1960 through It is recommended that one does not test the effects of RTAs in East and Southern Africa simultaneously. In the case of the WAEMU, significant export diversion effects reflect strong trade ties between former French colonies in West Africa through a tendency to export to member countries, which is detrimental to non-members. In short, three main findings emerge from our study. First, intra-bloc trade creation effects are found for most RTAs. There are increases in trade between member countries in the wake of the establishment of several RTAs. Second, export and import diversion effects are significant in many RTAs, regardless of whether they are formed by developed countries or developing countries. However, there are more export and import creations resulted from the formation of RTAs between developing countries. Finally, RTAs located in America and Europe appear to have more significant effects on intra-bloc trade as well as on extra-bloc trade than Asian and African RTAs. The EFTA is the only RTA witnessing that the increase in intra-bloc trade is entirely offset by a lower propensity to export and import. Besides, magnitudes of coefficient estimates for RTAs dummies variables in our study are smaller than ones found in previous studies on RTAs. The estimated impacts of RTAs on international trade are likely to be different when zero trade flows are taken into account by using the PPML estimator. We also observe that the coefficient estimates for extra-bloc trade effects for the ANDEAN, the CACM, the COMESA vary between different kinds of fixed effects specification (between results in Column (2) and Column (3)), from insignificantly positive coefficients to insignificantly negative ones. It means that these extra-bloc effects are very sensitive to the inclusion of the timevarying country fixed effects with three-year intervals or five-year intervals. There are probably strong evidence of other unobserved factors (e.g., regional and political instability) that cannot be totally controlled for by the time-varying country fixed effects and have impacts on these fixed effects otherwise. 5. Robustness check In Section 5.1, we test the robustness of RTA dummies coefficient estimates resulting from the PPML estimator by comparing with ones from panel data technique with fixed effects. In Section 5.2, we try to assess the effects of RTAs based on the level of economic development of member countries (North or South countries) Test for the robustness of RTAs effects resulting from the PPML estimator In a panel context, the gravity model is usually transformed to log-linear form as expressed in equation 2. As suggested by Baier and Bergstrand (2007), bilateral fixed effects along with timevarying fixed effects for exporter and importer are included in the model to overcome the RTAs 20

21 endogeneity bias and to take account of the Anderson and van Wincoop s multilateral resistance terms at the same time. Since our data set consists of 160 countries, it is unfeasible to include 25,440 bilateral dummies in the estimations with PPML technique. Unfortunately, we cannot add bilateral fixed effects to control for unobserved country-pairs factors and time-varying country fixed effects simultaneously in the model estimated by PPML technique. Column (4) in Table 4 provides the estimation outcomes resulting from panel data method with country-three-year period and bilateral fixed effects. One important drawback of countrypairs fixed-effects approach is that we cannot obtain the coefficient estimates for time-invariant bilateral factors such as distance, common language and contiguity. Moreover, as regards the total number of observations, since log-linear model estimated by panel data method excludes zero trade flows observations, it can only deal with 652,924 positive and nonzero dependent variables, whereas the PPML estimator with country-three-year period fixed effects is capable of handling a total of 1,096,603 observations. Thus, panel data method is likely to give up useful insight about zero trade flows between country pairs that could influence the RTAs impacts on bloc trade and international trade. We focus now on the comparison between the estimated coefficients of RTAs effects resulting from the PPML estimator and those resulting from panel data method. Regarding the RTAs effects on trade between member countries, we found that intra-bloc trade creation effects for ten RTAs (e.g., the ANDEAN, the ANZCERTA, the CAFTA-DR, the CARICOM, the COMESA, the EAC, the EU, the MERCOSUR, the SADC, the WAEMU) discovered by using the PPML estimator are consistent with results from panel data method. The magnitudes of the intra-bloc trade coefficients from the PPML estimate for most of these RTAs are greater than those from panel data method, except for the ANDEAN and the EU. The coefficient for the intra-bloc trade creation changes from significant in PPML technique to insignificant in panel data estimate for the EFTA, the NAFTA, the CACM. Likewise, the coefficient for the effects on intra-bloc trade for the PAFTA varies from negative in PPML estimate to positive in panel data estimate, while the ones for the ASEAN, the CIS, the SAFTA suddenly vary from positive to negative. Thus, these RTAs seem to be sensitive to the difference between PPML and panel data techniques. Regarding the extra-bloc trade effects of RTAs, significant import diversion effects for the ANDEAN, the CAFTA-DR, the EFTA, the EU found in panel data estimate are consistent with the PPML estimate. The panel data technique found the same import diversion effects for the ANZCERTA, the CACM and the NAFTA that are in line with the PPML estimate, but negligible. In addition, both methods find significant export diversion effects for the EFTA and the EU. The CIS witnesses export and import creation effects from both estimator. However, effects on extrabloc trade for several RTAs are likely to be very sensitive to the treatment of zero trade with the PPML estimator and the inclusion of bilateral fixed effects in panel data technique. We find that the ASEAN members have a significant level of openness concerning trade with non-members in panel data estimate through the increase in their propensity to export to the rest of the world, while these effects are insignificant from PPML estimate. There are negligible export creation effects for the CARICOM, negligible import diversion effects for the PAFTA and the SAFTA from panel data estimate that contrast with our previous findings by using the PPML estimator. Coefficients for extra-bloc trade of some RTAs also have opposite sign from PPML and panel data estimates, but they are negligible in panel data method. Some RTAs like the ANZCERTA, the CACM, the NAFTA, the PAFTA and the WAEMU lose their significant impacts on extra-bloc trade when the model is estimated by panel data method. Column (4) shows that the EU is the only case that has consistent effects on both extra-bloc trade 21

22 and intra-bloc trade with the PPML estimator. Moreover, we cannot clearly observe the trade patterns of RTAs based on levels of economic development of member countries or their location in Column (4), which is different from the findings of the PPML estimator. In sum, several RTAs keep a consistency in the effects on intra-bloc trade and extra-bloc trade in both the PPML estimation and the panel data estimation. Nonetheless, some RTAs are quite sensitive to the treatment of zero trade with the PPML estimator as they have different effects on extra-bloc trade than those resulting from the panel data method. The findings from the PPML estimator with country-three-year period fixed effects are still quite robust because they reasonably show similar trade patterns for groups of RTAs that are located on same continents or have same levels of economic development RTAs effects depending on the level of development of member countries In this section, we try to assess RTAs effects based on member countries characteristics. Do the level of economic development affect the impacts of RTAs on members trade? Since the establishment of the WTO in 1995, more developing countries (South partners) have involved in the formation of RTAs. Since then, more North-South RTAs and South-South RTAs are created all over the world by means of a dynamic participation of developing countries in Asia, South America and Africa. 2% 13.9% 15.9% 25.7% 19.2% 23.3% Immediate implementation Under or equal to 5 years 11 to 15 years 16 to 20 years 6 to 10 years over 20 years Figure 2: Duration of the implementation period of RTAs included in the study On the other hand, as we mentioned in previous sections, different RTAs are subject to different strategy for trade and tariff liberalisation. In some RTAs, member countries choose trade and tariff liberalisation beginning at the entry into force of RTAs. In others, partner countries choose this schedule taking place by the end of the implementation of RTAs, which means 13 We also perform the PPML estimator with country-two-year period fixed effects. The findings from this specification are consistent with those from the PPML estimator with country-three-year period fixed effects. For the sake of brevity, we omit to present these results here. 22

23 the final implementation of tariff eliminations undertaken by the slowest liberalising member (Acharya, 2016). Phase-in period for trade and tariff liberalisation are prominently applied in many RTAs. For this analysis, we expand our RTAs sample up to 245 RTAs all over the world until Figure 2 shows the breakdown of RTAs based on the length of their implementation period. Of those 245 RTAs included in this study, 15.9% RTAs have the implementation periods taking place immediately at their entry into force, while implementation periods up to five years occur in 19.2% and 23,3% had implementation periods varying from six to ten years. Of the RTAs studied, 41.6% RTAs have the transition period of trade and tariff liberalisation exceeding ten years. Figure 3 provides detailed insights of implementation periods of the RTAs studied based on types of RTAs. RTAs established between partners with similar characteristics on development level seem to undertake shorter implementation period for trade and tariff liberalisation. Of the RTAs studied, 65.8% North-North RTAs and 69.5% South-South RTAs have their transition period up to 10 years. By contrast, RTAs witnessing an asymmetry of economic development characteristics between members like North-South RTAs tend to undertake longer implementation period. This period that exceeds ten years happened in 57.3% North-South RTAs. North-North RTAs North-South RTAs South-South RTAs 3.1% 1.9% 7.3% 17.1% 17.1% 39% 19.5% 17.7% 36.5% 7.3% 11.5% 24% 9.3% 26.9% 19.4% 24.1% 18.5% Immediate implementation Under or equal to 5 years 11 to 15 years 16 to 20 years 6 to 10 years over 20 years Figure 3: Breakdown of RTAs implementation periods based on RTAs types In order to test the impacts of RTAs on trade during their implementation periods, we take into account one lagged level (five years after the date of entry in force of RTAs) and two lagged level (ten years after the date of entry in force of RTAs). We modify our gravity model in section 3.2 to include only RTAs dummies for intra-bloc trade and introduce new dummies to distinguish RTAs based on different levels of development of member countries. Consequently, the gravity 23

24 model employed in this analysis is described as follows: X i jt =β O (GDP it ) β 1 (GDP jt ) β 2 (DIS T i j ) β 3 e β 4(LANG i j ) e β 5(CONT IG i j ) e β 6(COMCOL i j ) e β 7(COLONY i j ) e α 1(North NorthRT A i jt ) e α 2(North NorthRT A i j,t k ) e δ 1(North S outhrt A i jt ) e δ 2(North S outhrt A i j,t k ) e θ 1(S outh S outhrt A i jt ) e θ 2(S outh S outhrt A i j,t k ) ψ i jt (5) where RTAs dummies are defined as: North NorthRT A i jt = 1 if both trading partners are developed countries and have joined the same RTA at time t, and 0 otherwise. North S outhrt A i jt = 1 if i and j join the same RTA at time t and i is developing country and j is developed country or vice versa, and 0 otherwise. S outh S outhrt A i jt = 1 if both trading partners are developing countries and have joined the same RTA at time t, and 0 otherwise. k = 1 if one lagged level is included or k = 2 if two lagged levels are included. The modification of our gravity model leads to the following theoretically-motivated gravity equation estimated by PPML estimator: X i jt = exp ( β O + β 1 ln (GDP it ) + β 2 ln (GDP jt ) + β 3 ln (DIS T i j ) + β 4 (LANG i j ) + β 5 (CONT IG i j ) + β 6 (COMCOL i j ) + β 7 (COLONY i j ) + α 1 (North NorthRT A i jt ) + α 2 (North NorthRT A i j,t k ) + δ 1 (North S outhrt A i jt ) + δ 2 (North S outhrt A i j,t k ) + θ 1 (S outh S outhrt A i jt ) + θ 2 (S outh S outhrt A i j,t k ) + λ it + λ jt + λ i j + ε i jt ) where λ it and λ jt are time-varying exporter and importer fixed effects, λ i j is country-pair fixed effects. Table 5 provides the empirical results concerning the effects of RTAs on member countries trade during the implementation period. Types of RTAs are based on the difference between members in the level of development. Our analysis also covers 160 countries spanning from 1960 through 2015 with five-year intervals. Column (1), (2) and (3) present the gravity estimates from equation 6 and based on the criteria of IMF for categorising North countries and South countries 14, while Column (4), (5) and (6) present the outcomes based on the criteria of World Bank for classifying developed and developing countries 15, for comparison purposes. For this analysis, we take account of three econometric techniques: Column (1) and (4) provide results from PPML estimator with country-year and bilateral fixed effects; Column (2) and (5) present results from panel data method with country-year and bilateral fixed effects; Column (3) and (6) report results from first difference technique with country-year fixed effects. (6) 14 According to the IMF World Economic Outlook, developed countries included in this study are: Belgium, Germany, Italy, United Kingdom, Austria, France, Netherlands, Luxembourg, Denmark, Ireland, Finland, Sweden, Spain, Greece, Portugal, Canada, Japan, Norway, Switzerland, Iceland, United States, Australia, New Zealand, Israel (1996-), Hong Kong (1996-), Republic of Korea (1996-), Singapore (1996-), Cyprus (2000), Slovenia (2006-), Malta (2007-), Slovak Republic (2008-), Czech Republic (2008-), Estonia (2010-), Latvia (2013-), Lithuania (2014-), Macao (2015) 15 According to World Bank World Economic Situation and Prospects, developed countries included in this study are: EEu-27, Iceland, Norway, Switzerland, Australia, Canada, Japan, New Zealand, United States 24

25 Table 5: Gravity estimates with two lagged levels of the RTA dummies over the period of Variables (1) (2) (3) (4) (5) (6) NorthNorthRTA i jt ** 0.223*** 0.133*** 0.327*** 0.440*** (0.0952) (0.0463) (0.0514) (0.0475) (0.0519) (0.0555) NorthNorthRTA i jt *** 0.125*** 0.105*** 0.228*** 0.215*** (0.0614) (0.0413) (0.0364) (0.0337) (0.0472) (0.0461) NorthNorthRTA i jt *** 0.114** * 0.254*** (0.0561) (0.0461) (0.0485) (0.0466) (0.0511) (0.0620) NorthSouthRTA i jt (0.0592) (0.0370) (0.0410) (0.0626) (0.0357) (0.0406) NorthSouthRTA i jt * (0.0583) (0.0505) (0.0504) (0.0519) (0.0499) (0.0504) NorthSouthRTA i jt *** 0.125*** *** 0.169*** (0.0548) (0.0519) (0.0456) (0.0555) (0.0514) (0.0463) SouthSouthRTA i jt 0.171*** 0.278*** 0.138*** *** (0.0549) (0.0475) (0.0516) (0.0854) (0.0542) (0.0542) SouthSouthRTA i jt *** 0.119** ** 0.143** 0.101* (0.0482) (0.0526) (0.0516) (0.0691) (0.0564) (0.0546) SouthSouthRTA i jt *** *** (0.0737) (0.0547) (0.0519) (0.0764) (0.0574) (0.0548) Constant 15.96*** 15.97*** (0.0159) (0.0160) Observations 226, , , , , ,770 Total North North RTAs effects Total North South RTAs effects Total South South RTAs effects R Within R Exporter-Year Yes Yes Yes Yes Yes Yes effects Importer-Year Yes Yes Yes Yes Yes Yes effects Country-pair Yes Yes No Yes Yes No effects First difference No No Yes No No Yes ***, ** and * denote statistical significance at the 1%, 5% and 10% levels, respectively. Robust standard errors are in parentheses. The dependent variable in PPML regression (1) and (4) are the export flows in level; the dependent variable in panel data regression (2) and (5), and first difference regressiong (3) and (6) are the natural log of the export flows. Coefficient estimates for country-year fixed effects are not reported for reasons of brevity. We include in this estimation two lagged levels (five years and ten years lagged). At first glance, based on the IMF s criteria concerning North and South partners, panel data (Column (2)) and first difference method (Column (3)) provide many significant RTAs impacts, while the PPML estimator (Column (1)) shows only significantly positive effects for South-South RTAs intra-bloc trade at time t. It may seem that RTAs impacts on members trade are very sensitive to the treatment of zero trade with the PPML estimator. Column (2) and Column (3) both agree on significant and positive impacts of RTAs between North partners on intra-bloc trade at their entry into force and during their implementation period of trade and tariff liberalisation over five or ten years. These outcomes confirm that North-North RTAs have shorter transition periods since the level of economic development of their members is more symmetrical than other types of RTAs. As regards North-South RTAs, panel data method found negligible coefficients for RTAs at the entry into force of RTAs and after five years, but significantly positive impacts of these RTAs after ten years. First difference method, however, found significantly positive impacts of these RTAs after five and ten years. These results show that North-South RTAs tend to increase trade 25

26 between member countries after at least five years after their entry in force. Furthermore, these outcomes seem in line with the fact that longer implementation periods occur in most North- South RTAs. In terms of South-South RTAs, we find highly significant and positive RTAs effects at their entry into force, and after five and ten years from panel data method. First difference technique leads to significantly positive RTAs impacts at time t and after five years, but negligible impacts after ten years. Comparing to the results based on the World Bank s criteria concerning North and South partners, we notice some changes. All three methods (Column (1), (2) and (3)) show North- North RTAs increase trade between their members from their date of entry into force. North- North RTAs in PPML estimator and panel data method exhibit an additional impact after five and ten years, while they only show an addition impact after five year in first difference method (Column (6)). Lagged effects of North-South RTAs and South-South RTAs remain consistent with our earlier findings based on the criteria of IMF. However, PPML estimator found counterintuitive significantly negative effects for South-South RTAs. In sum, panel data and first difference method show more significant effects for RTAs than PPML estimator. Results in the country-year and bilateral fixed specification show that RTAs, regardless of the nature of RTAs that reflects the North-North trade relations, the South-South or the North-South trade ties, tend to increase intra-bloc trade between member from their date of entry into force. North-South RTAs seem to undertake longer time in order to result in the increase of members trade. North-North RTAs have higher total average effects after ten years than North-South RTAs and South-South RTAs. This means that RTAs involving North and South partners take more years to liberalise trade and tariff between members, consequently, the beneficial impacts of these RTAs take more time to happen. Coefficents for RTAs impacts at the entry into force and the two lagged levels do not show the evolution of these effects over time, which means RTAs do not necessarily increase their intra-bloc over the years after the date of their entry in force. To test for the strict exogeneity of RTAs, we take into account the suggestion of Baier and Bergstrand (2007). We include in the regression a future level of RTAs. As Baier and Bergstrand (2007) proposed, there is not potential reverse causality between trade changes and RTAs changes when the future level of RTAs (RTA i jt+1 ) is uncorrelated with the concurrent trade flows. Column (1), (2) and (3) in Table 6 provide results from regressions based on IMF s classification of North and South partners. Only first difference method confirms the strict exogeneity of all types of RTAs by showing the effects of RTA i jt+1 on trade flows are small and insignificantly different from zero. By contrast, PPML estimator and panel data technique provide a negative future level of RTAs for North-North RTAs and North-South RTAs. It means firms in these RTAs have the propensity to delay trade in anticipation of new trade deals between member countries or new phases of tariff and trade liberalisation. As regards South-South RTAs, both Column (1) and (2) present a positive future level of RTAs, which exhibits anticipation effects of RTAs when firms promote trade before the date of their entry into force. Table 6: Gravity estimates testing for potential reverse causality between trade and RTAs Variables (1) (2) (3) (4) (5) (6) NorthNorthRTA i jt * *** * ** (0.0376) (0.0481) (0.0412) (0.0372) (0.0481) (0.0412) NorthNorthRTA i jt *** 0.224*** *** 0.225*** (0.0975) (0.0498) (0.0512) (0.0974) (0.0499) (0.0512) NorthNorthRTA i jt *** 0.126*** *** 0.128*** (0.0612) (0.0413) (0.0363) (0.0610) (0.0412) (0.0363) Continued on next page 26

27 Table 6 Continued from previous page Variables (1) (2) (3) (4) (5) (6) NorthNorthRTA i jt *** 0.112** *** 0.115** (0.0560) (0.0463) (0.0487) (0.0560) (0.0464) (0.0487) NorthSouthRTA i jt * * ** ** (0.0337) (0.0341) (0.0370) (0.0333) (0.0342) (0.0370) NorthSouthRTA i jt (0.0482) (0.0401) (0.0405) (0.0478) (0.0401) (0.0405) NorthSouthRTA i jt * * (0.0583) (0.0505) (0.0504) (0.0580) (0.0505) (0.0504) NorthSouthRTA i jt *** 0.124*** *** 0.118*** (0.0548) (0.0521) (0.0457) (0.0544) (0.0522) (0.0459) SouthSouthRTA i jt ** 0.139*** * *** (0.0476) (0.0540) (0.0560) (0.0491) (0.0606) (0.0622) SouthSouthRTA i jt 0.105** 0.179*** 0.136*** * 0.203*** 0.162*** (0.0527) (0.0514) (0.0509) (0.0557) (0.0582) (0.0574) SouthSouthRTA i jt *** 0.117** *** 0.117** (0.0480) (0.0526) (0.0515) (0.0503) (0.0604) (0.0593) SouthSouthRTA i jt *** *** (0.0746) (0.0547) (0.0518) (0.0787) (0.0633) (0.0566) Constant 15.96*** 16.00*** (0.0164) (0.0161) Observations 226, , , , , ,770 Total North North RTAs effects Total North South RTAs effects Total South South RTAs effects R Within R Exporter-Year Yes Yes Yes Yes Yes Yes effects Importer-Year Yes Yes Yes Yes Yes Yes effects Country-pair Yes Yes No Yes Yes No effects First difference No No Yes No No Yes ***, ** and * denote statistical significance at the 1%, 5% and 10% levels, respectively. Robust standard errors are in parentheses. The dependent variable in PPML regression (1) and (4) are the export flows in level; the dependent variable in panel data regression (2) and (5), and first difference regressiong (3) and (6) are the natural log of the export flows. Coefficient estimates for country-year fixed effects are not reported for reasons of brevity. In terms of RTAs lagged levels impacts, the results including the future level of RTAs are consistent with previous regressions (Table 5). These effects are significantly positive in panel data and first difference methods. However, the total average effects of South-South RTAs are greater than ones of North-North RTAs since the coefficient of future level of South-South RTAs is positive, whereas the one of North-North RTAs is negative. In Column (4), (5) and (6) of Table 6, we try to block the effects of RTAs that involve African countries (African South-South RTAs). As stated in Section 4, African RTAs have overlapping members and do not result in much effects on members trade. Therefore, we exclude these RTAs from regressions. First difference method finds strict exogeneity of North-North RTAs and North-South RTAs, but a significant tendency of firms in South-South RTAs to delay trade in anticipation of new trade deals. PPML estimator found only strict exogeneity of South-South RTAs, whereas panel data technique found a significantly negative coefficient for the future level of RTAs in all three types of RTAs. As regards lagged levels impacts of RTAs, most coefficients for RTAs are consistent with previous findings. Both panel data and first difference methods confirm that total average effects of North-North RTAs are greater than ones of other types of RTAs. In short, the causality between trade changes and RTAs changes is also sensitive to the 27

28 choice of econometric technique. Panel data and first difference seem to have consistent results with each other. 6. Conclusions This paper revisits the ex-post effects of RTAs on member countries trade by using the PPML estimator and including the Anderson and van Wincoop s multilateral resistance terms. The results concerning the average effects of RTAs over the period of are quite sensitive to the specification for fixed effects of the PPML estimator. Overall, the PPML estimate with country-three-year period fixed effects reveals significant intra-bloc trade creation for most of the RTAs. The Asian RTAs like the ASEAN and the SAFTA seem to do not lead to significant impacts on their intra-bloc trade between members because these RTAs have a long implementation period concerning trade liberalisation procedures. We also found that the overlapping RTAs in East and Southern Africa do not result in significant effects on members trade when they are tested at the same time. There are compatible extra-bloc trade patterns for groups of RTAs based on their levels of economic development or their location. Export diversion and import diversion effects are found for RTAs formed by developed countries (the EU, the NAFTA, the EFTA), import diversion effects are observed for RTAs established by developing countries in America (the CACM, the ANDEAN). A majority of RTAs showed evidence of trade diversion effects in terms of bloc exports and imports, regardless of the nature of RTAs that reflects the North-North trade relations, the South-South or the North-South trade ties. Our results featuring the increase in intra-bloc trade and trade diversion effects in terms of extra-bloc trade for most of the RTAs seem to be in line with previous studies like Soloaga and Winters (2001) and Carrère (2006). It appears that the propensity of regional integration around the world has improved the performance of intra-bloc trade for many RTAs, nonetheless, it is detrimental to the rest of the world. Our findings appear plausible in the light of the upsurge in RTAs around the world over the past two decades and the failure of the Doha Round of the WTO which aims to a better multilateral trading system. Our robust analysis also finds the discrepancy in impacts on members trade between RTAs based on the level of development of their member countries and the length of their implementation periods. It suggests that RTAs formed by developed countries and by countries that have symmetry of economic development characteristics result in greater increase in trade between members during shorter transition periods of trade and tariff liberalisation. Some caveats are necessarily considered in our future research. This paper cannot assess the RTAs effects of seventeen RTAs with the PPML estimator based on gravity model including both time-varying country fixed effects and bilateral fixed effects. These specifications of fixed effects may allow us to not only take account of the multilateral resistance terms but also to overcome the RTAs endogeneity bias generated by unobserved time-invariant factors among country pairs. Our further studies on this question should focus more on the dynamic effects of RTAs that can vary over time and on the drivers of successful RTAs according to both member countries and RTAs characteristics. 28

29 Appendix A. List of the 160 countries in the sample Figure A.1: World map indicating selected countries Table A.1: List of the 160 selected countries Albania Dominica Lao Sao Tome and Principe Algeria Dominican Republic Latvia Saudi Arabia Angola Ecuador Lebanon Senegal Argentina Egypt Lithuania Seychelles Armenia, Republic of El Salvador Luxembourg Sierra Leone Australia Equatorial Guinea Macao Singapore Austria Estonia Macedonia Slovak Republic Azerbaijan, Republic Ethiopia Madagascar Slovenia of Bahamas, The Fiji Malawi Somalia Bahrain, Kingdom of Finland Malaysia South Africa Bangladesh France Mali Spain Barbados Gabon Malta Sri Lanka Belarus Gambia Mauritania St.Kitts and Nevis Belgium Georgia Mauritius St.Lucia Belize Germany Mexico St.Vincent and the Benin Ghana Mongolia Grenadines Sudan Bolivia Greece Morocco Suriname Bosnia and Grenada Mozambique Sweden Herzegovina Brazil Guatemala Myanmar Switzerland Brunei Darussalam Guinea Nepal Tajikistan Bulgaria Guinea-Bissau Netherlands Tanzania Burkina Faso Guyana New Zealand Thailand Burundi Haiti Nicaragua Togo Continued on next page 29