Heterogeneous Economic Integration Agreement Eects

Transcription

1 Heterogeneous Economic Integration Agreement Eects Scott L. Baier, Jerey H. Bergstrand, and Matthew W. Clance Ÿ February 15, 2015 Abstract Gravity equations have been used for more than 50 years to estimate ex post partial eects of economic integration agreements (EIAs) on bilateral international trade ows. However, typically researchers estimate a single average (partial) eect on (the log change in) bilateral trade from an EIA; when eects are allowed to be heterogenous using instead a plethora of dummies, estimates are weak and imprecise. We oer four potential contributions. First, using a random coecients econometric approach, we show that the heterogeneity in various EIAs' partial eects on trade ows far exceeds that explained simply by variation in depth of the trade-liberalization agreement. Second, we use standard Armington- and Melitz-type general equilibrium trade models to motivate theoretically the roles of variable trade costs and xed export costs (with and without network eects), respectively, for explaining heterogeneous partial eects of EIAs on trade ows, as well as on intensive and extensive (product) margins. Third, we show empirically that the heterogeneity in EIAs' partial eects on the intensive margin is explained well just by distance and adjacency, capturing variable trade costs; however, the heterogeneity in EIAs' partial eects on the extensive margin is explained empirically by distance and adjacency, as well as several other cultural and institutional variables, capturing variable and xed exports costs. Fourth, in the context of a class of standard quantitative trade models, we show that such estimated heterogeneous eects can predict percent of economic welfare eects of EIAs and can potentially predict ex ante a potential EIA's partial trade-ow eect and general equilibrium welfare eect. Key words: International trade, economic integration agreements, gravity equation JEL classication: F1, F21, F22, F23 Acknowledgements: To be added. Aliation: John E. Walker Department of Economics, Clemson University. Address: John E. Walker Department of Economics, Clemson University, Clemson, SC USA. sbaier@clemson.edu. Aliation: Department of Finance, Mendoza College of Business, Department of Economics, and Kellogg Institute for International Studies, University of Notre Dame, and CESifo Munich. Address: Mendoza College of Business, University of Notre Dame, Notre Dame, IN USA. bergstrand.1@nd.edu. Ÿ Aliation: Department of Economics, University of Pretoria. Address:Department of Economics, University of Pretoria, Hateld 0028, South Aca. mwclance@gmail.com

2 The welfare eects in this class of [quantitative trade] model are linked to the change in the share of trade that takes places inside a country... Intuitively, because the initial ows are so small, even doubling trade with ex-colonies will result in very tiny changes in the share of expenditure that is spent locally. In contrast, adding even a few percentage points of trade with a major partner will be much more important for welfare. (Head and Mayer. Gravity Equations, Handbook of International Economics, vol. 4, 2015) 1 Introduction The gravity equation has been used for 50 years since Tinbergen (1962) to explain statistically ex post the cross-sectional and panel variation of bilateral international (aggregate goods) trade ows and the partial eects of economic integration agreements (EIAs) on such ows. However, the link between these ex post estimates and the welfare gains from trade liberalization has been at best tenuous. This paper addresses this shortcoming, picking up to a large extent where Arkolakis, Costinot, and Rodriguez-Clare (2012) left o. Typically, the trade ow from one country to another in a gravity equation is explained using the exporting and importing countries' gross domestic products (GDPs), bilateral distance, and an array of other explanatory variables, including dummy variables for EIAs. In fact, one focus of Tinbergen (1962) was to examine the partial eect of preferential trade agreements on trade ows. 1 While Tinbergen (1962) is typically cited as the rst published gravity-equation study of trade ows, researchers using the international trade gravity equation have seldom explored some of the novel contributions and anomalies of this seminal study. For instance, Tinbergen's rst specications included dummy variables for common membership in the British Commonwealth and for common membership in the Belgium- Netherlands-Luxembourg (BENELUX) economic union. Allowing the two agreements to have heterogeneous (partial) eects, Tinbergen found trivial impacts of these agreements 1In this study, we are concerned initially with partial equilibrium, not general equilibrium, eects; typically, partial eects are assumed to be the same across pairs (based on an estimated average treatment eect), but general equilibrium eects are sensitive to the levels of country-pairs' economic variables, cf., Anderson and van Wincoop (2003), Baier and Bergstrand (2009), and Bergstrand, Egger and Larch (2013). Also, our estimates of partial eects will account for several major advances in gravity-equation estimation, such as multilateral resistance (price) terms and endogeneity of EIA dummies. In the latter part of the paper focusing on welfare, we address general equilibrium eects. 2

3 on members' trade. However, in a later specication using a single dummy for membership in any preferential trade agreement, Tinbergen found that common membership increased trade for the typical pair by more than 100 percent. This completely overlooked contrasting result is another motivation for our paper, which we address theoretically and empirically. Hundreds of gravity-equation analyses have been published (and many more completed) in the last 50 years with thousands of estimates of (like Tinbergen, 1962) the bilateral trade impacts of common membership in some form of EIA. 2 Cipollina and Salvataci (2010) conducted a meta analysis of estimates of the (partial) eects of EIAs on trade ows and found a mean (median) eect of 0.59 (0.38), implying an increase of 80 (46) percent. The eects range from to including outliers. Once xed or random eects are introduced, the minimum estimate is 0.01 and the maximum estimate is Similarly, Head and Mayer (2015) in a meta analysis found mean (median) estimates of 0.59 (0.47). The vast heterogeneity in EIAs' estimated trade-ow eects motivated Baier and Bergstrand (2007) and Baier, Bergstrand, and Feng (2014) to rene estimation of gravity equations to generate more plausible, consistent, and precise EIA (partial) eects. However, more remains to be done. One of the limitations of Baier and Bergstrand (2007), or BB, is that the study estimates one sole average EIA partial eect, as in the latter specication in Tinbergen (1962). In reality, EIAs' eects on trade are likely to be quite heterogeneous across agreements. One obvious source of this heterogeneity is the depth of the agreement. Kohl, Brakman and Garretsen (2014) recently showed empirically that the degree of comprehensiveness of provisions and the degree to which such provisions are legally enforced aect an EIA's (partial) impact on bilateral trade. Baier, Bergstrand and Feng (2014), or BBF, addressed this shortcoming using the BB methodology but allowing four EIA dummies to distinguish between one-way preferential trade agreements (OWPTAs), two-way PTAs (TWPTAs), free trade agreements (FTAs), and a dummy for customs unions, common markets, and economic unions (CUCME- 2We use the term economic integration agreement to broadly capture any of one-way or two-way preferential agreements, free trade agreements, customs unions, common markets, or economic unions. 3

4 CUs). While the results revealed that deeper integration agreements had signicantly larger eects, we will show that there still remains enormous heterogeneity across agreements' eects within each type of EIA category and even within a given plurilateral EIA. A natural alternative approach to allow for heterogeneous EIA eects is to include individual dummies for each agreement; this has been used by several authors, cf., Frankel (1997), Ghosh and Yamarik (2007), Baier, Bergstrand and Vidal (2007), Baier, Bergstrand, Egger and McLaughlin (2008), and Eicher, Henn, and Paqpageorgiou (2012). However, this approach faces at least three limitations. First, such an approach often leads to weak estimates. The reason is that unless the EIA is plurilateral with numerous common memberships there is insucient variation in the RHS dummy variable. This was the dilemma Tinbergen (1962) faced, leading to the trivial EIA eects of the British Commonwealth and BENELUX economic union; there were only three countries in each agreement in his sample and only six 1's in each of the dummy variables. This was the same constraint leading BBF to combine the deeper integration agreements into one dummy variable (CUCMECU); there was limited variation in the right-hand-side (RHS) EIA dummy variables. Moreover, the vast growth in the number of EIAs in the world since 1990 has been primarily in bilateral, rather than plurilateral, EIAs, making estimation of ex post eects using gravity equations that much more dicult. Second, even within a plurilateral agreement, it is possible that the bilateral eect of a common EIA diers owing to variance in geographic, institutional, legal and cultural factors; typical gravity estimates ignore these considerations. Third, as will be discussed later, even if individual EIA dummies' partial eects could be estimated with precision and consistency, they remain ex post estimates. We will show that our approach in this paper will facilitate potential forecasting ex ante of EIA partial eects by country-pair and consequently general equilibrium welfare predictions, as we uncover observable determinants of these heterogeneous partial eects. This paper has three main goals. First, motivated initially by a random coecients econometric analysis that demonstrates the enormous heterogeneity in EIAs' partial eects across country pairs (even after accounting for dierent degrees of trade liberalization), we 4

5 use Armington- and Melitz-type models to explain why EIA partial eect estimates should be sensitive to the levels of country-pairs' variable and xed export costs (with or without network externalities), respectively. The notion that the (partial) trade-cost elasticity ɛ in Arkoloakis, Costinot, and Rodriguez-Clare (2012) is sensitive to the levels of country-pair variable and xed export costs has been overlooked in the international trade literature. Second, guided by these theoretical results, we show empirically that the heterogeneity in EIAs' eects can be explained well by (exogenous) observable factors commonly used to explain these costs. In particular, we show that distance and adjacency capturing variable transportation costs explain well the heterogeneity in EIA partial eects on the intensive (product) margin. Moreover, distance, adjacency, and typical gravity dummy variables re- ecting common institutional and cultural country characteristics capturing xed export costs explain well the heterogeneity in EIA partial eects on the extensive (product) margin. To the best of our knowledge, only two studies have estimated heterogeneous EIA eects using interaction terms like here to avoid the dilemma of a multitude of individual dummies noted above. Vicard (2011) investigated empirically interactions of numerous economic variables with EIA dummies, but the study was not guided by theory and so interaction eects lacked economic interpretation. Cheong, Kwak and Tang (2014) examined empirically interactions of EIA dummies with measures of GDP size and similarity and found signicant eects, but this study also lacked theoretical guidance. Also, both of those studies only looked at aggregate trade ows. Our study is unique by oering theoretical guidance from Armington and Melitz models to understand the roles of variable and xed export costs with or without network externalities for explaining heterogeneous EIA eects, for explaining dierential EIA eects quantitatively and qualitatively on intensive and extensive (product) margins, and for controlling for various degrees of EIA liberalization (as raised in Kohl, Brakman and Garretsen, 2014). Third, in the context of recent insights from Arkolakis, Costinot, and Rodriguez-Clare (2012), or ACRC, we show that two more readily observable statistics are largely sucient to explain quantitatively the (general equilibrium) economic welfare eects of an EIA trade 5

6 liberalization. ACRC showed for a certain class of quantitative trade models and under specic assumptions that two statistics were sucient to quantify the economic welfare eects for a country j of a decline in ad valorem trade costs from i to j (τ ); those statistics are the trade-cost elasticity (ɛ) and the share of domestic expenditure on domestic output in importer j (λ jj ). Unfortunately, researchers have found precise and consistent estimates of ɛ to be elusive to say the least. Moreover, at the aggregate level, even measurement of λ jj is dicult since aggregate trade ows of goods are measured on a gross basis, aggregate trade ows of services are measured poorly, and gross domestic product (GDP) is measured on a value added basis. 3 Guided by ACRC, we show here that only two (more readily observable) statistics the bilateral trade ow share (λ ) and the estimated partial EIA eect ( β ) are needed to explain quantitatively 85 percent or more of the importer's (net general equilibrium) welfare gain from an EIA between i and j. 4 Thus, as suggested in our introductory quote from Head and Mayer (2015), EIAs with major trading partners where the bulk of EIAs have occurred can explain a signicant portion of the the welfare gains from such trade liberalizations. 5 The plan of our paper is as follows. In section 2, we show using a random coecients econometric model that the variation in EIA eects across country-pairs far exceeds what can be explained by the depth of the EIA and the degree of trade liberalization. Moreover, preliminary empirical evidence from Baier, Bergstrand and Clance (2014) suggests that EIA dummies' coecient estimates are systematically related to various observable bilateral tradecost proxies. However, none of the theoretical foundations for gravity equations provided by the standard quantitative trade models Armington, Ricardian, Krugman, and Melitz, such as in Anderson and van Wincoop (2003), Eaton and Kortum (2002), Baier and Bergstrand 3This is one reason why internal trade is omitted in most gravity equations studies. At the disaggregate level, measurement of λ jj for many goods is more feasible. However, panel data sets for goods at such levels face limited time dimensions, cf., Bergstrand, Larch, and Yotov (2014). 4Moreover, guided by the approach in Baier and Bergstrand (2004), we show that 53 percent or more of the empirical likelihood of countries i and j having an EIA also can be explained by just these two statistics, λ and β. Once distance is included, this explanatory power rises to 85 percent also. 5Furthermore, the heterogeneity in EIAs' partial eects may well be more important for explaining welfare eects of EIAs than the heterogeneity of EIAs' eects associated with (general equilibrium) multilateral resistance terms and income changes, cf., Head and Mayer (2015). 6

7 (2001), and Chaney (2008), respectively can explain the sensitivity to the levels of tradecost proxies of the partial elasticity of trade to a variable (iceberg, or ad valorem) tari-rate decline, such as tari removals from a free trade agreement or customs union. In section 3, we show in the context of a simple Armington trade model how tari removals' eects on trade (at the intensive margin) that is, the trade-cost elasticity can be sensitive to levels of ad valorem bilateral variable export costs between two countries. In section 4, we extend the Melitz-type heterogeneous rms theoretical model in Krautheim (2012) to motivate how the interactions of exogenous factors inuencing xed export costs (such as bilateral distance as well as bilateral dummy variables capturing legal, institutional, and cultural characteristics) with endogenous xed export costs associated with network eects can additionally explain theoretically the sensitivity of the elasticity of the extensive margin of trade ows with respect to variable tari rates. In section 5, we pose six conjectures of the relationships between standard observable gravity covariates and their likely impacts on the eects of EIAs on trade ows, intensive product margins, and extensive product margins. Importantly, just as two-stage (or nonlinear) estimation has fundamentally altered gravity-equation analyses to account for zeros, heteroskedasticity, and rm heterogeneity to avoid endogeneity bias, our goal is to raise awareness of accounting properly for heterogeneous (partial) EIA eects. Section 6 addresses econometric issues and reports the empirical results. Since our theory suggests that EIAs' eects may be inuenced by factors inuencing both the intensive and extensive margins of trade, Section 6 employs the Hummels and Klenow (2005) productmargin-decomposition methodology to explore empirically how distance and other factors inuence such margins' EIA eects. We show that various factors inuence EIAs' eects on intensive and extensive margins of trade dierently, quantitatively and qualitatively. This section also provides a robustness analysis of our main results to lagged terms-of-trade eects, nontradable goods' cutos, and interaction eects by type of EIA. 6 6It is important to note that, although we focus on heterogeneous partial eects of EIA dummies, our analysis hold in principle for ad valorem tari rates as well, such as in Baier and Bergstrand (2001). Our focus empirically on heterogeneous EIA dummy coecients, rather than heterogeneous tari-rate elasticities, is due solely to the lack of high quality ad valorem tari-rate data and the empirical prominence of EIA dummies in the literature. Nevertheless, our theory will be cast with a focus on heterogeneous partial 7

8 Finally, Section 7 shows that the our approach to gravity-equation modeling now makes more plausible ex ante use of gravity equations for predicting the partial eects of future EIAs and their likely welfare eects. Studies such as BB and BBF can help policymakers predict future partial (and then general) equilibrium eects of a planned EIA; BB (BBF) predicts the average eect without (with) regard to type of EIA. However, those predicted partial eect estimates are homogeneous across country-pairs (based on average treatment eects). In this paper, we show that even the partial impact of a given type of EIA is sensitive to the geographic, legal, institutional, and cultural characteristics of each country pair. By estimating quantitatively the sensitivity of estimated partial eects to these characteristics, this enables gravity equations to more precisely inform policy makers ex ante of pair-specic predicted impacts of EIAs accounting for both heterogeneous general and partial equilibrium eects. Moreover, we will show that the heterogeneity of EIA partial eects helps to explain the likely welfare gains and predictability of EIAs, as well as the timing of EIAs. Put succinctly, previous gravity equations allowing for heterogeneous partial eects of EIAs on trade have been limited not just by weak estimates, but allowed only ex post evaluation. Our paper suggests a methodology for generating robust and precise partial eect estimates that can also be used potentially for ex ante trade and welfare analysis, once such estimates are linked to standard welfare approaches as in ACRC. 7 We conclude the paper in Section 8. 2 Heterogeneous EIA Eects In this section, we demonstrate that there is considerable heterogeneity in the (partial) effects of EIAs, independent of the depth of the EIA, using a random coecients econometric approach. First, for theoretical context we use a standard gravity equation in the spirit of ACRC to illustrate that most classes of quantitative trade models such as Armington, tari-rate elasticities. 7It should also be noted that traditional ex ante computable general equilibrium models typically use trade-cost elasticities previously estimated from empirical specications assuming homogeneous average partial eects. 8

9 Krugman, Ricardian or Melitz yield gravity equations with homogeneous (partial) EIA impacts. Second, we address econometrically how BBF can be extended to allow for random coecient estimates to illustrate the large heterogeneity of EIA eects even after accounting for the depth of EIA. The third section discusses the data and presents the results. 2.1 Gravity Recently, ACRC demonstrated that a gravity equation surfaces for a large class of quantitative trade models that feature four main assumptions: (1) Dixit-Stiglitz preferences; (2) one factor of production (typically, labor): (3) linear cost functions; and (4) perfect or monopolistic competition. 8 Trade models satisfying these four assumptions are Armington (cf., Anderson and van Wincoop (2003)), Ricardian (cf., Eaton and Kortum (2002)), Krugman (cf., Baier and Bergstrand (2001)), and Melitz (cf., Chaney (2008)). ACRC concluded that the gravity equation provides a common method for estimating the (homogeneous) trade-cost elasticity across these dierent approaches. Modifying here ACRC to a panel context and assuming a single aggregate industry, ACRC show that a gravity equation consistent with all four theoretical frameworks can be described by: X t = χ tn it (w it τ t ) ɛ Y jt Kk=1 χ kjt N kt (w kt τ kjt ) ɛ, (1) where X t is the trade ow from exporter i to importer j in year t, N it is the measure of goods in country i that can be produced in year t, w it is the wage rate in country i, τ t is (gross) variable trade costs of country i's products into j (i.e., τ t > 1), ɛ is the (trade-cost) elasticity of imports with respect to variable trade costs (ɛ < 0), Y jt is aggregate expenditure in j, and χ t captures all structural parameters other than τ t (for example, xed export costs from i to j, or F X t). For the purposes of this paper, the variables of interest are τ t, 8ACRC also note three macro-level restrictions: (1) trade is balanced; (2) aggregate prots are a constant share of aggregate revenues; and (3) the import demand system is Constant Elasticity of Substitution (CES). 9

10 F X t, and χ t. Typically, researchers have assumed that the formation of an EIA (such as a free trade agreement) between i and j lowers τ t. However, EIAs are broad agreements reaching beyond elimination of tari rates and variable trade costs; they likely also lower xed export costs, denoted F X t. Consequently, later in the paper, we will distinguish timevarying bilateral xed export costs associated with policy changes, denoted Ft XP, from time-varying bilateral xed export costs associated with non-policy, or natural, changes, denoted F XN t. 9 ACRC showed for this important set of quantitative trade models that the percentage change in welfare for country j (Ŵjt) from a trade liberalization could be computed from two sucient statistics: (i) the share of j's expenditures on domestically produced goods (λ jjt ) and (ii) the trade elasticity (ɛ, dened above). Specically, ACRC argued that: Ŵ jt = λ jjt 1/ɛ (2) where λ jjt is the importer's share of expenditures on domestic output (λ jjt = X jjt /Y jt ). Hence, the estimated partial eect of an EIA (through changes in τ t and/or F X,t) is central to estimating the welfare eects of an EIA. 2.2 Heterogeneous Partial EIA Eects In this theoretical context, BB showed that unbiased and precise estimates of EIAs on bilateral trade ows could be captured using the gravity-equation specication and ordinary least squares (OLS): 10 ln X t = α + η it + θ jt + ψ + βeia t + υ t (3) 9See Horn, Mavroidis, and Sapir, 2010, on the provisions covered in an anatomy of European Union and United States' preferential trade agreements. 10For now, we ignore zero trade ows, allowing a logarithmic transformation of equation (1). 10

11 where η it is an exporter-year xed eect, θ jt is an importer-year xed eect, ψ is a countrypair xed eect, and υ t is an error term. Equation 3 is commonly referred to as a xed eects model. However, it will be useful now to emphasize as well that this is also a xed parameters model (i.e., β is a xed parameter). Hence, BB was a xed eects, xed parameter model, and that is also the case to the best of our knowledge for virtually all like gravity analyses. A key insight of BB was to show methodologically and empirically the importance of the country-pair xed eect for controlling for the endogeneity of the EIA variable. As noted earlier, a limitation of equation (3) is that it imposes a common estimated average partial eect for all EIAs. In reality, EIAs are heteogeneous and their eects on trade ows are likely to be heterogeneous. In specications such as equation (3), this heterogeneity in EIAs' partial eects is captured in the error term, υ t, which is assumed to be uncorrelated with the other right-hand-side (RHS) variables. Two considerations arise when considering the potential heterogeneous eects of EIA dummies. One consideration is that a single EIA dummy cannot capture the heterogeneity among EIAs in their degrees of trade liberalization. Consequently, BBF accounted for heterogeneity in the eects of EIAs by depth of agreement, running a specication including separate dummies for one-way PTAs (OWPTA), two-way PTAs (TWPTA), FTAs, and a dummy combining customs unions, common markets, and economic unions (CUCMECU) due to the limited number of these more integrated EIAs in their sample ending in Hence, BBF ran the xed eects, xed parameters model: ln X t = α + η it + θ jt + ψ + β 1 OW P T A t + β 2 T W P T A t + β 3 F T A t +β 4 CUCMECU t + υ t (4) using OLS In this paper, we have extended that data set to 2011, enlarging substantially the number of EIAs with customs unions (CUs), common markets (CMs), and economic unions (ECUs), and so will treat each of those types separately. 12Ignoring zeros could potentially bias results, due to country selection; moreover, one must account for 11

12 However, a second consideration is that perhaps due to variable and/or xed bilateral export costs the partial eect on trade of EIAs with a given degree of trade liberalization may be heterogeneous. For tractability, suppose EIA t represents EIAs with a given degree of trade liberalization. Following Cameron and Trivedi (2005) (p. 774), we can consider the specication: ln X t = α + η it + θ jt + ψ + β EIA t + υ t (5) where the partial eect of an EIA on ln X t is allowed to be pair-specic. One way to interpret the heterogeneity in the β 's is to assume it is random. For instance, assume β = β + b where b is a zero mean random variable. In this case, the expectation of β is: E(β η it, θ jt, ψ, EIA t ) = β In section 2.3 below, we make this assumption to illustrate the enormous heterogeneity in values of β even after accounting for dierent degrees of trade liberalization. We will refer to these regressions as the random coecients model. Alternatively, suppose there exists a set of variables Z such that: E(ln X t α, η it, θ jt, ψ, β, EIA t, Z ) = α + η it + θ jt + ψ + β EIA t (6) potential bias due to rm heterogeneity in aggregate data, cf., Helpman, Melitz, and Rubinstein (2008). However, BBF showed that potential bias due to country selection and rm heterogeneity was largely cross sectional in nature and could be accounted for in panel data by the pair xed eects; see BBF and its Online Theoretical Supplement for a comprehensive discussion. Also, due to potential heteroskedasticity owing to Jensen's inequality, some studies have employed Poisson Quasi-Maximum Likelihood (PQML). Due to our specication using a very large number of xed eects, researchers have only been able to obtain convergence under PQML for a limited time series in the panel (i.e., a short panel), cf., Bergstrand, Larch, and Yotov (2014). Consequently, due to our long panel, this limitation allows us to only use OLS. We also note that Bergstrand, Larch and Yotov (2014) found, if anything, that OLS biased downward the EIA partial eect estimates relative to PQML estimates. 12

13 Without knowing the true values of the β, we take expectations over all variables to obtain: E(ln X t α, η it, θ jt, ψ, EIA t, Z ) = α + η it + θ jt + ψ +E(β α, η it, θ jt, ψ, EIA t, Z )EIA t We assume that the expected eect of an EIA between i and j, conditioning on all other variables, is given by: E(β η it, θ jt, ψ, Z ) = β + b Z (Z µ Z ) (7) where Z µ Z denotes the de-meaned values of Z. Absent knowledge of β, following Cameron and Trivedi (2005) we should estimate instead: E(ln X t α, η it, θ jt, ψ, EIA t, Z ) = α + η it + θ jt + ψ + βeia t + b Z (Z µ Z )EIA t (8) We refer to this as the mixed model approach. 13 One of the main goals of this paper is to identify the variables in Z, not just to determine the best linear unbiased predictors but also to produce ex ante trade and welfare eects. In fact, Sections 3 and 4 below will rst motivate theoretically in the context of Armington and Melitz models the roles of bilateral variable and xed export costs, respectively, in explaining the heterogeneous β. However, before we proceed to exploring potential sources of Z, it would be useful rst just to get a sense of how large the potential heterogeneity of β might be even without knowing the source (i.e., the Z ). Consequently, to illustrate how large the heterogeneity of β might be, we rst estimate equation (5) allowing for all six dierent types of EIAs using the random coecients approach. Estimation of equation (5) will account for the heterogeneity in dierent degrees of trade liberalization; however, without controlling at this stage for the Z µ Z, we anticipate bias in these preliminary (motivating) random coecient 13In some contexts, this is referred to as a random slopes model, cf., Wooldridge (2012). 13

14 estimation results Data and Results Nominal trade ows are from the United Nations' COMTRADE database for the years 1965, 1970, 1975, 1980, 1985, 1990, 1995, 2000, 2005 and Bilateral distances, adjacency, common language, common religion, common legal origin, and common colonial history (used later) are from the BACI data set. The data set for EIAs comes from Baier and Bergstrand's data set for There are 183 countries included in our data set. Appendix 1 lists the EIAs in our sample and (at its end) the countries included. Table 1 provides a decomposition of the data set into types of agreements. Note that the vast majority of observations have no economic integration agreement and less than 6 percent of the observations have FTAs, CUs, CMs, or ECUs. In Table 1, the subtotal of 726,468 observations indicates the total number of annual positive trade ow observations for 49 years from for the 33,306 country-pairs; the missing observations are composed largely of zeros. As discussed in BB and BBF, we only use observations from every ve years. The primary reason is that the number of xed eects related to equation (3) or (4) with potentially 33,306 (directional) bilateral trade ows and 49 years, or 1,631,994 potential observations (or even 726,468), would not be computable in STATA; restricting the data to every ve years from reduces the number of xed eects to a manageable level for estimation. A second reason is that Cheng and Wall (2005) and Wooldridge (2000) both argue in favor of using data drawn from a period longer than annually. For instance, Cheng and Wall (2005, p. 8) note that Fixed-eects estimations are sometimes criticized when 14Two further considerations are noteworthy. First, since gravity equations typically use binary variables to capture the policy change rather than an ad valorem variable such as τ t, the dummy variable's coecient estimate is a combination of ɛ and ρ, where ρ is an estimate of the eect of the EIA's formation on lowering τ t. For this study, we will use the dierences in partial eects across types of EIAs to gain insight about ρ. For example, the dierence in BBF between a deeper EIA and a one-way PTA (e.g., Generalized System of Preferences agreement) is (= ). For a common ɛ, this dierence informs us of the relative change of τ between the two types of agreements. Second, the heterogeneity of EIA eects can in principle be captured by introducing multiple variables such as τ t for each EIA. While feasible, recall that such ex post estimates are weak econometrically for the reason stated earlier. 15See jbergstr. 14

15 applied to data pooled over consecutive years on the the grounds that dependent and independent variables cannot fully adjust in a singe year's time (italics added). Consequently, our potential number of observations, based on positive trade ows, is 155,718. Table 2 presents several sets of results. Since this is the rst paper to also provide separate partial eects for all six EIA categories in the Baier-Bergstrand EIA Database, before providing the results from the random coecients specication equation 5 we provide the results rst for equations (3) and (4). Table 2 has ten columns, reporting the results of eight dierent specications of equations (3) and (4). Column (1) provides the names of alternative groupings of EIAs. Column 2 provides the expected coecient sign. The rst three specications in columns (3)-(5) are similar to equation (3), with no lags and combining dierent EIA types into one or two dummy variables. The fourth specication in column (6) is equation (4). The last four specications in columns (7)-(10) include additionally a veyear lag of each RHS variable. Previous studies such as BB and BBF have shown that, due to phase-ins of agreements as well as lagged eects of EIAs on terms-of-trade, EIA dummies tend to have lagged eects on trade ows, extensive margins and intensive margins. Specication 1 in column 3 uses a conventional grouping of EIAs: free trade agreements (FTAs, or 3), customs unions (CUs, or 4), common markets (CMs, or 5) and economic unions (ECUs, or 6). We denote this grouping EIA3-6, which reects the inclusion of EIAs of levels 3-6. This grouping is quite common and was used in BB. The coecient estimate of 0.56 is very similar to the partial eect estimated in BB; it implies the typical EIA using categories 3-6 increases trade (absent any general equilibrium eects) by approximately 90 percent. It is worth noting that the coecient estimate in column (3) of 0.56 is very similar also to the mean estimate of 0.59 in both the meta analyses in Cipollina and Salvatici (2010) and Head and Mayer (2013). So our baseline EIA coecient estimate is very similar to comparable estimates in the literature using non-random coecient estimates and ignoring interaction terms. Specication 2 in column 4 adds another EIA dummy to capture the eects of preferential trade agreements (PTAs). In the Baier-Bergstrand EIA database, a 1 denotes a 15

16 one-way PTA (OWPTA), including Generalized System of Preferences (GSP) agreements. Such agreements provide extensive liberalization to imports from one (typically, developing) exporter into another (typically, developed) importer. Much evidence recently shows that such one-way PTAs have not had signicant eects on trade. A 2 denotes a two-way preferential trade agreement (TWPTA), where there are preferences extended but it is not an FTA. In Specication 2, we include EIA1-2 as well as EIA3-6; EIA1-2 denotes a one if two countries have a one-way or two-way PTA. Not surprisingly, the coecient estimate for EIA1-2, 0.096, is considerably smaller than that for EIA3-6; there is less liberalization typically in agreements in EIA1-2. Specication 3 in column 5 uses EIA1-6. EIA1-6 includes all six levels of EIAs. Not surprisingly, the partial eect is roughly half of that for EIA3-6 in column 3 and roughly halfway between the coecient estimates in column 4. Specication 4 in column 6 includes six dummies, one for each of the six types of EIAs in the Baier-Bergstrand data set, which now has observations through The coecient estimates tend to reect the increasing order of perceived trade liberalizations underlying each type of EIA. The statistically insignicant partial eect of 0.02 for one-way PTAs (EIA1) is consistent with other studies examining such agreements' trade eects. The coecient estimate of 0.25 for two-way PTAs is statistically signicant. The statistically signicant FTA coecient estimate of 0.53 accords with other studies for FTAs. The statistically signicant customs union coecient estimate of 0.84 implies that such existing agreements tend to have deeper integration than FTAs. Common markets are expected to be even more integrated than FTAs and CUs, so the statistically signicant coecient estimate of 1.12 for EIA5 (CMs) is consistent with this notion. Finally, economic unions tend to be more integrated than common markets, but are dierentiated more along the lines of coordinated monetary and/or scal policies than deeper integration in trade. Consequently, the slightly lower coecient estimate for ECUs of is plausible in light of the estimate for the common market's coecient. However, the coecient estimates for CM and ECU are not 16

17 statistically dierent from each other. 16 The specications in columns 7-10 in Table 2 are analogous to the specications in columns 3-6 except for including additionally a ve-year lag of the EIA. We will not go through these results in detail, as the outcomes conform to those in earlier studies such as BB and BBF, whereby EIAs have lagged eects on trade ows due to phasing-in of agreements and lagged terms-of-trade eects. None of the results in Table 2's specications 7-10 are inconsistent with expectations. Note that the total EIA eect for common markets is 1.16 whereas that for economic unions is We now turn to the results of estimating equation (5). Because of the massive number of coecient estimates, we cannot use traditional reporting in the form of a table. For tractability, the results are presented in graphical form. For brevity, we present the results for the most interesting four types of EIAs: FTAs, CUs, CMs, and ECUs. For each type of EIA, we ordered coecient estimates from lowest value to highest value, and constructed curves to visualize the range of coecient estimates. For instance, for FTAs there are approximately 2,500 coecient estimates β 3. For each of CUs, CMs, and ECUs, there are smaller numbers of coecient estimates, as there are fewer numbers of such types of EIAs. Figure 1 provides the results from the random coecients model estimation. Figure 1 is divided into four panels. 17 The upper left panel shows a plot of the coecient estimates of β 3 for FTAs. The upper right panel shows a plot of the estimates of β 4 for CUs. The bottom left panel shows a plot of the estimates of β 5 for CMs. The bottom right panel shows a plot of the estimates of β 6 for ECUs. We note several distinguishing features of the results. The rst important conclusion from these results is that there is considerable heterogeneity in the (partial) eects of EIAs on trade ows. This is the rst paper to use a random coecients model to demonstrate the extent of this heterogeneity. Second, the heterogeneity in EIA eects cannot be explained solely by the depth of liberalization. FTAs have an average partial EIA eect of 0.33, CUs have an average eect of 0.92, CMs have an average eect of 0.91, and 16Once lagged EIA eects are included, the total EIA coecient estimates for common markets and economic unions are nearly identical. 17Enlarged versions of each panel are available in Appendix 3. 17

18 ECUs have an average eect of Note that these estimates vary from the corresponding EIAs' partial eects in Table 2. As noted in Section 2.2, this is to be expected; without the inclusion of the Z µ Z in the regressions, the estimated partial eects in the random coecients regressions are expected to be biased. Thus, while there is heterogeneity in the EIA eects across EIA types, there remains considerable heterogeneity across EIAs even within each type of agreement. Interestingly, as shown soon in Sections 3 and 4, the standard quantitative trade models that dominate the theoretical international trade literature of determinants of bilateral trade in such papers as Baier and Bergstrand (2001), Eaton and Kortum (2002), Anderson and van Wincoop (2003), and Chaney (2008) do not predict heterogeneous partial eects; the partial trade-cost elasticity is a parameter. Consequently, these preliminary empirical results call rst for extending the theoretical foundations for the gravity equation to determine factors that might explain such heterogeneous partial EIA eects. 3 A Simple Armington Trade Model with Only Variable Trade Costs Duty is not assessed on CIF charges. Customs and Border Protection (CBP) value is determined based on the "Price Paid" or "Payable" for the goods, which is usually on the bill of sale or invoice and bill of lading as the Freight On Board (FOB) price. (U.S. Customs and Border Protection website, 2014) One of the persuasive aspects of ACRC was rst demonstrating their main results in the simplest possible framework: an Armington trade model. The purpose of this section is to illustrate similarly in a simple Armington trade model how the elasticity of bilateral trade between a country-pair with respect to ad valorem taris the trade-cost elasticity (ɛ = 1 σ, in the Armington model) can be sensitive to the amount of distance between the country-pair. First, in a seminal article using the Armington model, Anderson and van Wincoop 18

19 (2003) show that the three assumptions of (i) an N-country endowment economy where each country produces a dierentiated good, (ii) consumers have identical constant-elasticity-ofsubstitution (CES) preferences, and (iii) market-clearing yield a gravity equation: X = Y iy j Y W Π 1 σ i τ 1 σ P 1 σ j, where Π 1 σ i = N Y j τ 1 σ j=1 Y W Pj 1 σ, P 1 σ j = N Y i τ 1 σ i=1 Y W Πi 1 σ, (9) where X is the nominal trade ow from country i to country j, Y i (Y j ) is the nominal GDP in i (j), Y W is world GDP, τ is ad valorem gross trade costs (including taris and transportation costs) on goods exported from i to j (τ > 1), σ is the elasticity of substitution in consumption, and: [ N ] 1/(1 σ) P j = (p i τ ) 1 σ i=1 (10) where p i is exporter i's Freight-On-Board (F OB) price of its dierentiated product. Second, following Hummels and Skiba (2004), we assume the price of a good produced in country i facing importer j, p, depends on the exporter's F OB price, p i, and a two-part trade cost that includes both an ad valorem gross tari rate, t > 1, and per unit shipping costs, freight : p = p i t + freight (11) Note that, consistent with the introductory quote, tari duties are applied to the F OB price. Equation (8) can be rewritten as: p = p i τ = p i (t + f ) (12) where f > 0 is the ad valorem-equivalent of freight charges per unit (f = freight /p i ). See Hummels and Skiba (2004) for empirical support for this specication. Substituting 19

20 equation (12) into equations (9) and (10) yields: X = Y iy j Y W (t + f ) 1 σ Π 1 σ i P 1 σ j (13) which is analogous to equation (18) in Anderson and van Wincoop (2004, p. 715), ignoring for brevity the z terms in their paper (which are not important here). Third, taking the logarithm of equation (13) and dierentiating ln X with respect to ln t yields the trade-cost elasticity for the bilateral tari rate: ln X / ln t = 1 (1 σ) (14) (f /t ) + 1 Consequently, in light of the additive relationship between f and t in equations (12) and (13), the trade-cost elasticity for tari rates is no longer parametric, as in all the so-called quantitative trade models noted earlier. Moreover, as well established in Hummels and Skiba (2004) and Hummels (2007), ad valorem bilateral freight costs f are highly positively correlated with bilateral distances. Hence, (in absolute terms) the elasticity of bilateral imports with respect to ad valorem tari rates is a negative function of bilateral distance. Since EIAs reduce tari rates, this suggests the hypothesis that the elasticity of (the intensive margin of) trade ows with respect to EIA may be negatively related to bilateral distance (and perhaps other bilateral trade impediments). 4 A Melitz-Type Model with Variable and Fixed Export Costs and Endogenous Network Eects The previous simple Armington trade model establishes that the elasticity of trade with respect to EIAs may be heterogeneous depending upon bilateral distance and possibly other bilateral trade impediments. However, the simple Armington model allows only for variable trade costs and changes in trade on the intensive margin. In this section, we provide a 20

21 more general model that also includes xed export costs, rm heterogeneity, and endogenous network eects and allows us to show how the elasticity of the extensive margin of trade with respect to tari rates may also be sensitive to distance and other bilateral trade factors, via multiple channels. Our theoretical model picks up where Krautheim (2012) left o. We account explicitly for exogenous xed export costs independent of and in addition to network spillovers. The main contribution of Krautheim (2012) was to develop a baseline Melitz-type model where export xed costs are endogenous and a function of the (endogenous) number of exporters from country i to country j in the representative industry (his sections 1-3). As Krautheim noted, the great advantage of his (baseline) model was to obtain a closed form solution. In his nal section 4, he notes, It is quite likely, however, that in reality some xed costs are entirely (or at least mainly) independent of the number of exporters (p. 33; italics added) and these independent and exogenous xed costs may inuence the elasticity of export xed costs with respect to the number of exporters. However, he does not provide a closed-form general equilibrium model of these inuences in his paper; this is the purpose of this section (and theory Appendix 2). Moreover, he concludes his last substantive section of the paper suggesting future empirical work should investigate the variability of trade-cost elasticities to changes in these exogenous (spillover-insensitive) xed export cost determinants. Such empirical work is another potential contribution of our paper. Our theoretical model (below and in Appendix 2) solves for the closed form solutions and structural gravity equation of this Melitz-type model allowing for exogenous xed export costs (linearly) independent of the endogenous (network-spillover-based) xed export costs. 18 This model yields a gravity equation with heterogenous EIA partial eects (on the extensive margin of trade) that are functions of factors inuencing exogenous xed export costs. We assume a world economy with N countries with L j denoting the (internationally immobile) population and labor force in country j. We assume a single aggregate industry 18Appendix 2 provides a non-trivial extension of the Krautheim's baseline model to include additively separable exogenous and endogenous export xed costs, and provides closed form solutions to a general equilibrium model with labor-market clearing and free entry and exit of rms. 21

22 with heterogeneous rms each producing dierentiated products under increasing returns to scale. Consumers (workers) are identical and have the utility function: ( U = ω Ω ) σ q(ω) σ 1 σ 1 σ dω (15) where q(ω) denotes the quantity consumed of product ω from the set of varieties Ω available and σ is the elasticity of substitution in consumption across varieties (σ > 1). Consumers maximize utility subject to a standard income constraint yielding a demand function in country j for variety ω imported from country i: q (ω) = ( ) σ ( ) p (ω) wj L j P j P j (16) where P j = [ ω Ω p(ω)1 σ dω] 1 1 σ, wj is the wage rate in country j, and hence w j L j is aggregate income in country j. Firms in country i are assumed to have heterogeneous productivities. Potential entrants face a xed entry cost, F E i. In order to sell in a foreign market j, a rm has to pay an additional xed export cost, F X. 19 We assume furthermore that xed export costs F X be decomposed linearly into xed costs associated with (what we term) natural (or nonpolicy) impediments into markets (such as marketing and distribution costs), F XN, and xed export costs associated with the destination market's trade policy impediments (such as regulatory costs), F XP. We assume that the costs for a rm with productivity ϕ in country i to sell q units of output in country j facing (gross) ad valorem iceberg variable trade costs τ (hence, assuming τ 1) is given by: can c(q ) = w iq τ ϕ + w j (F XN + F XP ) (17) 19 Krautheim (2012) uses instead C to denote these xed costs. 22

23 Facing demand curve equation (16), the price charged in j by a rm in i is given by: p (ϕ) = w iτ ρϕ (18) where ρ = (σ 1)/σ. Up to now, our economy is characterized by a standard Melitz model. Following Krautheim (2012), we introduce network eects into the natural xed export costs (F XN ). Distinct from the baseline model in Krautheim (2012), we assume that natural xed export costs are determined by an exogenous component (A XN ) and an endogenous component reecting network eects (n η ). We note now that we will demonstrate that the presence of exogenous xed export component A XN can inuence the eect of an EIA with or without network spillovers; moreover, network spillovers can inuence the eect of an EIA on trade without exogenous xed export costs. As in Krautheim (2012) we assume that the xed costs of selling a product from i to j are inversely related to the number of rms in i selling in j, n, which itself is endogenous to the model. Fixed costs are assumed to be: w j (F XN + F XP ) = w j [A XN + n η + F XP ] (19) where η is the elasticity of xed costs with respect to the number of rms in i exporting to j (as in Krautheim, 2012). Furthermore, we assume the exogenous component determining xed costs, A XN, is inuenced by geographic, legal, institutional, and cultural factors such as bilateral distance and the presence or absence of common land borders, ocial languages, predominant religions, colonial histories, and legal origins. The motivation for such factors inuencing exogenous xed exports costs is Helpman, Melitz, and Rubinstein (2008), or HMR, whereby these factors signicantly inuenced the probability of a country-pair exporting (i.e., the selection eect) and such factors just noted are arguably exogenous to a rm's xed export costs. 23

24 In this setting, the prots of rm ϕ in i to export to j (π ) are: π (ϕ) = Max 0, ( wi τ ρϕp j ) 1 σ w j L j σ w j[a XN + n η + F XP ] (20) Firms in i will choose to export as long as prots are positive. The marginal exporter from i to j, where prots approach zero, denes the cuto productivity (ϕ ): ( wi τ ρp j ) 1 σ w j L j σ (ϕ ) σ 1 = w j [A XN + n η + F XP ] (21) In Krautheim (2012), without the additive exogenous xed costs A XN + F XP, one can easily solve for the cuto productivity ϕ. However, the presence of the additive factor A XN +F XP makes the determination here of ϕ more complex. The closed form solution is provided in theory Appendix 2. We nd several additional theoretical insights beyond Krautheim (2012) by using the additive version of exogenous and endogenous xed export costs. As in that study, we assume a Pareto distribution for productivities, with the measure of productivity heterogeneity given by γ and let ϕ be the lower bound of the support of the Pareto productivity distribution. First, we nd that the eect of changes in variable trade costs (τ ) on the export cuto productivity is no longer parametric but is related to the share of exogenous export xed costs in total export xed costs: d ln ϕ = 1 1 γ ηs d ln τ (22) σ 1 where s = (A XN δ η i (ϕ / ϕ) γη + F XP ) + δ η i (ϕ / ϕ) = 1 γη 1 + AXN +F XP n η (23) and δ i = σ 1 L γσf E i (24) i 24

25 Note that s and exogeneous export xed costs are inversely related. In the case of Krautheim (2012) without the additive export xed costs (when A XN = 0), equation (22)'s coecient on d ln τ simplies to 1/[1 + F XP γ σ 1 η]. Note that a variable trade-cost (τ ) decline directly lowers the export cuto productivity (ϕ ) but also indirectly lowers ϕ by increasing the number of exporting rms (n ), consequently expanding the network eect and further lowering this cuto productivity. Note also that the presence of the additive exogenous xed export costs (A XN ) better ensures the positive relationship between τ and ϕ by scaling down the value of γη, imposing lower restrictions on the value of η. We show in Ap- σ 1 pendix 2 that the stability condition for the determination of the export cuto productivity is γ σ 1 ηs < 1. Hence, if there is a stable cuto productivity, the eect of d ln τ on d ln ϕ is positive. Finally, as shown in Appendix 2, the number of exporting rms from i to j is a function of the export cuto productivity: n = σ 1 L γσf E i (ϕ / ϕ) γ = δ i (ϕ / ϕ) γ (25) i Second, we can solve explicitly for the eect of exogenous xed natural export costs (A XN ) on the export cuto productivity: d ln ϕ = 1 (σ 1)(1 γ ηs σ 1 )(A XN + F XP + n η ) daxn (26) If the stability condition is met, then the relationship between da XN The eect of lower exogenous xed export costs A XN and d ln ϕ is positive. on decreasing the cuto productivity is enhanced, relative to Krautheim (2012), due to s < 1. Also, the introduction of A XN (and hence s ) imposes less restriction on the value of η needed to satisfy the constraint that γ/(σ 1) >1. Moreover, the eect of A XN level of exogenous xed export costs. on ϕ is diminished the larger is the existing The third result from our extension has important implications for estimation of gravity equations. Gravity equations are typically specied in multiplicative form in levels of vari- 25

26 ables or linearly in the logs of variables. Historically, coecient estimates have been assumed to be parameters, cf., equation (1) above. Using our theoretical model, we can show that the elasticity of bilateral trade (on the extensive margin) to a one percent change in ad valorem trade costs (τ ) is sensitive to the relative importance of exogenous xed export costs (A XN + F XP ) in total xed export costs: ( ) 1 ηs d ln X = γ 1 γ ηs d ln τ (27) σ 1 Recall that s is inversely related to exogenous export xed costs A XN + F XP. It is important to note that despite s being in both the numerator and denominator of the term in parentheses the sign of the coecient on d ln τ is unambiguously negative if, as in Krautheim (2012), we assume hence, η < 1. With γ γ η < 1. First, recall that σ 1 σ 1 > 1 for nite variance; γ σ 1 η < 1, the denominator is positive since by equation (23) s < 1. Second, since η < 1 and s < 1, the numerator is positive. Third, since in s, and consequently ηs, is magnied by γ σ 1 γ σ 1 > 1, any rise > 1 in the denominator, so that the term in parentheses must rise unambiguously; we provide this in Appendix 2. Fourth, although changes in A XN export costs A XN alter ϕ endogenously, given equation (23) any reduction in exogenous xed + F XP relative to endogenous xed export costs n η is likely to raise s. Hence, factors (unrelated to market size) that reduce exogenous xed export costs A XN, such as common cultural identities, relative to factors that alter network externalities, such as labor endowment size in exporting countries, are likely to increase the (absolute value of the) trade-cost elasticity of trade ows. 20 A fourth result is that we can decompose the log change in the trade ow into the log changes in the intensive margin (IM) and extensive margin (EM). We show in Appendix 2 that the intensive margin eect is: d ln IM /d ln τ = (1 σ) (28) 20 Note in equation (27) that if s = 1 (no exogenous export xed costs), the comparative static simplies to that in Krautheim (2012). Furthermore, if η = 0, it simplies to that in the Melitz model, γ. 26

27 and the extensive margin eect is: ( ) 1 ηs d ln EM /d ln τ = γ 1 γ ηs (1 σ) (29) σ 1 Fifth, even if network externalities did not exist, declines in xed export costs associated with EIAs (d ln F XP ) could lead to heterogeneous EIA eects. We show in Appendix 2 that: ( d ln EM /d ln F XP γ (σ 1) = σ 1 F XP F XP + A XN ) (30) implying the elasticity of EM with respect to F XP is sensitive to the level of A XN, where the latter may be sensitive to observable bilateral trade impediments. Consequently, it is important to emphasize that the heterogeneity in EIA eects due to export xed costs is associated with two channels, endogenous and exogenous export xed costs. 5 From Theory to Empirics: Hypothesized Observable Determinants of Heterogeneous EIA Eects So what observable factors might inuence unobservable non-policy variable and xed trade costs (f t and A XN t, respectively) discussed above? Beginning with Tinbergen (1962), the empirical gravity equation literature provides 50 years of econometric examination of observable non-policy (or natural) bilateral variables that likely aect trade ows via bilateral trade costs. Typical variables that have surfaced over decades are bilateral distance and dummy variables for common land border, primary language, primary religions, legal origins and colonial histories, cf., Head and Mayer (2013). Up until 2000, this literature has interpreted the channel of inuence of these variables on trade ows as the intensive margin. However, two pertinent issues suggest that some or all of these six what we will term standard gravity covariates might inuence xed trade costs. First, the trade literature since 2000 has called considerable attention to the theoretical importance of xed export costs for 27

28 explaining zeros in trade. Second, Helpman, Melitz and Rubinstein (2008, or HMR), Egger, Larch, Staub and Winkelmann (2011, or ELSW), and Baldwin and Harrigan (2011, or BH) have shown empirically that some of these six variables actually explain the extensive, as well as intensive, margin of trade. However, they also reveal that there are quantitative as well as qualitative dierences in the impacts of these variables on the two margins of trade. For instance, bilateral distance negatively inuences both the probability and volume of trade in all three studies. However, contiguity of nations (i.e., sharing a common land border) inuences positively the intensive margin, but negatively the extensive margin, in HMR and ELSW; contiguity was omitted in BH. Hence, we look to observable determinants of exogenous variable and xed trade costs to explain empirically bilateral variability of f and A XN, key factors in explaining heterogenous EIA eects in the context of our modied Melitz-type model. In this section, we oer six conjectures three strong and three weak about how these six observable factors may inuence unobservable variable and xed export costs, and consequently the elasticity of bilateral trade, the intensive margin, and the extensive margin to an EIA. Since the world is not so generous as to provide formal linkages from each of these six factors to natural xed export costs A XN and natural variable export costs f, we base our three strong conjectures upon results that are robust across HMR (Tables I and II) and ELSW (Tables 2 and 3) and we base our three weak conjectures upon results in HMR and ELSW for which there is less support; BH only included distance and language. The strong conjectures are for (bilateral) distance, adjacency, and language; the weak conjectures are for religion, colonial history, and legal origins. First, bilateral distance likely aects both variable and xed export costs positively. HMR and ELSW found economically and statistically signicant negative eects of distance on the probability of positive exports from i to j and on their level. 21 Regarding the intensive margin, our equation (25) suggests that the elasticity of the intensive margin of trade to the 21BH also found a negative eect on the extensive margin. Bernard, Redding and Schott (2011) also found negative relationships between bilateral distance and the intensive and extensive margins of rms and products. 28

29 decline in tari rates from an EIA may be sensitive to f. Hummels and Skiba (2004) and Hummels (2007) provide evidence that f and bilateral distance are strongly positively related, likely due to transportation costs; consequently, we expect the EIA elasticity of the intensive margin to be negatively correlated with bilateral distance. The tari-rate elasticity of the extensive margin is likely also to be negatively related to distance, via two channels. Bilateral distance has also been shown to be positively related to information costs, which could inuence A XN. Equation (29) suggests that a reduction in A XN (as a result of lower distance) might increase s, which tends to increase the (absolute) tari-rate elasticity of the extensive margin. Equation (30) suggests an additional channel independent of network externalities through which smaller distance lowers A XN the (absolute) EIA elasticity of the extensive margin. and consequently raises Second, one of the most common ndings in the large early literature using gravity equations is that other things constant adjacency of two countries increased their trade. However, most of those studies ignored zeros in international trade. HMR and ELSW both found that while adjacency increases the level of trade conditioned on positive trade adjacency of two countries decreases the probability of positive exports; these results were statistically signicant. We believe that physical adjacency of two countries should tend to lower variable export costs, f ; consequently, we expect a positive coecient sign for adjacency through the intensive margin channel. However, the negative common land border coecient for the probability of exporting, found in HMR and ELSW, is a bit of a mystery. HMR interpreted the negative eect of a common land border dummy on the probability of trade as adjacent countries have typically had more military border conicts; it is the case that measures of cumulative duration of wars between countries negatively inuence trade. However, we oer an alternative explanation for this negative impact. The wellknown McCallum (1995) border puzzle showed that an international border between U.S. states and Canadian provinces diminished trade, controlling for distance. It is possible that the dummy variable capturing a common land border is reecting this international border eect. If so, adjacency between two countries capturing the international border eect 29

30 could raise export xed costs and diminish the probability of trade. Consequently, we expect that sharing a common land border should increase A XN, decrease s and F XP /(F XP + A XN ), and decrease the impact of an EIA between i and j on their extensive margin. Third, HMR and ELSW found that one of the cultural factors that had the largest positive (and statistically signicant) impact on the probability of positive exports from i to j was having a common language. Intuitively, sharing a common language should reduce the costs of establishing a trade relationship. Historically in gravity equations (ignoring zeros), common language has been an economically and statistically signicant positive contributor to explaining trade ow levels as well. Interestingly, HMR and ELSW found that with their preferred two-stage estimation language had a statistically signicant positive eect on the probability of trade, but economically and statistically weaker eects on the level of trade (conditioned on positive trade), cf., HMR, Table II and ELSW Tables 2 and 3. Like HMR and ELSW, BH found a positive impact of language on the probability of trade. We expect that sharing a common language should reduce exogenous natural xed export costs A XN, increase s in equation (29), increase F XP /(F XP impact of an EIA between i and j on the extensive margin. + A XN ) in equation (30), and increase the We posit weaker conjectures for religion, legal origins, and colonial histories, largely due to the absence of religion and legal origins in ELSW and BH and ambiguous results for colonial history. Fourth, several earlier gravity equation analyses of trade ows (ignoring zeros) included a dummy for common religion and generally found positive eects on bilateral trade. One of the intriguing results of HMR was the nding using two-stage estimation that common religion had an economically and statistically signicant eect on the probability of positive exports from i to j, but no statistically signicant eect on the level of exports from i to j (conditioned on positive trade), cf., HMR, Table II. HMR consequently relied on common religion as their exclusion variable for their second-stage regression. Like sharing a common language, sharing a common religion creates a trust (and a network eect independent of numbers of exporters) that lowers exogenous natural xed export costs A XN. We expect that sharing a common religion should reduce A XN 30, increase s and F XP /(F XP + A XN ),

31 and increase the impact of an EIA between i and j on their trade ow. Fifth, another common nding in the large earlier literature using gravity equations (but ignoring zeros) is that two countries with a common legal origin tend to trade more. In HMR's two-stage estimation, while common legal origin retained its economically and statistically signicant positive eect on the intensive margin, its eect on the country extensive margin was much smaller quantitatively. ELSW and BH omitted a common legal origins dummy. However, one of the main purposes of an EIA is to harmonize legal issues (such as regulations) between members. Hence, the benets of an EIA are likely to be greater if two countries have dierent legal origins. Consequently, we hypothesize that sharing a common legal origin will tend to reduce the positive eect of an EIA on trade via the extensive margin. Sixth, in earlier gravity equation analyses ignoring zeros, sharing a common colonial history tended to increase trade, other things equal. However, HMR found in two-stage estimates that this relationship only held up for the level of trade conditioned upon positive trade. An interesting HMR nding is that sharing a common colonial history has no signicant positive or negative eect upon the probability of trade. A priori one could argue that such ties might reduce exogenous natural xed export costs and thus raise the probability of positive trade. However, one could also argue that such ties may have led to legal harmonizations that might have reduced the benets of an EIA, much like having common legal origins. Consequently, we hypothesize that sharing a common colonial history will tend to reduce the positive eect of an EIA on trade via the extensive margin. Hence, we use six standard gravity covariates to explain the heterogeneity in EIAs eects on members' trade ows beyond what can be explained by the depth of the EIAs' various liberalizations. Table 3 summarizes succinctly the expected eects of each of the variables on enhancing or reducing the trade, intensive margin, and extensive margin impacts of an EIA. In principle, we could account for many other economic, political, legal, and cultural dierences for explaining the eectiveness of an EIA on two members' trade; our goal here is to provide an initial empirical investigation of standard gravity covariates based upon our theoretical framework. Moreover, we will nd later that these relationships potentially 31

32 provide guidance for ex ante analysis of the eects of EIAs on economic welfare. 6 Econometric Specication and Empirical Results The rst section of this part presents the main econometric specication. The second section presents the main empirical results. The third, fourth, and fth sections present robustness analyses. Data sources were presented earlier. 6.1 Main Econometric Specication Following our discussion in section 2.2 earlier, the main specication introduces interaction terms motivated by our theoretical model. Our theory suggests that the eect of an EIA on bilateral trade is non-linear in factors inuencing variable and xed bilateral export costs. All interaction terms enter as de-meaned values. Consequently, our main specication is: ln X t = α + η it + θ jt + +ψ + β 0 EIA t + β 1 (EIA t ln DIST ) + β 2 (EIA t LANG ) + β 3 (EIA t RELIG ) + β 4 (EIA t ADJ ) + β 5 (EIA t LEGAL ) + β 6 (EIA t COLONY ) + ζ t (31) where ln DIST is the (de-meaned) natural logarithm of bilateral distance between i and j, LANG is a dummy assuming the value 1 if i and j share a common ocial language and 0 otherwise, RELIG is a dummy assuming the value 1 if i and j share a common ocial language and 0 otherwise, ADJ is a dummy assuming the value 1 if i and j share a common land border (are adjacent) and 0 otherwise, LEGAL is a dummy assuming the value 1 if i and j share common legal origins and 0 otherwise, and COLONY is a dummy assuming the value 1 if i and j share a common colonial history and 0 otherwise. 32

33 6.2 Main Empirical Results with EIA Interactions Table 4 provides the results of estimating equation (31), where we introduce EIA interaction terms using the de-meaned values of the variables. Equation (31) is estimated using OLS and panel data for every ve years from Column (2) of Table 4 provides also the expected coecient signs based upon Section 5's hypotheses, summarized in Table 3. Column (3) provides the coecient estimates for equation (31). Our EIA variable is our benchmark categorization, EIA3-6, described earlier. 23 Columns 2 and 3 in Table 4 reveal that the coecient estimates all have the expected qualitative eects, based upon the hypotheses noted above and discussion there. First, note the fall in the partial eect independent of the other variables (the rst line of coecient estimates in column 3). This reects the partial eect of an EIA if the country-pair was at the mean of all these bilateral variables. However, for example, most EIAs are regional in nature hence among close countries and share common cultural characteristics (such as primary languages and religions). So the EIA's eect with interactions must be evaluated for its particular pair's characteristics. Second, note that all six interaction eects are statistically signicant. Third, as noted, the interaction terms' eects are as expected. Countries that are closer have larger EIA (partial) eects; this is consistent with our hypotheses regarding the intensive and extensive margin eects. However, countries that are adjacent have a lower EIA eect. This is consistent with adjacency having a positive eect on the intensive margin of trade, but simultaneously having a negative eect on the extensive margin of trade due to higher export xed costs associated with an international border eect. Language has a positive eect on the EIA's impact; this is consistent with our hypothesis that a common language lowers xed export costs. Consider now our weaker conjectures. We nd that sharing a common religion increases 22See BBF (including their Online Appendix) on rationales for using OLS and rationalization of the benets of ve-year intervals. Potential biases due to issues of selection into exporting and rm heterogeneity are accounted for. The issue of potential heteroskedasticity bias is not accounted for using estimators such as Poisson Quasi-Maximum Likelihood (PQML) or Gamma; we address this later. 23It is important to recall that in the presence of the country-pair () xed eects the (separate) level eects of ln DIST and the other time-invariant bilateral (dummy) variables on X t are subsumed in the country-pair xed eects. 33

34 an EIA's impact on trade, likely by reducing exogenous xed export costs. Our results in column (3) of Table 4 are also consistent with expectations for common legal origins and colonial history. HMR found that common legal origins increased the probability of two countries trading. We nd that the eect of an EIA is diminished if two countries have common legal origins. Recall, we expected that an EIA which not only lowers ad valorem tari rates but often harmonizes nontari regulations between countries may have less of an eect on trade if the countries have similar legal origins. 24 With similar legal origins there may be less for two countries to gain from an EIA, since less legal harmonization via the EIA is obtained. In other words, having a common legal origin may be reducing the level of F XP, potentially osetting any eect of common legal origins on lowering A XN. We also nd that sharing a common colonial history decreases the impact of an EIA on trade. It is possible that similar to having common legal origins a common colonial history may reduce the benets of an EIA because the existence of earlier preferences may have reduced the level of F XP, potentially osetting the impact of a lower A XN In the next three sections, we pursue a sensitivity analysis. on the EIA's impact. Our robustness analysis addresses three issues for which the results may be sensitive. First, Table 2 introduced lagged eects of EIAs on trade ows. We investigate whether our interaction results are sensitive to lagged eects. Second, BBF introduced a decomposition between the intensive and extensive product margins using the decomposition technique in Hummels and Klenow (2005). We investigate whether the results for aggregate trade, the intensive margin and the extensive margin are consistent with our hypotheses. Third, Table 2 introduced six categories of types of EIAs. We investigate whether the results hold up for each type of EIA (and hence dierent degrees of liberalizations, i.e., dierent levels of reductions in τ ). For the rst two robustness analyses, our EIA categorization is EIA3-6, as used before. 24See Horn, Mavroidis, and Sapir (2010) on the various provisions of EIAs of the European Union and United States with trade partners. 34

35 6.3 Lagged Eects In Table 2, we found that EIAs have lagged eects on trade ows. Allowing for lags, in many cases the complete eects of EIAs were augmented. In this section, we augment the model in equation (31) to include ve-year lags of the RHS variables. Column 4 in Table 4 includes additionally lagged values of each of the RHS variables. Importantly, the addition of one ve-year lag reduces the size of the sample, so coecients are not directly comparable. We note several results. First, the eect of an EIA at the means of all the other RHS variables for the current period falls only slightly from 0.25 to 0.24, but remains statistically signicant. The lagged eect of 0.07 is statistically insignicant. The addition of a lag increases the total eect of distance on the partial eect from to and the lagged eect is statistically signicant. The addition of a lag increased the total negative eect of adjacency on the eect of an EIA. The addition of a lag had no material eect on the partial eect of common language on EIA's eect; the coecient estimate fell from 0.32 to 0.30 and remained signicant. For religion, the eect in the current period is diminished, but still statistically signicant; the lagged eect is economically and statistically insignicant. The eect of common legal origins on the partial EIA eect was marginally signicant in column (3); its eect becomes statistically insignicant in column (4). Finally, the eect of common colonial ties on the partial eect of EIA3-6 remains virtually unaltered from introducing a lag and the coecient estimate on the concurrent level remains statistically signicant. Since the results are quite similar with and without lags, for the remainder of the analysis we will ignore lags; this will preserve the larger number of observations and also facilitate computation later when we separate EIA eects by the six types of agreement. 6.4 Intensive and Extensive Margins Our theoretical framework suggests that the partial eects of EIAs on trade ows are potentially inuenced at both the intensive and extensive margins. Section 3 used a simple Armington model to show that the partial eect of an EIA on the intensive margin of trade 35

36 may be inuenced by natural variable export cost factors (inuencing f ). Section 4 used a Melitz-type model to show that the partial eect of an EIA on the extensive margin of trade may be inuenced by natural xed export cost factors (inuencing A XN ). Is it possible to determine the factors inuencing f versus those inuencing A XN? Are they the same factors? Are they dierent ones? In this section, we use the methodology employed in BBF to determine these relative inuences. BBF combined the approach in BB with the Hummels and Klenow (2005) margin-decomposition methodology to determine the impact of EIAs on aggregate trade, the intensive margin, and the extensive margin. However, in this paper, we estimate heterogeneous EIA partial eects. Hummels and Klenow (2005), or HK, was the rst paper to highlight a tractable method for decomposing transparently the extensive and intensive goods margins of trade for a large set of countries' bilateral trade ows using publicly available disaggregate trade data. 25 Let X t denote the value of country i's exports to country j in year t. Following HK, the extensive margin of goods exported from i to j in any year t is dened as: EM t = m M t X m W jt m M W jt X m W jt (32) where X m W jt is the value of country j's imports from the world in product m in year t, M W jt is the set of all products exported by the world to j in year t, and M t is the subset of all products exported from i to j in year t. Hence, EM t is a measure of the fraction of all products that are exported from i to j in year t, where each product is weighted by the importance of that product in world exports to j in year t Studies have also used country-specic data on individual plants (or rms) to study the extensive and intensive rm margins of trade liberalization, but such studies have necessarily been conned to particular countries because such data is widely known to be much more costly to access and such data sets have not been concorded for international comparisons, as noted in Helpman, Melitz, and Rubinstein (2008). See Eaton, Kortum, and Kramarz (2008) for a study of French rms, Treer (2004) for a study of Canada and the United States, and Pavcnik (2002) for a study of Indian rms. Another relevant theoretical and empirical piece with similar overtones is Arkolakis, Demidov, Klenow, and Rodriguez-Clare (2008). 26Alternatively, one could use an unweighted average, which would then be simply the fraction of all products exported from i to j. However, HK as well as researchers since then use the weighted average. A weighted average seems more appropriate since cars and pencils do not have the same values in trade. 36

37 HK dene the intensive margin of goods exported from i to j as: IM t = m M t X m t m M t X m W jt (33) where X m t is the value of exports from i to j in product m in year t. Thus, IM t represents the market share of country i in country j's imports from the world within the set of products that i exports to j in year t. One of the notable properties of the HK decomposition methodology is that the product of the two margins equals the ratio of exports from i to j relative to country j total imports: EM t IM t = m M t X mt m M W jt X W jmt = X t /X jt (34) where X jt denotes j's imports from the world. Taking the natural logs of equation (32) and some algebra yields: ln X t = ln EM t + ln IM t + ln X jt. (35) Consequently, the HK decomposition methodology yields that the log of the value of the trade ow from i to j in any year t can be decomposed linearly into (logs of) an extensive margin, an intensive margin, and the value of j's imports from the world. We note three issues regarding the HK methodology. First, the term ln X jt will be subsumed in the importer-time xed eect. Second, HK applied their methodology to only a cross section. By contrast, we are applying it to a time series of cross sections. Consequently, the trade weights used in constructing EX t and IM t will likely vary from year to year. To address this, BBF also considered in a sensitivity analysis a chain-weighting technique. However, their results were robust to this alternative weighting procedure. Third, there are numerous zeros in the variables and the results may be biased by ignoring the existence of rm het- Also, since we will use a time series of cross sections, we will consider later two alternative methods for xing the trade-share weights over time. 37

38 erogeneity. However, as discussed in BBF and their Online Appendix, our panel approach largely alleviates sample-selection bias and rm-heterogeneity bias as raised in Helpman, Melitz, and Rubinstein (2008). 27 Table 5 provides the results using EIA3-6 as our EIA dummy variable. Table 5 has seven columns with column (1) providing the variable names. Columns (2), (4) and (6) provide the expected coecient signs for the interaction terms for aggregate trade, the intensive margin, and the extensive margin, respectively, guided by our hypotheses in Section 5 and summarized in Table 3. Several results are worth noting. First, with the exception of the adjacency interaction term, the coecient estimates for the interaction terms in the aggregate trade regression column (3) not only conform to expectations, but also are qualitatively identical to and quantitatively similar to those in Table 4 column (3), despite a large loss in number of observations due to the zeros issue (associated with imposing a cuto in the data for traded vs. nontraded goods). Second, note that for the intensive margin results in column (5) the only two variables that have signicant eects on the EIA (partial) eect are bilateral distance and the adjacency dummy. Interestingly, our theoretical model in section 3 motivated heterogeneous intensive margin EIA trade eects via the sensitivity of the tarirate elasticity to levels of freight costs, f. The two variables of the six in Table 5 that are likely most related to freight costs are bilateral distance and adjacency; the other four variables are likely more related to information costs and consequently natural xed export costs, A XN. Third, note the importance of all six interaction terms for inuencing the partial EIA eect on the extensive margin of trade. Fourth, note the qualitatively dierent eects of the adjacency interaction term on the intensive margin's EIA eect versus the extensive margin's EIA eect. Adjacency has a signicant positive eect on the EIA intensive margin partial eect, consistent with the inuence of f on the EIA's eect. By contrast, adjacency has a signicant negative eect on the EIA extensive margin partial eect. This is consistent 27See BBF and their Online Appendix. Besedes and Prusa (2011) emphasize that extensive margin (intensive margin) eects may be overstated (understated) in examining eects of liberalizations using panel data, due to survival issues. Addressing this issue is beyond the scope of this particular paper. However, our results may not be biased excessively by this issue since we nd material intensive margin eects from EIAs. 38

39 with HMR which found that adjacency has a negative eect on the probability of trade, or the extensive country margin. Our result here for adjacency is consistent with our hypothesis that an international border may be an impediment to the extensive margin of trade. Figure 2 provides density plots illustrating the heterogeneous EIA (partial) eects for total trade (solid line), the intensive margin (dashed line), and the extensive margin (dotted line). Several points are worth noting. First, the dashed line shows the enormous dispersion of EIA total trade eects from -0.1 to 1.5, attributable to dierent geographic, cultural, and institutional characteristics shared by country-pairs likely associated with bilateral variable and xed export costs. The mean eect is slightly greater than 0.5, consistent with earlier ndings. Second, the other two lines indicate that the average intensive margin eect is less than the average extensive margin eect. Third, there is more dispersion in the extensive margin eects relative to the intensive margin eects. This may be attributable to more of the cultural and institutional factors having signicant inuence on the extensive margin eects relative to the intensive margins, cf., Table 5. Finally, as raised in Kehoe and Ruhl (2013), the eect of EIAs on the extensive and intensive margins of international trade ows are sensitive to the choice of cuto values determining traded from nontraded goods. As noted there, to characterize an extensive margin, one needs a denition of a nontraded good. Kehoe and Ruhl (2013) show for many trade liberalizations that using even an absolute cuto of 50,000 US dollars there was no extensive margin impacts of EIAs. Using their relative cuto approach, some country pairs' cuto for nontraded goods is several millions of US dollars, cf., Kehoe and Ruhl (2013, Table 7). We have also estimated the results discussed above using cutos of 25,000, 50,000, 100,000, 250,000, and 500,000 US dollars, in addition to the 1 million US dollar cuto used for Table 5. Table 6, for instance, provides the results using the 100,000 US dollar cuto. With regard to the statistically signicant coecient estimates, the results between the two tables are fundamentally the same, with the exception of the religion interaction term. In Table 5, religion has a signicant impact on EIA's extensive margin eect; by contrast, in Table 6 religion has instead a signicant impact on EIA's intensive margin eect. 39

40 6.5 Interactions by Type of EIA We showed in Table 2 that our approach could be applied using each of the six EIA types in the sample. We found that EIA partial eects were smaller for types of agreements with less trade liberalization, as expected. In this section, we investigate whether the interaction terms have the expected eects by EIA type. Thus, we determine here empirically whether there are heterogeneous impacts of EIAs at each level of degree of trade liberalization. Tables 7 and 8 present results using a USD 1 million cuto for nontraded goods. For robustness, Tables 9 and 10 present results using an alternative USD 100,000 cuto, as discussed just above. Table 7 presents rst the results without the interaction terms. Table 7's results benchmark this analysis against BBF. BBF found similar eects, although BBF grouped all customs unions, common markets and economic unions into one EIA type (CUCMECU). Here, we separate all six types of EIAs. The results for aggregate trade ows in column (3) of Table 7 correspond to the respective results in column (6) of Table 2, although here the sample size is smaller due to estimation also of intensive and extensive margin eects. Despite the considerable reduction in sample size from Table 2 to Table 7, the results are fairly similar. Notably, as in Table 2, there is a monotonic increase in the average (partial) EIA eect on trade in Table 7 with an increase in the degree of trade liberalization as one progresses from EIA1 to EIA5, with a slight drop from EIA5 to EIA6; even in with the larger sample in Table 2, there was a drop from EIA5 to EIA6. A distinguishing feature of the results for Table 7 is that we provide EIA eects also for intensive and extensive margins on all six EIA types; by contrast, BBF combined EIA4-6 into one category. Column (5) in Table 7 shows signicant EIA eects on the intensive margin for two-way PTAs and higher levels of integration. In fact, these partial eects increase monotonically with higher degrees of economic integration. Column (7) of Table 2 shows that EIA eects are signicant on the extensive margin for FTAs and higher levels of integration. There is not a monotonic increase in these partial eects by type of EIA; it turns out that the highest EIA eect on the extensive margin is 40

41 for customs unions. Table 8 provides the results by EIA type now including the interaction terms. Because of the large number of coecient estimates and t-statistics, we prioritize the results by common themes. First, we note that at the means of the trade-cost variables EIAs retain their economic and statistical signicance for total trade impacts for FTAs and higher levels of integration, and these partial eects increase monotonically with the degree of trade liberalization. Progressing from EIA1 to EIA6, these eects are -0.11, 0.01, 0.20, 0.61, 0.69, and Second, after accounting for dierences in levels of economic integration, the results for EIA-distance and EIA-adjacency interactions largely conrm results in Table 5 for the intensive margin EIA eects. For instance, the EIA-distance interaction's coecient estimate is statistically signicant for total trade for four categories: OW P T A, T W P T A, F T A, CM. Moreover, EIA-distance matters on the intensive margin EIA eect for all six EIA categories except customs unions. This accords well with the notion that distance raises f, which tends to lower the trade-cost intensive margin elasticity. Also, the other interaction variable that Table 5 suggested impacts the EIA intensive margin eect is adjacency, with a positive sign. Column (5) in Table 8 reports that the EIA intensive margin eect is aected positively by adjacency for FTAs and customs unions with statistical signicance; for common markets and economic unions, the eect is positive but not statistically signicant. Third, Table 8's column (7) reports the results of how the EIA extensive margin eect is inuenced by these interactions. Recall that in Table 5 all of the six interaction variables had coecient estimates that were statistically signicant with expected signs. For EIAdistance, we nd a negative and statistically signicant eect for FTAs and customs unions only. For EIA-adjacency, we nd expected negative eects for two-way PTAs and all higher levels of integration; however, we only nd a statistically signicant negative eect for FTAs. Both cultural variables have consistent eects on the EIA extensive margin (partial) eect. The coecient estimate for EIA-language is positive and statistically signicant for every EIA level except two-way PTA (though it is still positive for T W P T A). EIA-religion has a positive eect on the EIA extensive margin eect for every EIA level except OW P T A; it 41

42 is statistically signicant for T W P T A, F T A, CM, and ECU. EIA-legal has an expected negative coecient estimate for every category except customs union; it has statistically signicant negative coecient for OW P T A and F T A. Finally, EIA-colony had negative coecient estimates for all EIA levels except T W P T A and ECU; it is statistically signicantly negative for F T A, customs union, and common market. We also estimated the specications above using the alternative cuto of USD 100,000. The results analogous to those in Tables 7 and 8 just discussed are presented in Tables 9 and 10, respectively. For brevity, we will not provide a detailed discussion of these results as they are quite similar to those just discussed using the USD 1 million cuto. Regarding Table 9, the main notable dierence relative to the Table 7 results is, as expected, EIAs had larger impacts on the intensive margin when we reduced the nontraded good cuto value. This is consistent with our results for Tables 5 and 6, using the single EIA t dummy. Hence, EIA eects for extensive margin eects are smaller across types of EIAs. For example, in Table 7, the eects of common markets and economic unions were small but statistically signicant; in Table 9 they become statistically insignicant. Regarding Table 10, there are few changes relative to the Table 8 results than cannot be explained by the fact that with a lower nontraded good cuto there are larger impacts of the interaction variables on intensive margin EIA eects relative to extensive margin EIA eects. Table 10 is provided to corroborate this conclusion. Finally, as we did earlier for all EIAs, we present density plots of the trade, intensive margin, and extensive margin heterogeneous partial eects for FTAs, customs unions, common markets, and economic unions, in Figures 3-6, respectively. As before, we use the USD 1 million nontraded good cuto. The distinguishing feature of comparing the results is that the average extensive margin eects are larger than the average intensive margin eects for lower levels of trade liberalization, FTAs and customs unions. For common markets and economic unions, the average intensive margin eects are larger than the average extensive margin eects. The economic intuition for this result is intuitive. Deeper levels of economic integration, such as common markets and economic unions, have already likely overcome 42

43 export xed costs in earlier stages of integration. Consequently, it is the less liberalized EIAs such as FTAs and customs unions where the benets of having common cultural and institutional factors inuences to a larger extent the eect of an FTA or CU by reducing export xed costs. 7 EIAs and Welfare This section has three parts. First, we review the relationship between aggregate bilateral trade ows, trade creation, trade diversion, and economic welfare in a wide class of quantitative trade models and provide some novel empirical evidence using methodology in ACRC to show that trade diversion eects for the typical bilateral trade ow are very small relative to trade creation eects, suggesting that the bilateral trade-creation (average treatment) eects are likely highly correlated with economic welfare eects of EIA liberalizations. This sets the stage for showing that just two readily obtainable statistics the bilateral trade ow share (λ ) and the EIA partial eect ( β ) can largely explain (net general equilibrium) welfare eects of EIAs. Second, for robustness we establish the relationship between the probability of a pair of countries forming an EIA and the EIA partial eects. Third, we show that average treatment eects of EIAs on the treated (ATTs) have systematically declined over the past half century, consistent with the notion that the earliest EIAs in the post-world War II period were likely the most welfare improving. 7.1 Explaining Welfare Changes with Two Bilateral Statistics Theory As introduced in Section 2, ACRC showed in the context of many standard and inuential quantitative trade models that changes in economic welfare of representative country j (W j ) 43

44 are a function of two sucient statistics: 28 d ln W j = 1 ɛ d ln λ jj (36) where λ jj is the importer's share of expenditures on domestic output (λ jj = X jj /Y j ) and ɛ is the trade-cost elasticity (ɛ < 0). The intuition is that welfare changes from trade liberalizations in country j only depend on terms-of-trade changes. However, terms-oftrade changes can be inferred from changes in the relative demand for goods from dierent countries. While equation (36) suggests a simple way to quantify welfare changes using only two factors, empirically the exercise is not trivial. First, a large sub-literature in international trade has concerned itself with nding consistent and precise estimates of ɛ. While progress has been made, rm knowledge of ɛ is still remote. We still lack consistent and precise estimates of ɛ. Second, and seemingly less of a problem, one might think that measurement of λ jj = X jj /Y j is straightforward; this is not true. While customs oces in every country provide excellent reporting of X for goods ows, calculation of X jj /Y j is subject to several potential measurement issues. First, measures of X are essentially of aggregate goods trade ows (based upon underlying disaggregate trade ow data via customs oces). Since trade ows are measured on a gross (not value added) basis, X jj should be calculated as gross output minus exports (the sum of X jk over all k foreign markets). However, j's GDP, or Y j, is measured on a value-added basis, so that X jj and Y j are each measured on a dierent basis. Second, j's GDP includes goods, services and structures value added; however, trade ows on services are poorly measured. These problems can be resolved by disaggregated data. For instance, manufactures output and trade ows can be used; construction of variable X jj for manufactures does not face these measurement issues. Yet, such data is not as readily available and time-series of cross-sections of such data are quite limited. However, there is an alternative approach when ɛ and λ jj are not easily measured, such 28In their own paper, ACRC noted caveats to this strong result. Since then, other papers have shown important caveats as well. For brevity, however, we use the basic ACRC result. 44

45 as in our common context with aggregate goods trade ows. Following ACRC and their assumptions, we know that: N ln W j = ln P j = ln[ (w k τ kj ) ɛ ] 1/ɛ (37) k=1 where w k is country k's nominal wage rate (where country j's nominal wage rate, w j, is the numeraire as in ACRC). Dierentiating equation (37) with respect to a change in ln τ yields: N d ln Wj i = [λ d ln τ + λ kj d ln w k ] (38) k=1,k j where d ln W i j denotes the (log) change in country j's welfare from an EIA with country i. In our context, d ln τ ɛ = d ln e β EIA = β deia = β. Hence, equation (38) simplies to: d ln Wj i = 1 ɛ λ β N + ( λ kj d ln wk) ɛ (39) k=1,k j or d ln W i j = 1 ɛ λ β + φ (40) Equations (39) and (40) decompose the welfare eect into trade creation (the rst RHS term) and trade diversion (the second RHS term). While considerable advances have been made econometically using gravity equations for providing precise and consistent estimates of trade creation that is, EIA t on X t advances in estimation using gravity equations of the trade diversion eects has made little progress. A few studies have incorporated dummy variables to try to estimate trade-diversion eects. Beginning with Frankel (1997), some researchers have included a dummy variable if either the exporter i has an EIA with any country other than importer j (EIA it ) and a dummy variable if importer j has an EIA with any country other than exporter i (EIA jt ). Typically researchers have found if anything a negative eect of these dummies on X t. Ghosh and Yamarik (2004), Eicher, Henn, and Papageorgiou (2012), and Dai, Yotov, and 45

46 Zylkin (2014) all included dummies of this type. However, to date the empirical ndings largely indicate minimal trade diversion eects Econometric Issues and Empirical Results In this section, we estimate equation (40) using the welfare changes for importing country j from an EIA of i and j on the LHS and estimates of λ and of β on the RHS, where we treat φ as an error term. In a robustness analysis, we will try to account explicitly for variation in φ (= N k=1,k j λ kj d ln wk) ɛ using importer xed eects. Moreover, we will also later introduce exporter and importer xed eects as another robustness analysis. To anticipate the results, we will show that at least 85 percent of the variation in d ln W i j is explained by variation in λ and β. Hence, it is reasonable to treat φ as an error term. We now discuss the construction of all three variables. First, as mentioned, our empirical analysis earlier used interaction terms in a panel gravity equation to estimate 2,313 heterogeneous EIA partial eects, β. Second, d ln Wj i are generated from the following exercise. Using the Anderson and van Wincoop (2003) structural gravity model (cf., Head and Mayer (2013)), we calculate the level of welfare for the importing country j for all 2,313 observations for which we have EIAs; as standard, welfare is dened as the importer's nominal wage rate divided by the importer's CES price index. Given the actual trade ows and the model's structure, we can calculate each country's endogeneous wage rate level, w j, national incomes (Y j ), and importer's CES price index (P j ). For each of the 2,313 observations, we remove an EIA, say, EIA ; we then calculate the new bilateral trade ows, wage rates, national incomes, and importer's CES price index under each of the 2,313 new equilibria. This generates the data on d ln Wj i. Third, we combine the actual trade ows X with the initially computed values of Y j to calculate the λ. We estimate equation (40) using ordinary least squares (OLS). However, as in the gravity equation literature, the relationship between the variables of interest is multiplicative. For OLS, we follow the traditional gravity equation literature prior to Santos-Silva and Tenreyro 46

47 (2006) where we assume the error term, Φ, is multiplicative and rewrite equation (40) as: d ln W i j = 1 ɛ λ β Φ (41) Taking the logarithm of equation (41) yields a log-linear equation suitable for OLS: ln(d ln W i j ) = ln(1/ɛ) + δ 1 ln(λ ) + δ 2 ln( β ) + ln Φ. (42) where ln(1/ɛ) is embedded in the constant. Our theory suggests the hypothesis that δ 1 = δ 2 = 1. Table 11 reports the results of estimating equation (42) under four alternative specications. Specication (1) is equation (42), but constraining the coecients δ 1 and δ 2 to be equal. Column (3) shows that the coecient estimate for ln(λ β ) is positive and statistically signicant. Moreover, the coecient estimate of 0.92 is very close to the expected estimate of 1. Variation in ln(λ β ) explains 83 percent of the variation in ln(d ln Wj i ). In specication (2), we allow the coecient estimates for ln λ and ln β to be unconstrained. Column (4) shows that ln λ and ln β have positive and statistically signicant eects on ln(d ln Wj i ). Both variables explain 85 percent of the variation in d ln Wj i. This suggests that the correlation between the welfare eect and the direct partial eect is extremely strong. Naturally, we would expect a correlation because the welfare eect is a function of the partial eect; for an excellent discussion and quantitative analysis of the relationship of the partial eect to the welfare eect, see Head and Mayer (2013, section 4.3). Nevertheless, our result suggests that trade diversion plays a limited role empirically relative to trade creation in inuencing welfare. Finally, the coecient estimate on ln λ is economically very close to unity, as suggested by theory, though it is statistically dierent from unity (at the 1 percent signicance level). Specication (3) adds an importer xed eect to account for d ln X jj in equation (39). The R 2 value rises from 0.85 to 0.91 with the inclusion of the importer xed eect. More- 47

48 over, there is no material change in the estimated coecients relative to specication (2). For completeness, specication (4) includes an importer xed eect and an exporter xed eect. As for specication (3), the R 2 value rises from 0.91 to 0.94 with the inclusion of the importer and exporter xed eects. Once again, there is no material change in the estimated coecients relative to specication (2). On net, the results suggest that welfare changes from an EIA are well-approximated by the partial eect estimates of an EIA (ln β ), alongside measures of X /Y j (ln λ ). However, since the data used for the LHS variable in the regressions just reported (d ln Wj i ) are generated from a general equilibrium model that incorporates the partial eect estimate, we evaluate next the robustness of these results. We do this by examining the roles of ln β and ln λ for explaining an empirically generated measure of the potential welfare gain from an EIA between i and j, suggested by the methodology in Baier and Bergstrand (2004): probit estimates of the likelihood of an EIA. 7.2 Probability of an EIA and the EIA Partial Eect As just noted, one of the constraints of the previous welfare regressions is that the welfare changes are functions of the partial eects by construction. The purpose of the preceding analysis was to show that potential trade diversion eects played little role. However, there is another way to show that ln β and ln λ are useful and readily available variables for predicting welfare changes from an EIA. Baier and Bergstrand (2004) provided a framework for predicting the probability that a pair of countries would have an EIA. Based upon a general equilibrium model, the authors showed that the welfare of two countries' representative consumers would be enhanced by an EIA the closer they were to each other, the more remote they were from the rest-of-the-world (ROW ), the larger and more similar their economic sizes, and the larger their dierences in relative factor endowments. Following the qualitative choice model of McFadden (1975, 1976), they showed that these economic factors would also be related to the probability of having an EIA. Their results indicated that the 48

49 country-pairs that tended to have EIAs tended to have the economic characteristics consistent with such EIAs being welfare improving. Moreover, the econometric model predicted correctly 85 percent of the 286 EIAs in 1996 among the 1,431 country-pairs and predicted correctly 97 percent of the remaining 1,145 pairs with no EIA. The econometric framework we employ is the qualitative choice model of McFadden (1975, 1976). A qualitative choice model can be derived from an underlying latent variable model. For instance, let y denote an unobserved (or latent) variable, where for similcity we ignore the observation subscript. As in Wooldridge (2000), let y represent the dierence in utility levels from an action (the formation of an FTA), where: y = ς 0 + xς + ɛ (43) where x is a vector of explanatory variables (i.e., common economic characteristics), ς is a vector of parameters, and error term ɛ is assumed to be independent of x and to have a standard normal distribution. In the context of this model, y = min( U i, U j ). Hence, both countries' consumers need to benet from an EIA for their governments to form one. Since y is unobservable, we dene an indicator variable, EIA which takes the value 1 if two countries have an EIA (indicating y > 0), and 0 otherwise (indicating y 0). The response probability, P, for EIA is: P (EIA = 1) = P (y > 0) = G(ς 0 + xς) (44) where G( ) is the standard normal cumulative distribution function, which ensures that P (EIA = 1) lies between 0 and 1. In this context, we predicted the probabilities of country-pairs having EIAs for the nine years 1970, 1975,..., 2010 using similar economic characteristics. Although the pseudo-r 2 values range between 34 to 80 percent, the relationship between the economic characteristics with the probabilities are qualitatively very similar across the nine years, and are consistent 49

50 with ndings in Baier and Bergstrand (2004). As expected based upon the theoretical framework in Baier and Bergstrand (2004), the likelihood of an EIA is negatively related to distance, positively related to economic size and similarity, and positively related to having a common primary language and common primary religion. In the context of the Baier- Bergstrand model, the country-pairs that tend to have EIAs tend to have the economic characteristics consistent with such EIAs being welfare improving. The probit results are available in Appendix 3. Our goal in this section is to determine whether ln β and ln λ can also explain the variation in the probabilities of EIAs, which serve as proxies for the welfare gains of a country-pair from an EIA. Table 12 presents the results of ve alternative specications. The number of observations is limited to 2,360 as in the previous section, as these are the number of estimates of ln β from our earlier results. The rst specication, shown in column (3), regresses ln P (EIA t ) on ln λ, ln β, and a constant. Two results are worth noting. First, both variables have the expected qualitative relationship with ln P (EIA t ). Second, as ln P (EIA t ) serves as a proxy for d ln Wj i, we cannot assign a specic expected quantitative value for the coecients. Third, we note that the R 2 value is 53 percent. One possible concern, however, is that the ln β works well to explain the probability of an EIA because the ln β themselves will tend to be higher when variable and xed export costs are low, as our theory suggested. Consequently, ln β may have an economically and statistically signicant eect simply because ln β and ln P (EIA t are inuenced by common variables, such as bilateral distance, adjacency, etc. To address the robustness of our results, we considered several other specications. Column (4) adds the log of bilateral distance to the regression. Column (4) shows that while ln DIST helps to explain ln P (EIA t ) ln λ and ln β still have signicant explanatory power. Moreover, adding only distance increases the explanatory power from 53 percent to 85 percent. In the next sensitivity analysis, we included bilateral distance as well as all the other variables used earlier to explain variable and xed export costs (and which are determinants of the predicted probabilities). Column (5) shows that although all these observables are statistically signicant in explaining 50

51 ln P (EIA t ) ln λ and ln β still retain signicant explanatory power. Moreover, the coecient estimates for ln λ and ln β hardly change at all. The specication in column (6) adds an importer xed eect to control as in the earlier analysis for variation in trade diversion, φ. As shown in column (6), this has no material eect on the explanatory power of ln λ and ln β. Finally, for completeness, the specication in column (7) adds both an importer and exporter xed eect. Although the coecient estimate for ln λ becomes negative and signicant, the coecient estimate for ln β has no material change. 7.3 Timing of Average Treatment Eects on the Treated We consider one other approach to show that the partial eects of EIAs are consistent with welfare improvement from EIAs. As discussed in BB and BBF, one of the benets of using panel data to estimate partial eects is that it can account for the endogeneity of EIAs. However, one of the drawbacks of estimating partial treatment eects using panel data is that the treatment eect is assumed to be constant across all years; in our analysis, that is As now apparent, a main thrust of our paper has been to show that partial eects vary by country-pair; however, we have not allowed any variation in treatment eects over time. In this section, we examine in particular the average treatment eects on treated pairs (ATTs) across nine 5-year periods. We now provide the motive for estimating the ATT for each of nine 5-year periods. Following the logic of Baier and Bergstrand (2004), if country-pairs' governments are maximizing welfare of their consumers, then the EIAs that are formed will likely have economic characteristics consistent with welfare maximization. For instance, Baier and Bergstrand (2004) showed theoretically that if country-pairs are physically close and have large and similar GDPs, they will benet more from an EIA. Baier and Bergtrand (2009) showed for nine cross-sections 5-years apart ( ) that countries that are close and have large and similar GDPs tend to have EIAs. Moreover, Bergstrand, Egger and Larch (2014) show that the countries that are closer and have larger and more similar GDPs have higher prob- 51

52 abilities of forming EIAs sooner. If indeed country-pairs' governments are maximizing their consumers' welfare, then not only is the probability higher that they will form an EIA earlier but in the context of our paper the ATTs of the EIAs should be larger in earlier years than later years. In this section, we examine the ATTs from our panel data estimates. That is, using the methodology described earlier, we have estimates β for every country-pair that was treated. We then group the partial eects by the 5-year periods in which these EIAs formed: , ,..., We then take the average of the partial eects for each of these 5-year periods. These are what we term here the ATTs for each period. 29 These are reported in Table 13. Table 13 has four columns. The rst column provides the 5-year period in which all EIAs were formed in that period. 30 The second column provides the ATT for EIAs formed during that period. The third column provides the standard deviation of the partial eects during that period. The fourth column provides the number of estimated partial eects that were negative during that 5-year period. The most important nding is the following. The ATTs in earlier years are the highest. Moreover, the ATTs decline nearly monotonically over time. Just as Bergstrand, Egger and Larch (2014) found that the country-pairs that were closest, largest in GDP, and most similar in GDP three important economic characteristics found in Baier and Bergstrand (2004) to increase the welfare gains from an EIA formed the earliest, we nd the the ATTs are the highest among the earliest EIAs. This is consistent with other evidence that the most welfare improving EIAs were the earliest to form. A second notable nding is the small number of negative partial eects. While the 29Note that the approach here uses a parametric approach to derive ATTs. This diers from the nonparametric approach used in Baier and Bergstrand (2009). In that paper, the authors used nonparametric (matching) techniques to estimate ATTs. Such techniques were useful to identify in a time-series of crosssections the average treatment eects on treated pairs over time also, accounting (in cross-sections) for the endogeneity of EIAs using treated and control groups. See that paper for such an alternative technique, beyond the scope of this already lengthy paper. 30The Baier-Bergstrand data set is the most comprehensive data set of EIA formations from 1950 to 2010 for 195 countries in the world. 52

53 number increases over time due to the rapid increase in the number of EIAs, the number of negative partial eects is small. This is consistent with the ndings in Figure 2 earlier. 8 Conclusion The following summarizes the contributions of this paper: (1) Using a random coecients econometric model, we demonstrated extensive heterogeneity in EIAs' partial eects on trade ows using gravity equations controlling for endogeneity using the panel techniques of BB and BBF. (2) Using an Armington theoretical trade model, we demonstrated why the heterogeneity in partial eects of EIAs in gravity equations may be related to factors associated with variable trade costs such as freight charges, inuencing the intensive product margin of trade. (3) Using a Melitz-type theoretical model, we demonstrated why the heterogeneity in partial eects of EIAs may be related to factors associated with variable and xed export costs, inuencing the extensive product margin of trade. (4) Using the approach in BBF, we demonstrated that six standard covariates explaining trade ows (used in HMR) have economically and statistically signicant impacts on the eects of EIAs on trade ows, with distance and adjacency explaining the intensive margin heterogeneous EIA eects and distance, adjacency, language, religion, legal origins, and colonial history explaining the extensive margin heterogeneous EIA eects. Moreover, the qualitatively dierent impacts of institutional variables versus cultural variables on the extensive margin eect is consistent with the novel aspect of our Melitz model distinguishing policy from non-policy exogenous export xed costs, as well as distinguishing exogenous from endogenous export xed costs. (5) In the spirit of ACRC, we showed that 85 percent or more of the variation in welfare eects for an importer from EIAs could be explained by two more readily observable statistics: the share of the trade ow from i to j relative to importer j's expenditures (λ ) and the partial eect ( β ). 53

54 (6) Moreover, in the spirit of Baier and Bergstrand (2004), we showed that 85 percent or more of the variation in the probability of an EIA between i and j could be explained by the above mentioned two factors, along with bilateral distance. (7) We showed that the average treatment eects on the treated pairs were highest for the EIAs that formed the earliest, and then declined nearly monotonically, consistent with Bergstrand, Egger, and Larch (2014) and the notion that the earliest EIAs were formed by country-pairs with likely the most to gain in welfare from their EIAs. [TO BE COMPLETED.] 54

55 References Cameron, A., and P. Trivedi (2005): Microeconometrics - Methods and Applications. Cambridge University Press, Cambridge, United Kingdom. [TO BE COMPLETED.] 55

56 Table 1: Data Description 1 Integration Index Count Percent of Total Percent of subtotal 0 (None) 567, (1-way PTA) 94, (2-way PTA) 23, (FTA) 25, (Customs Union) 7, (Common Market) 5, (Economic Union) 2, Subtotal 726, Missing observations 905, Total 1,631, Total observations are based upon 183 countries (183 x 182 = 33,306) for 49 years ( ). Missing observations include country pairs with zero trade value and/or one country (or both) of a bilateral pair did not ocially exist. See data source at 56

57 Table 2: EIA Coecient Estimates (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Variables Expected Sign EIA3-6 t *** 0.60 *** 0.33 *** 0.34 *** (16.85) (16.91) (7.82) (7.46) EIA3-6 t *** 0.31 *** (6.03) (6.23) EIA1-2 t *** 0.02 (3.19) (0.56) EIA1-2 t (1.42) EIA1-6 t *** 0.13 *** (10.74) (3.99) EIA1-6 t *** (5.25) OW P T A t (0.62) (0.12) OW P T A t (-0.61) T W P T A t *** 0.10 ** (6.53) (2.28) T W P T A t *** (3.82) F T A t *** 0.32 *** (14.22) (6.96) F T A t *** (4.07) CU t *** 0.57 *** (8.55) (4.81) CU t *** (3.01) CM t *** 0.82 *** (15.39) (8.79) CM t *** (3.08) ECU t 1.03 *** 0.59 *** (10.27) (3.77) ECU t *** (3.25) R N 155, , , , , , , ,132 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. 57

58 Table 3 (1) (2) (3) (4) (5) Variables Strong/Weak Expected Sign Expected Sign Expected Sign Hypothesis Trade Intensive Extensive EIA t S EIA t ln DIST S EIA t ADJ S? + EIA t LANG S EIA t RELIG W EIA t LEGAL W 0 EIA t COLONY W 0 Notes: S (W) denotes strong (weak) hypothesis. 58

59 Table 4 (1) (2) (3) (4) Variables Expected Sign Without Lags With Lags EIA t *** 0.24 *** (4.53) (3.52) EIA t (0.80) EIA t ln DIST *** ** (-5.92) (-2.11) EIA t 5 ln DIST *** (0.70) EIA t ADJ? *** * (-3.42) (-1.65) EIA t 5 ADJ? *** (-3.13) EIA t LANG *** 0.30 *** (3.68) (2.71) EIA t 5 LANG (-0.66) EIA t RELIG *** 0.23 ** (3.47) (1.97) EIA t 5 RELIG (-1.23) EIA t LEGAL * (-1.91) (-0.91) EIA t 5 LEGAL (-0.30) EIA t COLONY *** ** (-2.86) (-2.07) EIA t 5 COLONY (-0.07) Fixed Eects: Exporter-Year Yes Yes Importer-Year Yes Yes Country-Pair Yes Yes R N 155, ,713 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. 59

60 Table 5 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive EIA t *** ** *** (6.94) (2.31) (4.08) EIA t ln DIST *** *** *** (-9.88) (-3.83) (-5.18) EIA t ADJ? *** *** (0.58) (4.23) (-3.00) EIA t LANG *** *** (3.82) (0.47) (2.92) EIA t RELIG *** *** (4.51) (1.39) (2.75) EIA t LEGAL *** *** (-2.80) (0.60) (-2.92) EIA t COLONY ** *** (-2.52) (1.02) (-3.04) Fixed Eects: Exporter-Year Yes Yes Yes Importer-Year Yes Yes Yes Country-Pair Yes Yes Yes R N 70,173 70,173 70,173 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. Cuto for nontraded goods is $1,000,000; this aects the sample size. 60

61 Table 6 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive EIA t *** *** (6.03) (6.03) (0.20) EIA t ln DIST *** *** ** (-7.69) (-5.44) (-2.46) EIA t ADJ? *** *** (-1.14) (3.40) (-4.30) EIA t LANG *** *** (3.81) (1.04) (2.76) EIA t RELIG *** *** (5.55) (4.08) (1.56) EIA t LEGAL *** * (-3.02) (-1.14) (-1.88) EIA t COLONY *** *** (-3.19) (1.49) (-4.54) Fixed Eects: Exporter-Year Yes Yes Yes Importer-Year Yes Yes Yes Country-Pair Yes Yes Yes R N 103, , ,147 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. Cuto for nontraded goods is $100,000; this aects the sample size. 61

62 Table 7 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive OW P T A t ** (-2.33) (-1.50) (-0.77) T W P T A t *** *** (3.39) (2.59) (10.78) F T A t *** *** *** (19.42) (7.43) (10.93) CU t *** *** *** (16.75) (5.36) (9.98) CM t *** *** *** (24.17) (16.22) (7.14) ECU t *** *** *** (14.53) (11.98) (2.57) Fixed Eects: Exporter-Year Yes Yes Yes Importer-Year Yes Yes Yes Country-Pair Yes Yes Yes R N 70,173 70,173 70,173 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. Cuto for nontraded goods is $1,000,000; this aects the sample size. 62

63 Table 8 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive OW P T A t *** ** (-3.75) (-2.56) (-1.13) OW P T A t ln DIST *** *** 0.01 (-3.44) (3.84) (0.22) OW P T A t ADJ? (0.80) (-0.73) (1.30) OW P T A t LANG *** *** (-0.18) (-2.71) (2.12) OW P T A t RELIG (-0.65) (0.78) (-1.21) OW P T A t LEGAL *** *** (0.00) (3.17) (-2.65) OW P T A t COLONY (0.52) (11.41) (-0.72) T W P T A t (0.25) (0.43) (-0.14) T W P T A t ln DIST *** *** (-3.55) (-3.57) (-0.07) T W P T A t ADJ? (0.96) (1.25) (-0.22) T W P T A t LANG (0.00) (-1.47) (1.24) T W P T A t RELIG * * (0.59) (-1.72) (1.95) T W P T A t LEGAL (-0.77) (10.56) (-1.14) T W P T A t COLONY (-0.36) (-1.11) (0.62) F T A t *** *** (5.94) (3.30) (2.33) F T A t ln DIST *** * *** (-8.12) (-1.76) (-5.57) F T A t ADJ? 0.12 *** *** ** (16.76) (14.46) (-2.24) F T A t LANG *** ** (3.31) (10.79) (2.21) F T A t RELIG *** *** (3.45) (-1.07) (3.88) F T A t LEGAL ** *** (Continued) 63

64 Table 8 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive (-2.13) (1.55) (-3.13) F T A t COLONY * ** (-1.89) (0.94) (-2.41) CU t *** *** (3.46) (2.99) (0.49) CU t ln DIST ** (-1.36) (1.03) (-2.04) CU t ADJ? ** (0.57) (2.11) (-1.27) CU t LANG *** *** (4.62) (0.45) (3.61) CU t RELIG * (1.72) (0.18) (1.34) CU t LEGAL (0.69) (0.32) (0.32) CU t COLONY *** ** (-3.31) (-0.41) (-2.51) CM t *** *** (5.95) (-1.51) (6.42) CM t ln DIST *** *** 0.29 *** (-2.86) (-7.25) (3.62) CM t ADJ? (-0.16) (1.45) (-1.36) CM t LANG *** ** (0.09) (-2.63) (2.27) CM t RELIG (1.33) (0.03) (1.13) CM t LEGAL (-0.32) (0.20) (-0.45) CM t COLONY ** (-1.12) (1.31) (-2.07) ECU t *** (4.37) (1.16) (2.82) ECU t ln DIST * 0.24 (0.01) (-1.93) (1.63) ECU t ADJ? (-0.45) (1.28) (-1.46) ECU t LANG *** * *** (2.57) (-1.94) (3.85) ECU t RELIG *** *** (Continued) 64

65 Table 8 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive (2.61) (3.14) (-0.36) ECU t LEGAL (-1.08) (-0.41) (-0.59) ECU t COLONY (1.26) (0.77) (0.44) Fixed Eects: Exporter-Year Yes Yes Yes Importer-Year Yes Yes Yes Country-Pair Yes Yes Yes R N 70,173 70,173 70,173 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. Cuto for nontraded goods is $1,000,000; this aects the sample size. 65

66 Table 9 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive OW P T A t (-0.92) (-0.72) (-0.22) T W P T A t *** ** *** (5.27) (2.46) (2.85) F T A t *** *** *** (17.46) (14.91) (3.17) CU t *** *** *** (12.82) (9.15) (3.84) CM t *** *** (21.55) (20.88) (1.58) ECU t *** *** (14.28) (15.22) (-0.19) Fixed Eects: Exporter-Year Yes Yes Yes Importer-Year Yes Yes Yes Country-Pair Yes Yes Yes R N 103, , ,147 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. Cuto for nontraded goods is $100,000; this aects the sample size. 66

67 Table 10 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive OW P T A t *** * (-2.81) (-1.73) (-1.11) OW P T A t ln DIST *** *** 0.01 (-2.81) (-3.30) (0.28) OW P T A t ADJ? ** 0.23 (0.39) (-0.58) (0.92) OW P T A t LANG * (-1.31) (-1.87) (0.47) OW P T A t RELIG ** * (0.11) (2.12) (-1.87) OW P T A t LEGAL *** *** (-1.11) (4.51) (-5.29) OW P T A t COLONY (0.96) (0.32) (0.62) T W P T A t (1.21) (-0.48) (1.63) T W P T A t ln DIST *** *** 0.02 (-3.37) (-3.94) (0.44) T W P T A t ADJ? 0.25 *** * (2.82) (1.06) (1.76) T W P T A t LANG (0.08) (-0.60) (0.64) T W P T A t RELIG * (0.69) (-1.13) (1.75) T W P T A t LEGAL ** ** (-0.36) (2.04) (-2.25) T W P T A t COLONY ** (-2.42) (-1.50) (-0.96) F T A t *** *** (5.78) (6.68) (-0.54) F T A t ln DIST *** *** *** (-5.67) (-2.81) (-2.86) F T A t ADJ? *** *** (0.65) (4.36) (-3.39) F T A t LANG *** (3.08) (1.53) (1.61) F T A t RELIG *** *** (4.77) (1.76) (2.95) F T A t LEGAL ** *** (Continued) 67

68 Table 10 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive (-2.19) (0.78) (-2.86) F T A t COLONY *** *** (-2.93) (1.12) (-3.91) CU t *** *** (2.85) (3.07) (-0.09) CU t ln DIST (-0.76) (0.04) (-0.77) CU t ADJ? * (0.18) (1.91) (-1.60) CU t LANG *** *** (4.38) (1.07) (3.26) CU t RELIG * (1.74) (1.22) (0.56) CU t LEGAL (0.25) (-0.57) (0.78) CU t COLONY *** *** (-3.07) (-0.35) (-2.67) CM t *** *** (4.29) (0.12) (4.09) CM t ln DIST *** *** 0.28 *** (-2.94) (-6.80) (3.52) CM t ADJ? (-0.02) (0.72) (-0.70) CM t LANG * (0.04) (-1.65) (1.58) CM t RELIG ** (2.03) (1.49) (0.58) CM t LEGAL (-1.03) (-1.33) (0.24) CM t COLONY ** *** (-2.01) (1.31) (-3.19) ECU t ** ** (2.06) (2.29) (-0.13) ECU t ln DIST ** * (-1.99) (-1.93) (-0.15) ECU t ADJ? (-0.05) (-0.37) (-0.14) ECU t LANG ** * ** (2.39) (-0.03) (2.35) ECU t RELIG ** *** * (Continued) 68

69 Table 10 (1) (2) (3) (4) (5) (6) (7) Expected Sign Expected Sign Expected Sign Variables Trade Trade Intensive Intensive Extensive Extensive (2.22) (4.41) (-1.94) ECU t LEGAL (-1.37) (-1.23) (-0.18) ECU t COLONY (-0.22) (0.96) (-1.11) Fixed Eects: Exporter-Year Yes Yes Yes Importer-Year Yes Yes Yes Country-Pair Yes Yes Yes R N 103, , ,147 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. Cuto for nontraded goods is $100,000; this aects the sample size. 69

70 Table 11: Determinants of (Logs of) Welfare Changes (1) (2) (3) (4) (5) (6) Expected Coecient Variables Value ln λ β *** (105.02) ln λ *** 1.00 *** 0.93 *** (104.49) (119.66) (68.47) ln β *** 0.58 *** 0.58 *** (18.62) (22.41) (21.22) Constant? 3.10 *** 3.28 *** 3.63 *** 0.45 (48.99) (53.69) (5.68) (0.49) Fixed Eects: Importer Exporter No No No No Yes No Yes Yes R 2 N , , , ,266 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. 70

71 Table 12: Determinants of (Logs of) Probabilities of EIAs (1) (2) (3) (4) (5) (6) (7) Expected Coecient Variables Sign ln λ *** 0.04 *** 0.12 *** 0.14 *** ** (19.76) (4.63) ln β *** 0.52 *** 0.65 *** 0.62 *** 0.52 *** ln DIST (43.77) (20.52) *** (16.66) *** (19.57) *** (16.65) *** ADJ? (-69.52) (-54.01) *** (-65.03) (-24.14) *** LANG? (-12.92) *** (-21.54) 0.13 *** (-7.83) 0.37 *** RELIG? (-4.60) *** (3.48) *** (7.90) *** LEGAL? (-5.30) *** (-8.32) *** (-3.36) *** COLONY? (-9.22) 0.96 *** (-6.50) 0.66 *** (-1.21) 0.48 *** Constant (8.53) (7.41) (5.54) 0.29 *** 9.03 *** 9.96 *** *** 8.97 (3.62) (67.57) (56.85) (26.62) (18.80) Fixed Eects: Importer Exporter No No No No No No Yes No Yes Yes R 2 N , , , , ,360 Notes: *, **, and *** denote p < 0.10, p < 0.05, and p < 0.01, respectively. 71

72 Table 13 (1) (2) (3) (4) Period ATT Stand. Dev. Number Negative

73 Figure 1