Unilateral versus Regional Integration: What Matters More for Economic Development?

Unilateral versus Regional Integration: What Matters More for Economic Development? (JEL Category: F16, F43, I47) (Keywords: regional trading arrangements, trade, development, institutions) February 2008 Steven Yamarik and Sucharita Ghosh California State University The University of Akron Department of Economics Department of Economics Long Beach, CA 90840 Akron, OH 44325-1908 phone: (562) 985-4634 phone: (330) 972-7549 e-mail: syamarik@csulb.edu e-mail: sghosh@uakron.edu Abstract In this paper, we test for the individual impacts of unilateral and regional integration on long-run economic development. We build a four-part empirical model that links bilateral trade and regional trade arrangement formations to the cross-country determination of per capita income levels. Using a deep determinants approach where per capita income depends upon integration, institutions and geography; we find that regional integration has a positive and robust effect on per capita income, while unilateral trade has a negative and fragile impact. The point estimates indicate that a 1 percent increase in the percentage of intra-bloc trade (relative total trade) raises the level of per capita income by 1.5 to 2.0 percent. We would like to thank Scott Baier, Jeff Bergstrand, Daniel Bernhofen, Jay Shimshack, Chih Ming Tan, and participants of workshops and seminars at IUPUI, Brandeis University, California State University at Long Beach, Clark University, Cleveland State University, Kent State University, and the University of Connecticut for their helpful suggestions and comments.

1 The issue is to know first whether there is an interest for individual countries to get into these (regional trading) agreements, what do they have to gain with these agreements and maybe what do they have to lose. Francois Bourguignon at Global Economic Prospects 2005 press briefing on November 16, 2004 1. Introduction During the past 50 years, trade liberalization has occurred along three fronts: multilateral, regional and unilateral. At the multilateral level, membership in the GATT / WTO has increased from 23 at its inception in 1948 to 151 in 2007 (WTO, 2007a). At the regional level, the number of regional trading agreements (RTA) has increased from 2 in 1960 to 380 in 2007 (WTO, 2007b). At the unilateral level, Wacziarg and Welch (2003) estimate that the percentage of nations having an open trade policy, as defined by Sachs and Warner (1995), increased from 16% in 1960 to 73% in 2000. Similarly, the world-wide trade share (trade volume/gdp) doubled from 24% to 50% during the same time period (WDI, 2006). In response, an empirical literature has emerged to test the link between trade liberalization and per capita income. One branch of this literature (e.g. Frankel and Romer, 1999 and Irwin and Terviö, 2002) examines the relationship between the trade share (openness) and per capita income. These researchers find that a higher trade share increases the level of real GDP per capita. A second branch (e.g. Edwards, 1998 and Lee, Ricci and Rigobon, 2004) investigates the connection between trade policy and income. They find that higher tariffs, import duties and other non-tariff barriers reduce per capita income. A third branch (e.g. Easterly and Levine, 2003 and Rodrik, Subramanian and Trebbi, 2004) examines the respective roles of integration, institutions and geography in determining per capita income. These authors find that trade has no significant impact on the level of real GDP per capita once institutions are included. 1 1 One exception is Dollar and Kraay (2003) who estimate a positive coefficient for the trade share in the presence of institutions.

2 However, most research in the trade and growth literature uses unilateral measures of integration. There are reasons to believe that regional integration can impact per capita income. For instance, Corden (1972) shows that the formation of a regional trading agreement (RTA) can shift production in favor of intra-regional trade which in turn can lead to greater economies of scale. Moreover, RTAs can erode market power of domestic firms in participating countries by encouraging competition within the bloc (Baldwin and Venables, 1995). Furthermore, RTAs can lead to technology spillovers either directly through greater intellectual property rights (IPR) enforcement or indirectly thorough greater trade flows. Although some of these regional effects may be captured by unilateral measures, it would nonetheless be important to see what role, if any; regional integration plays in the determination of per capita income. In this paper, we test for the individual impact of unilateral and regional integration on long-run economic development. We build an empirical model that links bilateral trade flows and RTA formations to the cross-country determination of per capita income levels. At the bilateral level, we estimate unilateral trade shares and RTA partnership pairs as a function of geographical, historical and political factors. These bilateral variables are summed up across trading partners to generate cross-country measures of unilateral and regional trade. We then include the two trade variables in a cross-country levels regression to disentangle the impact of unilateral and regional integration on per capita income. A few previous studies have attempted to estimate the link between regional integration and per capita income. de Melo et al. (1992, 1993) and Vamvakidis (1998) include a dummy variable for RTA membership into a cross-country growth regression. Under most specifications, the coefficient for the RTA dummy is insignificant. Vamvakidis (1999) includes an RTA dummy along with the unilateral trade share. He finds that an increase in the trade share leads to higher growth, while RTA membership has no impact. In contrast, Schiff and Wang

3 (2003) find that Mexico received significant total factor productivity gains from trade with its NAFTA-partner the US, but no gains from trade with other OECD countries. These measures of unilateral and regional integration are likely endogenous. Trefler (1993) and Grossman and Helpman (1995) argue that trade policy is not set in a vacuum but rather is endogenously determined by political and special interests. Empirical work by Dutt and Mitra (2002) find that geographical and political factors explain tariff rates across a sample of countries. Krugman (1991) contends that RTAs form among natural trading partners, which are geographically close to each other and remote from other potential partners. Using bilateral data, Magee (2003) and Baier and Bergstrand (2003, 2004) test the natural trading partners hypothesis by estimating the determinants of regional trade arrangements. They find that geographical, economic and political factors explain the formation of RTAs. Moreover, the impact of regional trade arrangements on per capita income may vary across agreements. As with unilateral trade, the amount of growth-enhancing productivity spillovers and technology transfers in RTAs is likely to be correlated with the volume of intra-bloc trade. For instance, one would expect Mexico to benefit more from NAFTA (with Canada and the U.S.) than from the Group of Three (with Columbia and Venezuela) since the U.S. accounts for around three-quarters of Mexican total trade. In our paper, we control for the endogeneity and magnitude of unilateral and regional integration by constructing a four-part empirical model. The first part of the model describes the bilateral formation of RTA partnership pairs. The second part specifies a gravity equation of bilateral trade based upon Anderson and van Wincoop (2003). The decision to form an RTA pair depends upon relative geography, historical and political factors; while the trade share depends upon geography, historical and unobserved fixed country effects. We estimate the first two parts of the model simultaneously using a Heckman two-step procedure to control for

4 endogeneity bias. The third part sums up the predicted bilateral trade shares and RTA pairs to construct cross-country instruments of unilateral and regional integration. We measure unilateral integration using the total trade share and regional integration as the percentage of intra-bloc trade (relative to total trade). The fourth part is a per capita income equation. We estimate the per capita income equation using two-stage least squares (2SLS) to control for the endogeneity of integration and institutions. We find that regional integration has a positive effect on per capita income, while unilateral integration has a limited and fragile impact. We begin with a Frankel and Romer (1999) specification where unilateral and regional trade is included with measures of geography. We estimate a positive and significant coefficient for unilateral and for regional trade. The significance of each trade variable, along with the low correlation between the two, suggests that the positive impact of regional integration operates outside of its effect on the unilateral trade share. We then use a deep determinants specification of Rodrik, Subramanian and Trebbi (2004) where per capita income depends upon integration, institutions and geography. By controlling for institutions, we find that regional integration continues to have a positive and robust effect, while unilateral trade has a negative and fragile impact. The point estimates imply that a 1 percent increase in the percentage of intra-bloc trade (relative to total trade) raises the level of per capita income by 1.5 to 2.0 percent. The remainder of the paper is organized as follows. In section 2, we discuss the potential dynamic benefits of regional integration. We then construct the empirical model in section 3 and present the estimation strategy in section 4. In section 5, we present the estimation results. We conclude in section 6.

5 2. Dynamic Benefits of Regional Integration Lawrence (1996) makes a distinction between shallow and deep types of integration. 2 Shallow integration or trade liberalization involves the lowering or removal of border barriers to trade like tariffs, quotas and other non-tariff barriers. Deep integration, on the other hand, is the removal of internal or behind-the-border barriers that discourage the efficient allocation of international production within the region. These internal barriers include customs procedures, regulation of domestic production, product standards that differ from international norms, regulation of inward investment, competition policy, protection of intellectual policy, and government procurement rules (Evans et al., 2004). Shallow integration is likely to generate static welfare gains. Unilateral trade liberalization will lower domestic prices of imported and imported-competing goods and thus help consumers. Moreover, unilateral trade liberalization will lead to a more efficient reallocation of resources across industries and thus benefit the economy. Regional trade liberalization, on the other hand, can have either a positive or negative impact on welfare depending upon the relative amounts of trade creation and diversion (Viner, 1950). Summers (1991); Frankel, Stein and Wei (1995) and others argue that the beneficial effects of RTAs depends positively upon the geographic proximity and trade dependence upon members (Natural Trade Partners). Deep integration can generate additional growth or dynamic gains. The establishment of investment rules and dispute resolution and the harmonization of tax policy, antitrust rules, product standards, immigration laws and regulation can increase the flows of portfolio investment, foreign direct investment (FDI), labor, and intermediate goods. In addition, FDI and 2 Hoekman (1998) contends that the two types of integration are not mutually exclusive in that you can have four types of RTAs: (i) shallow integration but no deep integration, (ii) no shallow integration but deep integration, (iii) shallow and deep integration, and (iv) no shallow nor deep integration.

6 competition rules can reduce monopoly powers of domestic firms. In a neoclassical growth model, an increase in investment and skilled labor will raise the transitional growth rate. 3 In an endogenous growth model, FDI can increase the rate of technology diffusion (Glass and Saggi, 1998); while intermediate goods trade can raise the rate of technology adoption (Coe and Helpman, 1995). Similarly, by reducing the market power of domestic firms tied to current production processes, technology adoption can be raised (Parente and Prescott, 2000). Regional trading arrangements contain elements of deep integration and thus may deliver dynamic benefits. In particular, RTAs possess five possible institutional features that are deep: (1) merchandise trade rules, (2) services liberalization, (3) investment rules and rights, (4) intellectual property rights, and (5) labor mobility. Table 1 provides a matrix that connects each institutional feature to its potential dynamic benefit(s). Merchandise Trade Rules RTAs contain merchandise trade rules that reduce trade costs of goods between member countries by removing tariffs and quotas, creating more efficient customs operations, and establishing standards and conformity in the marketplace (Global Economic Prospects, 2005a). The trade rules are mostly shallow integration and thus expected to lead to static effects through trade creation and diversion. However, there are potential dynamic gains from reducing trade costs. First, the reduction in trade costs could increase the size the product markets. Larger product markets could increase the flow of investment as firms inside the RTA leave some countries and go to others depending upon comparative advantage, location and agglomerations of each member; while firms outside the RTA locate production inside the bloc to take advantage of larger markets and to access the lower internal tariff rates. Second, larger product markets 3 A net inflow of labor will increase the transitional growth rate only if the increase in human capital outweighs the increase in total population (c.f. Mankiw, Romer and Weil, 1992).

7 could increase FDI and their knowledge spillovers. Third, merchandise trade rules can lower the price of imported capital and intermediate goods. Depending upon the amount of trade creation and diversion within this sector, there could be an increase in the flow of technology to member countries. Lastly, the increased competition in the product markets reduces monopoly powers of the domestic firms. By reducing the barriers to technology adoption, competition could increase access to the global pool of knowledge. Services Liberalization From 1995 to 2005, 32 services agreements have been signed by RTAs under the General Agreement on Trade in Services (WTO, 2006c). These services agreements contain provisions that allow firms in member countries to provide economic services to other members and more importantly to establish foreign branches or subsidiaries. In general, services liberalization can enlarge the number of competitors and carries fewer income losses than merchandise trade rules because the barriers to service trade are often prohibitive and not revenue gaining. As a result, countries are likely to gain from regional services liberalization since the trade diversion losses are small (Mattoo and Fink, 2002). As with merchandise trade, services liberalization will create larger service markets with economies of scale potential. The economies of scale potential are especially large in service industries such as international transport and communications. The increase in market size and economies of scale will raise the return on investment, especially FDI. With fewer restrictions to prevent entry, firms from developed nations will establish foreign subsidiaries in developing economies and bring new production processes and knowledge. Moreover, since many services are used as intermediate inputs, services liberalization can benefit all sectors of the economy by lowering the price, raising the quality and increasing the variety of these intermediate inputs. Moreover, greater competition in the services sector could increase the adoption of new

8 technologies as monopoly powers of domestic firms are broken down. Mattoo, Rathindran, and Subramanian (2006) estimate that full liberalization of the financial and telecommunications sectors generate a 1.5 percent increase in GDP growth ceteris paribus. Investment Rules and Rights RTAs often include new investment rules and investor rights. The new investment rules entail provisions that provide equal treatment of foreign and domestic investors (national treatment); bans on discrimination among investors from member countries (nondiscriminatory treatment); rights to invest in businesses in all sectors, except where expressly prohibited (preestablishment); bans on trade-related investment measures (TRIMs); and investor rights to take foreign governments to dispute resolution for violating the investment provisions (investor-state dispute resolution). As a consequence, these new rules and rights can facilitate access to markets by relaxing restrictions on market entry and providing investor protections. New investment rules and rights can raise investment flows and FDI knowledge spillovers. As with merchandise trade, the preferential treatment of investment inside the RTA leads to investment creation within the trading bloc and investment diversion from outside the bloc. Levy Yeyati, Stein and Daude (2005) find strong evidence of FDI creation but inconclusive evidence of FDI diversion for 13 FTAs. Intellectual Property Rights Regional trade arrangements can also include intellectual property rights (IPR) that go beyond those of the WTO. Under the WTO, intellectual property such as copyrights, trademarks, trade secrets and patents are covered by the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPs). Signed in December 1994, TRIPs established minimum levels of protection on intellectual property afforded by member governments. Under TRIPs, any future bilateral and multilateral agreements can only establish higher standards of intellectual

9 property rights, known as TRIPs-plus. The TRIPs-plus provisions, contained in many US and EU bilateral regional trade arrangements, extend the patent term for delays, require patent protection for plants and animals, limit the use of compulsory licenses, extend pharmaceutical patent protection, and provide arenas for dispute settlement (Global Economic Prospects, 2005b). The enforcement of TRIPs and TRIPs-plus would raise the expected return of the production of knowledge-based goods and services such as pharmaceuticals, chemicals, food additives and software. As a result, investment and technology transfer in those sectors would increase. The impact of IPR on technology however depends critically on how technology is diffused. If technology is acquired primarily through the details of the patents, then IPR will increase the rate of innovation. However, if technology is acquired primarily through imitation, then IPR will raise the cost of imitation and the rate of innovation would decrease. This question is even more pronounced in technologically-advanced intermediate inputs such as machinery, chemicals, software and so on. Empirically, Eaton and Kortum (1996) found that the value of patent rights is typically high in the OECD countries. Labor Mobility Regional trading agreements vary by the degree of labor mobility allowed between member nations. At the one end, the European Union (EU) and European Free Trade Association (EFTA) allow full labor mobility, with very limited exceptions. At the other end, the South Asian Association for Regional Cooperation (SAARC) and the Central European Free Trade Agreement (CEFTA) have no effective provisions for labor mobility. The regional agreements also tend to discriminate in favor of skilled workers. For example, several Latin American agreements (Group of Three, Brazil FTAs) provide for the movement of some skilled workers, but not for unskilled workers.

10 Regardless, labor mobility provisions have the potential to raise economic growth by increasing the inflow of labor and reducing barriers to technological adoption. First, depending upon the nature of growth and the skill set of immigrants; the inflow of labor could decrease the transitional growth rate, increase the transitional growth rate or increase the steady-state growth rate. Second, if the migration of labor increases the stock of human capital, then member countries will be more able to adopt new technologies. As a result, barriers to the adoption of the global pool of knowledge would fall. 3. Empirical Model We develop a four-part empirical model to estimate the effects of unilateral and regional integration on per capita income. The first part of the model develops a gravity equation that describes bilateral trade flows. The second part describes the endogenous formation of bilateral regional trade arrangements. The third part sums up the bilateral variables to create crosscountry measures of unilateral and regional integration. The fourth part estimates the impact of integration, institutions and geography on the level of per capita income across economies. Bilateral Gravity Equation The gravity equation has long been the workhorse model used to explain bilateral trade flows. The gravity equation can be derived from models of complete specialization and identical preferences (Anderson, 1979 and Anderson and van Wincoop, 2003); product differentiation with imperfect competition (Helpman, 1987); and incomplete specialization and trading costs (Haveman and Hummels, 2004). We use the Anderson and van Wincoop model of complete specialization, homothetic preferences and iceberg trade costs. Our intention here is to adopt a bilateral gravity equation that is grounded in trade theory, parsimonious in its specification, and easily summed up across trading partners.

11 Each country i produces a single good i but derives utility from all goods 1,, i, j, N. Each country i bears a trade cost c ij when exporting their good i so that country j pays p ij = p i (1+c ij ) for a good produced at p i. Assuming that the bilateral trade costs are symmetric, c ij =c ji, the equilibrium price level P in country j is P = S ( c / P) (1) 1 σ 1 σ j i ij i i where S = Y / Y is country i's share in World GDP and s is the constant elasticity of i i World substitution. We denote bilateral variables with lower-case letters and cross-country variables with upper-case letters. Anderson and van Wincoop refer to P as 'multilateral resistance' variables as they depend on all bilateral costs {c}, including those not directly involving i. Using (1), the following gravity equation emerges: t ij YY i j ij = Y c PP World i j 1 σ (2) where t is the bilateral trade flow. 4 Taking logs of each side of (2), we get t ij ln = (1 σ)ln cij (1 σ)ln Pi (1 σ)ln P. (3) j YS i j The left-hand side is the log of bilateral trade divided by the relative economic size of countries i and j. The right-hand side is the log of bilateral trade costs minus the log sum of the implicit price levels. Letting bilateral trade costs c depend stochastically upon geographic factors g (distance, common borders) and RTA membership, we can rewrite (3) as 4 Note that Anderson and van Wincoop use t to represent trade costs and x to represent trade flows.

12 t ij ln = (1 σ) β0 + (1 σ) β1rta + g (1 σ) ρ (1 σ)ln P (1 σ)ln P + (1 σ) u YS i j ij ij i j ij (4) where u is a random error and ( β0, β1, ρ) are coefficients to be estimated. Anderson and van Wincoop estimate a variant of equation (4) using a customized constrained minimization technique. Instead, Rose and van Wincoop (2001) and Feenstra (2004, Ch. 5) adopt a fixed individual country effects specification: t ij ln = µ + αrta + g λ + δ + δ + e YS i j ij ij i j ij (5) where d i is a fixed effect for country i, d j is a fixed effect for country j and e is the error term. 5 Feenstra (2002) shows that the estimation of (5) is easier to implement than (4) and yields similar results. Moreover, unlike paired fixed effects, individual fixed effects permit us to include timeinvariant variables like distance and common border. Bilateral RTA Formation A regional trade arrangement, like any trade policy, is a potentially endogenous variable. In the words of Lawrence (1998), free trade areas may well be an endogenous variable that is, a response to, rather than a source of, large trade flows (p. 59). For instance, Krugman (1991) shows that most RTAs tend to form among natural trading partners nations that are geographically close and alike. de Melo, Panagariya and Rodrik (1993) show how regional integration can arise from strategic interaction. Magee (2003) and Baier and Bergstrand (2003, 2004) estimate a qualitative choice model and find that countries are more likely to form an RTA (i) the closer in distance, (ii) the more remote from the rest of the world, (iii) the larger and more economically similar, and (iv) the bigger difference in factor endowments. 5 Eaton and Kortum (2002) and Baier and Bergstrand (2007) also use fixed individual country effects to account for multilateral price terms.

13 Following Baier and Bergstrand, we model the determination of an RTA using a qualitative choice model of McFadden (1975, 1976). Let u * represent an unobservable variable capturing the difference in utility levels from forming an RTA versus not forming an RTA: u = π + π g + π z + v (6) * ij 0 1 ij 2 ij ij where g are geography and historical variables, z are political measures and v is an error term (uncorrelated with g and z). With u * unobservable, we define an indicator variable rta, which is 1 if the two countries are in a regional trade arrangement (u * > 0) and 0 otherwise (u * = 0). The response probability for rta is Prta = g z = Pu > g z = Φ + g + z (7) * ( 1, ) ( 0, ) ( π π π ) ij ij 0 1 ij 2 ij where F is the standard normal cumulative distribution function. The standard errors for the estimates of ( π 0, π 1, π 2 ) are asymptotically standard normal. Aggregate Unilateral and Regional Trade We next examine the relationship between the bilateral trade and RTA variables and their aggregate or cross-country counterparts. The observed cross-country unilateral trade share is defined as the sum of aggregate exports and imports divided by GDP. For each country i, we take a weighted average of the bilateral trade share, t / YS, where the weights are the GDP ij i j share of country j, to arrive at the cross-country trade share: T / Y = S ( t / YS ). (8) i i j ij i j j i We use predetermined GDP shares S as weights to create instruments for the trade share.

14 We measure cross-country regional integration as a trade-weighted average of bilateral RTA memberships: RTA = ( t / T ) rta (9) i ij i ij j i where t ij is bilateral trade between countries i and j and T i is the aggregate trade of country i. As before, we use predetermined bilateral and cross-country trade shares as weights to create instruments for the regional integration variable. The RTA variable (9) has three important properties. First, the value of RTA increases with the amount of trade conducted inside each regional agreement. This property is consistent with the idea that the dynamic benefits of regional integration are carried by or at least correlated with the volume of intra-bloc trade. Second, there is no systematic relationship in the construction of the unilateral trade share (8) and the regional RTA variable (9). As a consequence, movements in the RTA variable will not simply be mimicking movements in the unilateral trade share. Third, the use of predetermined trade shares in conjunction with predicted values for rta will create exogenous instruments for the observed RTA variable. Cross-Country per Capita Income Equation The final piece of our empirical model is a cross-country per capita income equation. We adopt the deep-determinants approach of Rodrik, Subramanian and Trebbi (2004). 6 Under this approach, a distinction is made between the proximate (i.e. capital accumulation and education) and deep (i.e. geography and institutions) sources of economic growth. Rather than attempting to isolate the impact of each proximate source, a deep-determinants approach 6 Easterly and Levine (2003), Dollar and Kraay (2003) and Glaeser et al. (2004) also follow a deep determinants approach in their analysis.

15 specifies that the level of per capita income Y/POP depends upon integration TRADE, institutions I and geography G: ln( Y / pop ) = b + btrade + bi + bg + ε (10) i i 0 1 i 2 i 3 i i where e is an error term. In this paper, we measure TRADE as the unilateral trade share T/Y and regional integration RTA using (9). 4. Estimation Strategy and Identification We adopt a three-part estimation strategy. First, we estimate the gravity model (5) and bilateral RTA formation (7) equations using a Heckman two-step procedure. Second, using equations (8) and (9), we sum up the predicted bilateral trade shares and RTA variables to obtain our cross-country instruments Tˆ i and RTA i and actual membership RTA. Third, we estimate the cross-country output equation (10) using two-stage least squares (2SLS). Average Treatment Effect In the first part, we are estimating the effects of an endogenous binary variable ( rta ) on a continuous endogenous variable ( t / YS ). A basic problem in such a situation is that we do not ij i j observe the same individual (bilateral pair of nations) in both the treated (RTA membership) and untreated (no RTA membership) states at the same time. We instead observe one outcome for each country pair (i, j): ij ln( t / SY) = ( rta )ln( t / SY) + (1 rta )ln( t / SY) (11) 1 0 ij j i ij ij j i ij ij j i where the superscript refers to the outcome (1=RTA, 0=no RTA). Each trade share outcome is determined by a separate gravity model equation:

16 ln( t / SY) = µ + α rta + g λ + δ + δ + e = η + xβ + e (12) 0 0 0 0 0 0 0 0 0 0 ij j i ij ij i j ij ij ln( t / SY) = µ + α rta + g λ + δ + δ + e = η + xβ + e (13) 1 1 1 1 1 1 1 1 1 1 ij j i ij ij i j ij ij where η is the intercept term and x are a vector of geography and historical variables and country-specific dummies. In equations (12) and (13), we allow both the intercept and slope coefficients to vary across regimes. The average treatment effect (ATE) is defined as the expected difference of bilateral trade shares under the two outcomes: 1 0 ATE( tij / SY j i) E ln( tij / SY j i) ln( tij / SY j i) xij, zij, rta ij (14) Substituting in equations (12) - (14) into (11), we get that the observed trade t is 0 1 0 0 ln( t / SY) = η + E ( ) ij j i rta e e ATErta x β ij ij ij + + + ij ij 1 0 0 1 0 1 0 rtaij( xij x)( β β ) + eij + rtaij( eij eij) E rtaij( eij eij ). (15) where x are the demeaned values of x. Wooldridge (2002, Ch. 18) describes four IV procedures to consistently estimate (15). Each procedure is based upon the correlation between the treatment indicator rta with (i) the error term 0 e, (ii) the differences in how factors affect trade flows for partners with RTAs versus those without (x- x ), and (iii) differences in unobservables for partners with RTAs versus those without 1 0 ( e e ). The last correlation occurs due to selection bias. We choose a "Heckman" two-step procedure (Procedure 18.4 in Wooldridge) to estimate (15). The Heckman procedure allows the treatment indicator rta to be correlated with 0 e, (x- x ) and 1 0 ( e e ). In the first step, we run a probit on (7) to form the predicted probabilities ˆΦ and ˆ φ for all i,j. In the second step, we use OLS to estimate

17 0 0 1 0 ln( tij / SY j i) η ATE rtaij xijβ rtaij( xij x)( β β ) = + + + + ρ rta [ φˆ / Φˆ ] ρ (1 rta )[ φˆ /(1 Φ ˆ )] + error. (19) 0 ij ij ij 1 ij ij ij ij The terms [ φˆ / Φ ˆ ] and [ φˆ /(1 Φ ˆ )] are called the inverse of Mill's ratios or hazard rates and control for unobservable heterogeneity and selection bias. A test of joint significance of? 0 and? 1 can test for the endogeneity of rta. The Heckman two-step procedure does require that the error structure 1 0 (, ve, e ) is independent of (x,z) and trivariate normal in (6) and (19). We were willing to make this assumption because the IV estimates obtained under a less-restrictive alternative (Procedure 18.3) were very imprecise. The imprecision was a byproduct of using the predicted values ˆΦ as instrument for rta and all interaction terms. Moreover, Wooldridge (2002, Ch. 18) argues that the Heckman two-step procedure is likely to be more efficient than Procedure 18.3 because it is based on Etrta (, xz, ) as opposed to Etxz (, ). Cross-Country Unilateral and Regional Trade We next construct actual and predicted cross-country unilateral and regional trade variables. We use the ratio of nominal exports plus imports to nominal GDP to measure the actual trade share. For the predicted trade share, we use equation (8) and add up the predicted bilateral trade share tˆ / Y S from (5), weighed by the 1960 GDP shares S. Equation (9) is used to create the actual and predicted regional integration variables. For actual regional integration, we use the realized values of rta, t and T to construct RTA. For predicted regional integration, we add up the predicted probit values rta from (7), weighed by the 1970 trade shares t and T,

18 to construct RTA. We use the 1960 GDP share and 1970 trade share in the construction of the predicted values to maintain exogeneity of our instruments. 7 Two-Stage Least Squares We estimate the cross-country income equation (10) using 2SLS. There are three potential endogenous variables: T/Y, RTA and I. In the first stage, we run the following three regressions: T Y = c + cg + ct / Y i + crtai + cz+ ε (19) / i 0 1 i 2 3 4 i i RTA = d + dg + d T / Y i + drtai + dz + µ (20) i 0 1 i 2 3 4 i i I = e + eg + et / Y i + ertai + ez + ω (21) i 0 1 i 2 3 4 i i where G are cross-country geography measures and Z are instruments for institutions. The set of variables ( T / Y, RTA, Z) are the excluded instruments, which allow us to identify the three endogenous variables. In the second stage, we regress Y on the three fitted values of (19)-(21) and G. An instrumental variable must satisfy two requirements: it must be orthogonal to the error term (validity) and it must be correlated with the included endogenous variable (relevance). We use the Hansen J statistic to test for the orthogonality of the instruments when there are more excluded instruments than endogenous variables (overidentification). Relevance can be examined through the first-stage R-square and F-statistics. However, the recent literature on weak instruments (c.f. Stock, Wright and Yogo, 2002) has shown that mere instrument relevance is insufficient. In other words, rejection of the null of underidentification does not ensure reliable IV inference. 7 For those countries like the former Soviet republics, Czech Republic, Slovakia and Yemen that did not exist in 1970; we use the GDP share and trade data for the first available year.

19 With more than one endogenous variable, we use the Shea (1997) partial R-square and the Stock and Yogo (2002) weak instrument test to determine the strength of our instruments. The Shea partial R-square records the additional explanatory power of the excluded instruments. Unlike the conventional partial R-square, the Shea partial R-square takes the intercorrelations of the instruments into account. The Stock and Yogo weak instrument test compares the Cragg- Donald statistic to critical values based upon the worst possible case of weak instruments. The Cragg-Donald statistic is the minimum eigenvalue of the generalized F-statistic from the firststage, reduced form regression. Stock and Yogo (2002) present a testing procedure where the Cragg-Donald statistic is compared to pre-determined critical values under which the size of a nominal 5% test about a parameter of interest where actually r percent. Specification and Identification Table 2 summarizes the specification and identification scheme for the bilateral RTA (7), bilateral gravity (5), institutions (21), and per capita income (10) equations. There are four categories of variables: endogenous, geography, endowment and political. Each variable can be measured bilaterally (ij), cross-country (i) or both (ij) / (i). We choose our specifications to be parsimonious and consistent with earlier work. Furthermore, when possible we include crosscountry counterparts to any commonly-used bilateral variables and visa versa. For instance, we include the number of island nations in the bilateral equations and a dummy for an island nation in the per capita income equation. The bilateral RTA equation includes geography, historical and political variables, while the bilateral gravity equation includes geography and historical factors. The geography factors are log of bilateral distance; dummy for a common border; log sum of surface area and remoteness; and the number (0,1,2) of island nations and landlocked nations. The historical variables are dummies for a colonial link, common colonizer and common language. The political factors are

20 dummies for both free, current political alliance and correlation of voting in the United Nations. The political variables are excluded from the gravity equation to achieve identification. We consider three different specifications of the per capita income equation (10). First, we use the Frankel-Romer (FR) specification where Y/POP depends upon trade, log of population and log of surface area. The FR specification has been used by Irwin and Treviö (2002), Noguer and Siscart (2005), and Alcalá and Ciccone (2006) to test the link between unilateral openness and per capita income. Second, we consider a geography (GEO) specification where all the bilateral geography measures are summed up to form a set of cross-country geography variables G. The geography variables G are log of population, log of surface area, log of remoteness, and dummies for island and landlocked. Third, we use the Rodrik, Subramanian and Trebbi (RST) specification where per capita income depends upon integration measured by trade share; geography measured by distance from equator; and institutions measured by rule of law. We include our regional integration variable in each specification. The institutions and growth literature recognizes that growth may lead to better institutions. To correct for this endogeneity, researchers have proposed different instruments to identify institutions. In two important studies, Acemoglu, Johnson and Robinson (2001, 2002) argue that the actions of European colonizers shaped the institutions of their colonies. As a consequence, they use settler mortality rates or population density in 1500 as instruments for current institutions of former colonies. Mauro (1995) and La Porta et al. (1998) argue that language and legal origin play a critical role in determining institutions. In response, Hall and Jones (1999); Dollar and Kraay (2003); and Rodrik, Subramanian and Trebbi (2004) use European and English language fractionalization as instruments. Moreover, Easterly and Levine (2003) and Glaeser et al. (2004) also include legal origin in their instrument set.

21 We use European and English language fractionalization as instruments for institutions. First, unlike settler mortality rates, these instruments are available for a large sample of countries. Second, European and English language fractionalization are commonly-used and thus allow us to directly compare our results to other papers, especially Rodrik, Subramanian and Trebbi (2004). Third, diagnostic tests found that this instrument set was the most relevant and strong. We conduct robustness tests to alternative instrument sets later in the paper. Data We use bilateral and a cross-country data to conduct our empirical analysis. The bilateral data set is a series of five panels for 1980, 1985, 1990, 1995 and 2000. In each panel, there are 5,000 to 10,000 bilateral observations. The trade flow t is an average of imports plus exports for each country in the pair. The trade flow data is drawn from the IMF Direction of Trade. We then divide trade flow by YY i j / Y World to get the bilateral trade share (see equation 2). The geography and historical data are from Mayer and Zignago (2006) and the CIA World Factbook (2006). The political variables are from Polity IV (2002), Gibler and Sarkees (2004) and Gartz, Jo and Tucker (1999), respectively. The cross-country data set is a cross-section of countries. Appendix A presents the basic data. The dependent variable is the log of real GDP per capita in PPP-adjusted dollars in 2000. We measure trade share T as nominal openness (= nominal trade / nominal GDP) averaged for 1960-2000. The GDP, population and trade data are from PWT 6.2 in Heston et al. (2006). Institutions I are measured as the rule of law indicator of Kaufmann, Kraay and Mastruzzi (2005) averaged from 1995 to 2000. The geography variables are taken from the CIA Factbook (2006). European and English language fractionalization are from Hall and Jones (1999) with data for the former Soviet Republics obtained from Gordon (2005).

22 Unfortunately, there is no straightforward way to determine which RTAs are deep and which are not. First, the concept of depth is not easy to measure. There are a multitude of incongruent policies ranging from the elimination of contingent protect to the complete liberalization of factor flows that make up the depth of integration. For instance, Hoekman (1988) lists 9 specific policies for deep integration, while Evans et al. (2004) count 8 broad features. Second, deep integration is less about reducing trade barriers and more about the harmonization of regulatory practices in the trading bloc. As a result, it becomes quite difficult to talk about the degree of harmonization. Third, like shallow integration, there often can be a large disparity in the policies called for in the regional agreements and those actually implemented. We therefore use all RTAs that were in force from 1980 to 2000. Appendix B lists the regional trading arrangements. We include those RTAs notified to the GATT / WTO and also some that were not notified. The data were obtained from Medvedev (2006) and WTO (2007b). We did exclude those regional agreements that have no binding economic policies like the Arab Maghreb Union (AMU), Asia-Pacific Economic Cooperation (APEC), Organization of Eastern Caribbean States (OECS) and other agreements. We believe that the inclusion of all types is more defensible than including some and excluding others based upon incomplete and imprecise measures of depth. Furthermore, if we introduce a bias by including RTAs that are not deep, we are likely to underestimate the positive impact of regional integration on the level of per capita income.

23 5. Empirical Results Bilateral Estimation Results We begin by estimating the bilateral RTA equation (7) using probit. Table 3 presents the results for 1980, 1985, 1990, 1995 and 2000. We estimate each year individually to allow for differences in the slope coefficients. The geography and historical variables are included in both the RTA and trade share equations, while the political variables are included only in the RTA equation. In general, the coefficients have the correct sign and are mostly significant. For geography, two economies are more likely to form an RTA if they are close to each other; large in area; remote from other countries; and an island(s). For historical variables, two economies are more likely to create an RTA if they share a common colonizer and a common language. For political variables, two countries are more likely to form an RTA if they are both free; in a current political alliance; and share similar voting patterns in the United Nations. We next estimate the bilateral gravity equation (5). Table 4 presents the OLS and Heckman two-step results for each year. The OLS regressions include the geography and historical variables; while the Heckman regressions include the geography and historical variables, interaction terms and individual fixed country effects. The regression diagnostics are shown at the bottom. The OLS results explain 0.14 to 0.30 of the variation in the trade share, while the Heckman results explain around one-half. The F-test statistics indicate evidence of selection bias, especially in 1990 and 1995, and variation in the slope coefficients? across the treated and non-treated. The results show that the bilateral trade share increases in border, landlocked, colonial link, common colonizer and common language; while it decreases in bilateral distance. The coefficients for area, remoteness and island change signs and are generally insignificant.

24 Turning to the treatment variable rta, the trade creation estimate is much larger for Heckman (with fixed effects) than for OLS (without fixed effects). Using OLS, the coefficient for rta is 0.44 to 1.48 in value. However, using the Heckman two-step procedure, the coefficient for rta increases to 1.01 to 4.93 in value. Baier and Bergstrand (2007) also find that the trade creation estimate becomes larger and more variable when individual fixed effects are included. Nevertheless, our objective here is not to estimate the amount of trade creation, but rather to produce fitted bilateral values of the trade share and RTA membership that can be used to create instruments for cross-country analysis. Predicted Trade Shares and Regional Integration We next construct actual and fitted trade share and regional integration variables for our cross-country analysis. As with Rodrik, Subramanian and Trebbi (2004), we compute the actual trade share T using data for 1960 to 2000. Unlike the unilateral trade share, intra-bloc RTA trade varies markedly across time as new trading blocs are formed. Therefore, we create weighted averages of actual and fitted RTA variables using observed outcomes rta and the predicted probabilities rta for 1980, 1985, 1990, 1995 and 2000. We then take a simple mean across time for each country to produce our cross-country variables RTA and RTA. Table 5 presents the correlation matrix of the cross-country variables. Each cell contains the Pearson correlation coefficient (r). There are two results of interest. First, there is little correlation between unilateral and regional integration (r = 0.19). At first blush, the low correlation suggests that RTA trade could exert an independent effect on per capita income. Second, the positive correlation between the fitted and actual cross-country variables (shown as highlighted elements) is moderate to strong (r = 0.31 to 0.72). The relatively strong correlation indicates that the fitted values will serve as good instruments.

25 Cross-Country per Capita Income Results We are now ready to estimate the per capita income equation (10). Table 6 presents the OLS results for the Frankel and Romer (FR) and geography (GEO) specifications. The FR specification is shown in the first six columns, while the GEO specification is displayed in columns 7-9. In columns 1-3, we measure the dependent variable as the log of real GDP per capita in 1985 from PWT 5.6 to compare our results with those of Frankel and Romer (1999). We then use the log of real GDP per capita in 2000 from PWT 6.2 in columns 4-9. The OLS results show that both unilateral and regional integration are positively correlated with the level of per capita income. In each specification, the coefficient for the unilateral trade share and for regional trade is positive and strongly significant. The point estimate of the trade share is 0.73 to 1.09, which is close to the 0.85 found by Frankel and Romer (1999). The point estimate for RTA trade, on the other hand, ranges from 1.85 to 2.52 in value. However, because of the possibility of endogeneity bias, the OLS results cannot be interpreted as accurate or causal. Table 7 presents the 2SLS results for the FR and GEO specifications. The top panel presents the second-stage estimates for each regression; while the bottom panel shows the firststage results for the regressions in columns 3, 6 and 9. As before, we measure the dependent variable as the log of real GDP per capita for 1985 in columns 1-3 and for 2000 in columns 4-9. In the first-stage regressions, the coefficients for fitted trade share and fitted RTA trade are positive and strongly significant. The two instruments are clearly relevant with high R-squares and F-statistics. Moreover, the values of the Shea partial R-squares and the Cragg-Donald statistic indicate the instruments are not weak. In particular, the Cragg-Donald statistic exceeds the r = 0.10 critical value and thus we reject the null hypothesis of weak instruments that would lead to size distortions of 10% or more.

26 The use of 2SLS raises the estimated impact of integration, especially unilateral trade, on the level of per capita income. The point estimate of the unilateral trade share nearly doubles in value to 1.33 to 2.38. As before, our estimate of the trade share is close to the 1.85 found by Frankel and Romer (1999). Similarly, the coefficient for RTA trade increases to 2.21 to 3.65 in value. The increase in the point estimates suggest that the attenuation bias from endogeneity of integration is negative The results of Tables 6 and 7 suggest that the positive impact of regional integration on the level of per capita income operates outside of its effect on unilateral trade. If the two trade variables were linked, then the inclusion of the two would lead to a decrease in the significance of one or even both of them. Moreover, the point estimate for the unilateral trade share remains relatively stable and close to the estimate found by Frankel and Romer. However, there is likely omitted variable bias due to the low R-square. The institutions and growth literature has shown the importance of institutions in explaining cross-country income differences. We therefore include our RTA trade variable in a deep determinants specification. Table 8 shows the OLS results. We include the integration variables in columns 1-3, add institutions (rule of law) in columns 4-6, and then add geography (distance from equator) in columns 7-9. The results are similar to those found by Rodrik, Subramanian and Trebbi (2004). The coefficients for integration, institutions and geography are all positive and strongly significant. The one exception is column 9 where the coefficient for RTA membership is not significant at the 10% level. Table 9 presents the 2SLS results for the RST specification. The top panel shows the second-stage estimates for each regression; while the bottom panel shows the first-stage results for the regressions in columns 3, 6 and 9. The Hansen J statistic and the corresponding p-value indicate that the instruments are orthogonal. In the first-stage regressions, the coefficients for the