The World Bank Economic Review, 37(3), 2023, 366–388 https://doi.org10.1093/wber/lhad005 Article The Economic Impact of Deepening Trade Agreements Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Lionel Fontagné , Nadia Rocha, Michele Ruta, and Gianluca Santoni Abstract This paper explores the economic impacts of preferential trade agreements, conditional on their level of am- bition. It clusters 278 agreements, encompassing 910 provisions over 18 policy areas and estimates the trade elasticity for the different clusters. These elasticities are used in a series of general-equilibrium counterfactual situations for endowment economies, revealing that deepening existing agreements (the intensive margin of re- gional integration) could boost world trade by 3.9 percent and world GDP by 0.9 percent. The expected gains from deepening agreements within or across regions vary depending on the initial depth of agreements and the size of regional markets. JEL classification: F14, F15 Keywords: preferential trade agreements, deep integration, structural gravity, general equilibrium 1. Introduction The content of preferential trade agreements (PTAs) has largely changed over time. Trade agreements in the 1950s focused on few policy areas, mostly regulating border measures such as tariffs and quotas, and included a limited number of regulatory requirements and commitments in these areas. While many modern PTAs deal primarily with border measures, a growing number of the recent trade agreements, like the Comprehensive and Progressive Agreement for Trans-Pacific Partnership (CPTPP) or the African Continental Free Trade Area (AfCFTA), are “deep”—i.e. they are complex legal documents covering a large array of border and behind-the-border policies and including liberalizing commitments, trans- parency, enforcement, and other regulatory requirements. Despite the heterogeneous content of modern trade agreements, the large body of economic literature on PTAs mostly relies on dummy variables to identify their trade effects (Limão 2016). This approach fails to capture the multidimensional nature of the depth of trade agreements and can lead to measurement error bias in assessing their economic effects. Lionel Fontagné (corresponding author) is an advisor to the Bank of France, a chaired professor at the Paris School of Economics, and a scientific advisor to CEPII, Paris, France; his email address is lionel.fontagne@banque-france.fr. Nadia Rocha is lead economist at the World Bank; her email address is nrocha@worldbank.org. Michele Ruta is deputy division chief at the International Monetary Fund; his email address is mruta@imf.org. Gianluca Santoni is economist at CEPII, Paris; his email address is gianluca.santoni@cepii.fr. The views expressed in this paper are the authors’ and do not necessarily reflect those of the institutions they belong to. A previous version circulated under the title “A General Equilibrium Assessment of the Economic Impact of Deep Trade Agreements.” We thank seminar participants at the World Bank Economics of Deep Trade Agreements Seminar Series and two anonymous referees for insightful comments. The data underlying this article are available on Mendeley at https://data.mendeley.com/datasets/d63grb8csw. A supplementary online appendix is available with this article at The World Bank Economic Review website. © 2023 International Bank for Reconstruction and Development / The World Bank. Published by Oxford University Press. The World Bank Economic Review 367 This paper takes a novel look at the economic impact of PTAs taking into account the depth of these agreements. It builds on the recent strand of the structural gravity literature (Anderson and Van Wincoop 2003) and explores the uneven economic impacts deriving from the diverse provisions con- tained in PTAs—the intensive margin of regional integration. Different provisions have different effects on trade depending on how they affect trade costs. They may also affect differently PTA members and non-members. For instance, while discriminatory provisions such as rules on antidumping duties or ex- Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 port taxes ease restrictions to trade between members only, non-discriminatory provisions such as rules on subsidies or competition policy reduce trade costs for both members and non-members by allowing all foreign firms to benefit from undistorted competition in the PTA members’ markets. In other words, the varying content of PTAs may be an important determinant of the uneven trade effects observed in Baier, Yotov, and Zylkin (2019). The purpose of this paper is to provide a new method to assess these differential effects and to offer a quantification using new data on the content of PTAs. The quantification analysis relies on a comprehensive characterization of the provisions included in PTAs based on a new database compiled by the World Bank (Mattoo, Rocha, and Ruta 2020). It uses information on all policy domains (excluding tariffs) covering objectives, substantive commit- ments, regulatory requirements, and enforcement procedures included in legal texts and annexes of the 278 PTAs in force and notified to the WTO up to 2018. Examples of policy areas covered in the database include competition policy, state-owned enterprises (SOEs), subsidies, public procurement, tech- nical barriers to trade (TBTs), sanitary and phytosanitary standards (SPS), labor rights, and environmental rules. Such a very rich set of information has to be collapsed into broad categories of PTAs in order to be tractable in a general-equilibrium framework accounting for the complex impact of PTAs with vari- ous levels of ambition on the world matrix of bilateral trade costs. This paper goes beyond the dummy approach—which does not capture depth—used in the large majority of the literature, and the use of syn- thetic indicators of depth such as the count of provisions in PTAs (Mattoo, Mulabdic, and Ruta 2022). It starts by defining statistically significant groupings. It relies on a clustering approach to identify groups of trade agreements based on the provisions’ content. In doing so, it opts for the iterative “k-means++” algorithm developed by Arthur and Vassilvitskii (2007), which ensures greater accuracy by randomizing starting points at each replication. Given the underlying distribution of provisions in each of the 18 policy areas in the data, the silhouette width criterion (which evaluates cluster fit on within-group cohesion and between-group separation) recommends three clusters. This method defines a natural grouping of agreements grounded on a transparent statistical approach based solely on the content of the treaties. Since PTAs within a group are expected to have a similar dis- tribution of provisions by policy areas (as clustering maximizes within-group cohesion) their impact on trade costs within members and between members and third countries is expected to be different. More- over, as PTAs in different groups are expected to have a statistically different distribution of provisions (as clustering maximizes between group separation), the marginal contribution of each policy area is ex- pected to be different across clusters. In order to gauge the differential contribution of policy areas in the classification, the paper provides evidence on the policy markers of different clusters. In a second step an explicit bilateral trade function is estimated, taking stock of the classification of agreements. Once more, we let the data speak and estimate the mean impact on trade for the PTAs be- longing to each cluster. Note that a large part of the literature estimating the impact of PTAs on trade may be flawed because it does not control for the appropriate benchmark in terms of trade cost, i.e. domestic sales (Yotov 2022). Importantly, the gravity equation is taken seriously and the paper integrates internal trade flows in the estimation of the trade impacts of the different types of PTAs. We compile the largest database for which internal and international trade is available in a panel, based on the last release of UNIDO data. 368 Fontagné et al. A challenge when assessing the trade impact of PTAs is endogeneity. Countries self-select in signing a PTA due to unobservable bilateral linkages (Baier and Bergstrand 2004). By the same token, the networks of firms and their joint involvement in global value chains (GVCs) contribute to shape the geography of PTAs. “Lobbying for globalization” (Blanga-Gubbay, Conconi, and Parenti 2020) is ascertained in the US case: large firms benefit more from regional integration because it reduces the trade frictions between affiliates and improves market access; consequently, they spend more on lobbying for PTAs.1 Hence, the Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 endogeneity concerns that have been raised for PTAs in general are likely to bind also for their depth, which is a challenge to be addressed here. This issue is addressed in different ways. First, the paper relies on a conservative fixed effects strategy. Second, it controls for the indirect intensity of bilateral GVC relationships, which captures only the bilateral income generated throughout production linkages with third countries (i.e. excluding the potentially endogenous component related to direct trade). Third, it reallocates randomly country pairs across groups of PTAs of different content and shows after a large number of replications that the distribution of parameters for each group of PTAs is not statistically different. The last step of the analysis illustrates how to perform counterfactual exercises in a theoretically con- sistent way, taking stock of general-equilibrium effects. First, all the existing PTAs are moved away from their cluster towards the most ambitious one (i.e. a PTA cluster associated to higher trade) and the paper examines the economic effects of such policy reform of preferential trade agreements. Since only the con- tent of the agreements is changed, keeping the network of agreements unchanged but only their content, this exercise captures a variation in the intensive margin of PTAs. It is shown that deepening all existing trade agreements could boost world trade by 3.9 percent and world GDP by 0.9 percent relative to the baseline. Few countries are negatively affected by the deepening of trade cooperation, but overall gains are significantly positive for most of the countries, illustrating the importance of the intensive PTA margin for international trade cooperation. The paper then simulates the economic impacts of deepening all the existing agreements of each region from their current level of ambition to the highest level of ambition, and repeats the exercise sequentially for PTAs signed within or between regions. Countries in the East Asia and Pacific region would mostly benefit from deepening preferential trade agreements within the region, while countries in the Middle East and North Africa region would benefit more from deepening agreements with partners outside the region. All other regions fall somewhere between these two extremes. Given the diversity of policy preferences may be larger across than within regions, the low hanging fruits when it comes to deepening trade agree- ments may be mostly regional. But the gains of these different integration strategies (within or between regions) would ultimately differ across countries. The last step simulates the impact of the extensive margin of regional integration for PTAs of different levels of ambition. The rest of the paper is structured as follows. Section 2 describes the rich data set used in our exercise and characterizes the clusters of PTAs in terms of commonalities in their content. Section 3 presents the methodology used to measure the ex post impact of the different types of PTAs and characterizes these clusters in terms of their impact on trade. Section 4 addresses the issue of endogeneity in the relationship between trade and deep PTAs. Section 5 shows how to exploit the information on the trade impact of PTAs of different types in a general-equilibrium framework for an endowment economy in order to simulate counterfactuals for the world economy. Section 6 concludes. 1 As evidenced in the case of US antidumping duties (Bown et al. 2021), trade frictions propagate throughout global production networks. Similarly, Blanchard, Bown, and Johnson (2016) and Bown, Erbahar, and Zanardi (2021) show respectively that offshoring shapes the optimal trade policy of a country and that the importing countries tend to remove antidumping duties from their main partner countries in GVCs. The World Bank Economic Review 369 2. Clustering PTAs Based on Their Content 2.1. Data on the Provisions Contained in PTAs The analysis in this paper is based on a new database on the detailed content of deep trade integration— i.e. the depth of commitments that countries take in PTAs (Mattoo, Rocha, and Ruta 2020).2 In partic- ular, this database provides information on the content of 18 policy areas most frequently covered in PTAs (fig. 1). The list of policy areas mapped includes border measures such as antidumping duties or Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 export taxes, and behind the border measures such as technical barriers to trade, competition policy and environmental law, among others. The analysis focuses on a sample of 278 trade agreements that were signed between 1978 and 2018 and that are currently in force and notified to the WTO. For each agreement and policy area, the database provides a series of questions covering aspects such as stated objectives and substantive commitments, as well as aspects relating to transparency, procedures, and enforcement. The number of provisions coded (910 in total) varies by policy area, reflecting differences in terms of coverage and complexity across policy areas that are negotiated in the agreements (table 1). The share of provisions included in each policy area across agreements ranges between 7 percent on average for antidumping to more than 30 percent on average for policy areas such as competition policy or services. More than one-half of the agreements considered for this analysis cover 50 percent or more of the provisions included in policy areas such as subsidies, sanitary and phytosanitary measures, competition policy, rules of origin, and services. The share of agreements covering 50 percent of provisions is much lower (below 30 percent) for policy areas such as antidumping, labor-market regulation, intellectual property rights, and visa and asylum. The rich information on the content of PTAs poses the challenge of how to aggregate it to define and quantify the agreements’ overall depth. Different approaches can be contemplated from the simple count, Figure 1. Classifications of PTAs by Objective of Policy Area 300 Number of Agreements 100 200 0 s As ts s En pe an es nm P D ta cy Pu Ex erp e La Mo Pr t Ta s ul al s Se ons O rrie as n r M me ure s t R Ca t nt um s ro es s n m ne s to ure bl po rise on S aw bo ve oc xe e l P Inv vice d i vi titi d C h eg pit at Ba M ig ar nt o me Vi ert me en oli om g c d sidi En Tra ch sa les sto an Rig yl C pin ren St al ary f Or i lL at t r b Te yto Ru u ro on e m fe sa y u C o r f t d d u re r ti- f P an p d ke An arif ic c it n w i T ua at lit e- ct ci ni lle Ph Fa te d In e an ad Tr ry ta ni Sa Source: Deep Integration Handbook (Mattoo, Rocha, and Ruta 2020). Note: Number of agreements including a given type of provision. 2 The methodology and data are available at https://datatopics.worldbank.org/dta/index.html. 370 Fontagné et al. Table 1. Content of Preferential Trade Agreements, 1978–2018 Average Agreements Agreements Agreements Agreements Agreements Policy Number of coverage by with zero with < 25% with 25–50% with 50–75% with > 25% area provisions policy domain provisions coverage coverage coverage coverage Antidumping 36 0.10 2 267 9 0 0 Competition policy 35 0.37 2 85 121 59 11 Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Countervailing duties 14 0.21 1 194 67 16 0 Environment 48 0.16 11 196 51 20 0 Export taxes 46 0.20 68 94 106 10 0 Intellectual Property Rights 120 0.07 170 78 28 1 1 Investment 56 0.18 168 6 49 55 0 Labor market 18 0.16 128 98 9 18 25 Migration 30 0.13 169 31 59 19 0 Movement of capital 81 0.21 132 26 78 42 0 Public procurement 95 0.15 111 98 14 54 1 Rules of origin 38 0.36 35 35 126 82 0 Sanitary and Phyto-Sanitary 53 0.14 47 168 61 2 0 State-Owned Enterprises 52 0.23 57 66 150 4 1 Services 62 0.33 136 3 21 78 40 Subsidies 36 0.33 10 51 176 41 0 Technical Barriers to Trade 34 0.18 51 136 83 8 0 Trade facilitation 52 0.25 30 120 90 38 0 Source: Note: Coverage reports of the share of non-zero provisions within each cluster and policy area. or the coverage ratio of the provisions included in an agreement, where it is assumed that the relative importance of each provision is the same across policy areas and agreements. Alternative methods to assign different weights to different provisions according to their commonality or explanatory power across agreements include principal components analysis or machine-learning algorithms. This analysis uses an agnostic statistical procedure to classify PTAs into an optimal set of groups, where agreements present both the maximum similarity in terms of provisions included within groups, and the maximum difference between groups. 2.2. The Classification Algorithm The 278 agreements covered in the data set include a large number of clauses that have been grouped into 910 types of provisions across 18 policy areas. Not all provisions are present in all the agreements and when they are their content may be of various ambition. This extremely detailed information can be used to partition PTAs into distinct groups (clusters) based on the similarity of the underlying treaties (presence or not of individual provisions as well as their ambition). The purpose of the classification is to group PTAs up to the point where the dissimilarity between clusters and the similarity within clusters is maximal. The typical approach to solving this problem in a multidimensional space is to rely on Euclidean distance.3 Groups of PTAs are clustered based on the proximity of their scores obtained for the different categories of provisions. These scores are a simple metric of the ambition of the PTAs for each area. The optimal 3 An analogy may help: an electric car producer may want to partition its potential clients into different groups in order to better target advertising investments. Information is needed on age, gender, income, location in an urban or rural area, daily commuting time, and time spent surfing on social networks with a smartphone. The algorithm would, e.g., identify the group of high-income technophile consumers, which would differ from consumers who make the automobile an ostentatious object of consumption, and also distant from the group of environmentalist consumers who prioritize the reduction of CO2 emissions. Other analogies could be found in medicine or biology. The World Bank Economic Review 371 number of clusters is then identified, i.e. the number that simultaneously maximizes the distance between groups and the cohesion within them, by a statistical criterion: the silhouette width (Rousseeuw 1987). The silhouette width measures the separation between clusters by evaluating how similar agreements within a cluster are to each other with respect to those in the nearest group. Figure S2.2 in the supplementary online appendix reports the average silhouette width by number of clusters. The silhouette reaches its maximum when the data are partitioned across three clusters. Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 To identify the composition of each cluster, a state-of-the-art statistical classification method is used— the k-means++ clustering algorithm. It is a non-hierarchical iterative clustering method that partitions the data into a number of predetermined clusters (defined on the silhouette width) based on the dissimilarity matrix measuring the Euclidean distance between agreements across the 18 policy areas covered in each of the treaties under analysis. In this approach, groups’ centers are defined randomly and, at each iteration, the group center is chosen based on a probability proportional to the minimal distance to the closest pre- viously defined center, ensuring greater accuracy of the resulting classification.4 Eventually, the 278 PTAs are partitioned across three clusters as follows: cluster #1 (29 agreements), cluster #2 (96 agreements), and cluster #3 (153 agreements).5 These clusters are going to be used in the next section to assess the impact of different PTAs on trade. Formally, our data can be represented as a matrix composed of 278 agreements (columns) and 910 provisions (rows). Each cell informs us of the presence or ambition of a given provision in a given agree- ment. Notice than the rows will eventually be grouped into “areas,” e.g. competition or services policy, in order to reduce the dimensionality of the matrix.6 There are 18 such areas. For the sake of clarity, let us first give an example of how the data is structured in the services policy area. The provision covering the obligations needed for a juridical person “to be considered a service sup- plier of a party to the agreement” allows for 6 different options across the 278 agreements. These options range from the most restrictive as “being incorporated under the domestic law of the party and having substantive business operations in the territory of a member” (coded with value 1), to the most liberal as “being owned or controlled by natural persons of the other party” (coded with value 6). The PTA between Australia and New Zealand (year 1983) by including the most liberal formulation of the provi- sion (i.e. “being owned or controlled by natural persons of the other party”) is an illustration of deeper cooperation in the service domain than, for example, the Andean Community (1988), whose founding treaty mentioned the stricter requirement of being incorporated under the domestic law of the party and having substantive business operations in the territory of a member. In order to obtain the score that will help cluster the agreements, we first need to normalize the data, as within a policy area provisions are coded by increasing level of ambition (1, 2, …, n)—see our previous example of the services policy area—while other provisions are coded with a dichotomic variable (1, 0). For each provision in the matrix composed by 278 agreements (columns) and 910 provisions (rows), the mean of each provision across agreements is calculated. This normalization has the double advantage of (i) harmonizing the measuring scale across provisions (not all provisions are binary); (ii) factoring in the frequency of the provision across agreements (if a provision is not present in a certain agreement it is coded as zero). The normalized scores assigned to each of the 910 provisions are then aggregated using the simple average across all the provisions that are included in each of the 18 policy areas mapped. Lastly, the k-means++ algorithm is applied to a reduced matrix composed by 18 rows (defined by the policy domains) and 278 columns (defined by the agreements). 4 The classification is performed using the kmeapp function in R. Although it is allowed for a fairly large number of iterations, i.e. 5,000, the algorithm converges to a stable classification after few rounds. As a robustness check, an alternative classification method is tested in Section 3.2. 5 The complete list of PTAs within each cluster is available in the Mendeley repository associated with this article. 6 The implementation of the k-means clustering poses some difficulties for data sets with high levels of dimensionality. 372 Fontagné et al. Table 2. Practical Example of Provisions Aggregation (Raw Matrix) Cluster #1 Cluster #2 Cluster #3 PTA-1 PTA-2 PTA-3 PTA-4 PTA-5 Average Policy area Provision (1) (2) (3) (4) (5) (6) Area xx a 3 1 0 0 0 0.8 Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Area xx b 2 1 0 1 0 0.8 Area xx c 1 1 1 1 1 1 Coverage by PTA 1 1 0.33 0.66 0.33 Coverage by cluster 1.00 0.67 0.5 Source: Authors’ construction. Note: PTA stands for preferential trade agreement. Table 3. Practical Example of Provisions Aggregation (Normalized Matrix) Cluster #1 Cluster #2 Cluster #3 PTA-1 PTA-2 PTA-3 PTA-4 PTA-5 Average Policy area Provision (1) (2) (3) (4) (5) (6) Area xx a 3.75 1.25 0.00 0.00 0.00 Area xx b 2.50 1.25 0.00 1.25 0.00 Area xx c 1.00 1.00 1.00 1.00 1.00 Score by PTA 2.42 1.17 0.33 0.75 0.33 Score by cluster 2.42 0.75 0.54 Source: Authors’ construction. Note: PTA stands for preferential trade agreement. Table 2 provides an example to illustrate how the data on the content of preferential trade agreements are normalized and aggregated in order to generate a set of clusters. We consider a hypothetical situation with a total of five PTAs that need to be grouped into three clusters which include only one policy area (area xx), which comprises three coded provisions (a, b, c). These provisions take values 1 to 3 according to their level of ambition and take value 0 if they are not included in the agreement. The first step is to construct a matrix with normalized scores capturing the average occurrence of each provision across PTAs. Consider provision a. This provision is present in PTA-1 and is coded with the highest level of ambition (score=3). Provision a is also present in PTA-2, but with the lowest level of ambition (score=1). Provision a is absent from PTAs 3 to 5 (code=0). The average occurrence of provision a across the PTAs is therefore 0.8 ((3 + 1 + 0 + 0 + 0)/5). This value also captures the frequency of provision a across agreements. A similar exercise can be done with provisions b and c, where the average occurrence is equal to 0.8 and 1, respectively. As a second step, the score provided to each provision is normalized by dividing the current score by the average occurrence (frequency). The normalized scores for provisions a, b, and c included in agreements PTA-1 to PTA-5 are provided in table 3. The normalized score of provision a in agreement PTA-1 is equal to 3.75 (3/0.8) and captures the relative occurrence and intensity (gradient 1, 2, …, n) of this provision in this agreement. Finally, all the provisions’ scores within each policy area are aggregated to reduce the dimensionality of the matrix.7 7 Table S2.1 provides a subset of actual data in the raw and normalized matrix for the area of services. Note that provision #850 is coded 5 in EFTA (high ambition) and only 2 in the Chile–Japan PTA, which illustrates the aforementioned (1, 2, …, n) coding. In cluster #1 the final score is 0.840; it is 0.846 in cluster #3 and 1.362 in cluster #2. The same exercise pertains for the 17 other policy areas and for other PTAs in order to obtain the score for each area and each PTA (an 18*278 matrix) thereafter used for the clustering. The World Bank Economic Review 373 Figure 2. Marginal Probability of Being in Cluster #1 and Cluster #3 Anti-dumping Countervailing Duties Competition Environment Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Export Taxes IPR Investment Labor Market Movement of Workers Movement of Capitals Public Procurement Rules of Origin SPS STE Services Subsidies TBT Trade Facitlitation -0.200 -0.100 0.000 0.100 0.200 Cluster # 3 Cluster # 1 Source: Authors’ calculation. Note: Marginal effect of the 18 different provision areas from a linear probability model for being in the corresponding cluster. The model controls for the decade of signature of the agreements. Cluster #2 is the reference group. 2.3. The Marginal Effect of Provisions on the Probability of an Agreement In order to provide an illustration of the policy content of PTAs in each cluster, a simple linear probability model will ask what the marginal effect of the 18 policy areas on the probability of an agreement being in clusters #1 or #3 with respect to cluster #2 is. Specifically, we run two separate regressions for the probability of an agreement being classified in cluster #1 or #3. The total number of observations in each regression is accordingly 278. On the right-hand side, we include the 18 scores by policy domain used in the clustering exercise. As additional controls we also include a dummy for the decade in which the agreement has been ratified and a series of dummies for the income level of participants: high–high, low– low, high–low income. In order to fix the reference group, a dummy variable equal to 1 for the agreements classified in cluster #2 is included in both regressions. Taking all these elements on board, fig. 2 plots the 18 provision areas on the vertical axis, and on the horizontal axis the marginal effect of each on the probability of a PTA belonging to cluster #1 or cluster #3. It illustrates that antidumping and competition provisions play an important role in cluster #1 relative to provisions in areas such as labor regulation. Following that line of reasoning, deepening a PTA classified in cluster #3 into a PTA classified in cluster #1—the type of counterfactual considered below— would on average require the introduction of ambitious provisions on antidumping or competition. Public procurement and movement of workers are also the types of provisions discriminating between those two types of PTAs. 374 Fontagné et al. 3. Ex Post Quantification of the Impact of Different Types of PTAs In this section, the trade impact of the three clusters of PTAs (identified with the k-means++ algorithm) is estimated in a gravity framework in order to recover the partial effect on trade of PTAs associated with the different clusters. Robustness of our results to alternative classifications is also tested. 3.1. The Structural Gravity Estimation Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Specifically, the following structural gravity model is estimated using Poisson pseudo maximum likelihood (PPML) with panel data:8 3 Xi j,t = exp βz PTAz i j,t + β Transitory_PTAsi j,t z=1 2000 + βT INTL_BRDRi j ∗ T + πi,t + χ j,t + μi j + i j,t , (1) T =1978 where Xij,t includes both intra-national and international yearly manufacturing trade flows from 1978 to 2018.9 Including intra-national sales is critical as the domestic economy is the most appropriate bench- mark for trade integration (Yotov 2012). Otherwise, the estimated coefficients would suffer from a missing variable bias. The data on inter- and intra-national trade come from, respectively, the UN-ComTrade and UNIDO-Indstat databases.10 Following standard practice, intra-national flows are filled in using linear interpolation between non-missing data and extrapolating remaining missing values using gross output to value added ratios as in Head and Mayer (2021). In each year, only countries with non-missing intra- national trade flows enter the estimation sample. =1 z=2 z=3 Our main variable of interest, PTAij,t , is split across cluster groups: PTAzi j,t , PTAi j,t , and PTAi j,t . The dummy INTL_BRDRi j takes the value 1 in the case of an international trade flow. This dummy is in- teracted with decades indexed by T (leaving the period after 2010 as reference). Exporter-time, π i,t , and importer-time, χ j,t , fixed effects control for time-varying multilateral resistance terms, while directional bilateral fixed effects, μij , control for time-invariant unobserved characteristics of the country pair po- tentially leading to self-selection into PTAs (Baier and Bergstrand 2007).11 As an additional control, the variable Transitory_PTAsi j,t is also included to control for agreements that are no longer in force. The results presented in table 4 provide the elasticity of bilateral trade to the different types of PTAS (as grouped by the clustering). This elasticity is here estimated within sample alternatively for the raw data and for the sample including extrapolated data. The set of elasticities obtained for the extrapolated data will be introduced in a second step as a parameter in the procedure. In doing so, we are able to exploit a squared data set of 112 countries in the general-equilibrium counterfactual evaluation for the year 2018. Restricting the sample to non-extrapolated domestic sales would constrain the analysis to 8 The variable Xij,t is in levels in column (1) to (4) of table 4 and in shares of absorption at destination—namely Xij,t / i Xij,t —in the remaining estimation tables. See discussion below. 9 An issue is that the adjustment of exporters to the inception of a PTA is not instantaneous. While Anderson and Yotov (2016) use four-year intervals, with our data it would be difficult to use such a strategy as each individual agreement would hardly be implemented at the first date of a given interval. Averaging across agreements with different com- mencement dates would not adequately capture the impact of the depth of agreements on trade volumes. Accordingly a specification relying on sequential data is adopted, as suggested by Egger, Larch, and Yotov (2022). 10 ComTrade is used to retrieve information on international trade flows from 1978 up to 2018, whereas Indstat provides data on gross output used to compute domestic sales (as the difference between production ant total exports). Due to the limited information on international trade, transition economies enter the estimation sample in 1992, after the collapse of the Soviet Union. 11 Our results are robust to symmetric country-pair fixed effects a shown in table S2.2. The World Bank Economic Review 375 Table 4. PPML: Gravity Estimations of the Elasticity of Trade to PTAs by Cluster Xijt Xijt /Xjt Dep var: (1) (2) (3) (4) (5) (6) (7) (8) PTAij,t 0.569 0.588 0.273 — — — — — (0.047) (0.054) (0.040) Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 =1 PTAz i j,t — — — 0.606 0.525 0.484 0.542 0.492 (0.046) (0.041) (0.073) (0.052) (0.091) =2 PTAz i j,t — — — 0.227 0.205 0.260 0.241 0.308 (0.045) (0.030) (0.035) (0.037) (0.046) =3 PTAz i j,t — — — 0.112 0.096 0.135 0.109 0.154 (0.054) (0.041) (0.040) (0.048) (0.049) Transitory_PTAsi j,t — 0.082 0.104 0.203 0.206 0.130 0.233 0.145 (0.050) (0.046) (0.054) (0.037) (0.043) (0.046) (0.053) INTL_BRDR ∗ 1980 — — −0.855 −0.843 −1.077 −0.989 −1.063 −0.975 (0.044) (0.043) (0.048) (0.062) (0.076) (0.094) INTL_BRDR ∗ 1990 — — −0.528 −0.529 −0.792 −0.774 −0.787 −0.762 (0.034) (0.034) (0.038) (0.045) (0.056) (0.064) INTL_BRDR ∗ 2000 — — −0.168 −0.169 −0.261 −0.315 −0.256 −0.302 (0.027) (0.027) (0.028) (0.033) (0.043) (0.047) Intra-national flows Raw Raw Raw Raw Raw Extrapolated Raw Extrapolated Period 1978–2018 1978–2018 1978–2018 1978–2018 1978–2018 1978–2018 1978–2018 1978–2018 Frequency Yearly Yearly Yearly Yearly Yearly Yearly Yearly Yearly N. country ID 136 136 136 136 136 143 136 143 Observations 338,685 338,685 338,685 338,685 338,685 591,320 338,685 591,320 FEs it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij Three-way Three-way correction correction Source: Authors’ calculation. Note: Exporter-time (it), importer-time (jt), and exporter-importer (ij) fixed effects are always included. From columns (1)–(5), standard errors in parentheses are clustered by country pair. In columns (7) and (8), both standard errors and point estimates are corrected using the Weidner and Zylkin (2021) procedure, implemented in Stata with ppml_fe_bias. In columns (5) and (7), missing values in domestic sales are linearly interpolated and the remaining missing values are extrapolated using the evolution of a country total exports. PTA stands for preferential trade agreements. PPML: Poisson pseudo maximum likelihood. INTL_BDR: dummy taking the value 1 for international trade flows. only 61 countries, unevenly distributed across regions. The broader data set is favored as it improves the coverage and reliability of the regional general-equilibrium results. Columns (1)–(5) and (7) exploit the raw UNIDO data used. There are 338,685 observations for 136 countries for which we observe internal flows at least once over the estimation period (61 countries in the year 2018). In order to expand internal flows coverage, in column (6) we rely on a broad sample obtained by extrapolating domestic sales, which leads to 591,320 “observations” for 143 countries (112 in the year 2018).12 Column (1) is the standard estimation strategy with importer-time, exporter-time, and dyadic fixed effects, and controlling for internal flows. Column (2) replicates column (1) by adding the control for transitory PTAs, which does not significantly affect the PTAs coefficient. Column (3) adds the control for internal flows (with a dummy taking the value 1 in the case of international flows interacted with a decade specific indicator variable).13 Comparing with columns (1) and (2) shows that this dramatically reduces the trade impact of PTAs. The second result is that the negative impact on commerce of crossing the border decreases progressively, which is the other side of the “globalization” coin. 12 The list of 61 and 51 additional countries is given in the table S2.5. 13 The years 2001 to 2018 represent the excluded period. 376 Fontagné et al. The impact of PTAs shown in column (3) is then split by clusters of PTAs. The estimated elasticity ranges from 0.112 to 0.606, with cluster #1 having the largest impact on trade. This cluster is referred to as the one of (revealed) “deep” PTAs, the second as “medium,” and the third as “shallow” PTAs. The comparison of column (3) and column (4) shows a key result of our paper: using a well-specified gravity equation, the usual approach relying on a single dummy for the presence of PTAs misses the differentiated impact of PTAs having different ambition. The average point estimate for the usual dummy approach is Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 0.273, while we get 0.112, 0.227, and 0.606 respectively for clusters #3, #2, and #1.14 The impact of such a difference in quantifying the impact of regional integration is indeed of primary importance: the use of a single dummy variable would overestimate the impact on trade and GDP of the new low-ambition agreements, while underestimating this impact for the high-ambition agreements.15 One issue to be addressed is that if the bilateral factors are incompletely captured or proxied, they will be loaded into the importer and exporter fixed effects used later in the counterfactuals. This may be the case here since the bilateral factors are time invariant (so as not to capture variation in signed agreements). In this case, the model would fit large trade flows quite well but small flows not so well (Egger and Nigai 2015). Correcting for this potential bias caused by unobservable trade costs that operate through general-equilibrium constraints allows the impact of EPAs to be correctly identified (Kharel 2019).16 As our database also includes domestic sales, starting from column (5) results control for the relative size of the trading partners using an estimator on shares. The PPML estimator, in fact, assigns more weight to countries with large import volumes in the identification of the parameters (Eaton, Kortum, and Sotelo 2013; Head and Mayer 2014). A way to eliminate differences in the penalization of large and small trade flows is to normalize trade flows by destination country total absorption (Eaton, Kortum, and Sotelo 2013); this is what is done for specifications in columns (5) and (7) of table 4 with the raw data, and in columns (6) and (8) using the extrapolated data.17 While columns (1)–(5) and (7) rely on raw data, columns (6) and (8) report the estimation results on extrapolated data. Both the point estimates of each cluster and the difference between them are very similar across the two estimation samples. In order to ensure a broader and more representative set of countries in the general-equilibrium simulations, the sample and estimated elasticities in column (6) of table 4 are chosen. Finally, an important econometric issue is that our three-way fixed effect panel PPML procedure with time-invariant country-pair, time-varying exporter and importer fixed effects may be subject to an 14 Assuming a value for the elasticity of substitution of σ = 5 (see e.g. estimations in Fontagné, Guimbard, and Orefice (2022)) the baseline PTAs estimates can be expressed as tariff-equivalent effects. Deep PTAs are equivalent to a tariff reduction of 11.4 percent (i.e. [exp (0.606/(−5)) − 1]*100 = 11.4); medium PTAs are equivalent to a reduction of 4.4 percent and shallow ones are equivalent to a reduction of 2.2 percent. 15 This is illustrated in table S2.3 by comparing columns (3) and (4) with columns (7) and (8) respectively. In this table showing the results of the extensive margin of regional integration for East Asia and Pacific–region countries, the impact of the PTAs of medium ambition is very close to the impacts that would be obtained with a simple dummy approach, meaning that the usual approach is not capturing the richness of the agreements signed (e.g. 2.29 percent additional total exports instead of 2.30 percent) for Australia. In contrast, Australia’s signing of new superficial (columns (5) and (6)) or ambitious (resp. (1) and (2)) PTAs would make a difference, as the impact on exports would be 1.31 percent versus 4.56 percent, respectively, which would not be captured by the usual single-dummy approach. 16 We are indebted to an anonymous referee for stressing this issue. 17 As shown by Sotelo (2019), Poisson estimation on the market-share variable with country fixed effects is equivalent to the multinomial PML proposed in Eaton, Kortum and Sotelo (2013). It is important to note that the interpretation of coefficients estimated with PPML in shares is the same as that in levels: the only difference lies in the way the observations are weighted. In calculating counterfactuals, we use a PPML in levels to which we apply PTA elasticities and trade frictions both estimated with market-share weights. The use of the PPML in levels in the counterfactual exercise is imposed by the need to comply with the model’s constraints, according to which the sum of trade inflows must equal expenditure. The World Bank Economic Review 377 incidental parameter problem when the number of periods is small (such as ours), which biases the estimated parameter for PTAs in gravity equations, as well as their confidence intervals.18 Accordingly, column (7) replicates column (5), using raw data, by relying on the fix developed by Weidner and Zylkin (2021), which confirms that point estimates and standard errors are both larger when the incidental parameter problem is properly addressed. The same is done in column (8), by replicating the estimates in column (6), using extrapolated data. Importantly, the statistical significance of our estimated parameters Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 is confirmed. Estimates in column (6) yield a robust set of elasticities for a broad set of countries to simulate the welfare effect of trade policy shocks. Indeed, the identified trade elasticities to PTAs in column (6) will be the parameters introduced in the general-equilibrium gravity model, jointly with the country-pair fixed effects and the international border effects. Two types of counterfactuals for regional integration can be contemplated: one set of counterfactuals on the intensive margin and one on the extensive margin. On the intensive margin—which is the key issue addressed in this paper—we simulate the effect of deepening existing agreements. This can be done for all PTAs in the world, or for all PTAs involving countries in a given region (with other countries within the same region, or alternatively with countries in different regions). Practically, in the counterfactual policy scenario for these simulations, positive entries in either PTAz=3 or PTAz=2 are switched to zero while the corresponding entries in PTAz=1 are set to 1.19 On the extensive margin, we simulate the effect of “signing” missing agreements with all trade part- ners for the East-Asia-Pacific (EAP) region. This is done assuming that new agreements are, alternatively, shallow (all zero entries are switched to PTAz=3 , other entries are unchanged), medium (respective to PTAz=2 ), or deep (to PTAz=1 ). We also provide, for the sake of comparison with the usual single-dummy approach, a counterfactual where countries in this region sign PTAs with all their trading partners, with no consideration of the depth of the agreements signed. Evidence from new (Canada and the EU: CETA) or renegotiated (renegotiation of the NAFTA: USMCA) PTAs suggests that the complexity of PTAs has increased through time. As the clustering al- gorithm centers on the provision content and not on the timing of the agreement, we test whether the trade effect of deep PTAs (i.e. cluster #1) has changed over time. The estimation results by decade are reported in fig. S2.1. Until the end of the 2000s, the estimated elasticity of deep PTAs appears to be fairly stable, but it increased during the 2010s. However, confidence intervals are larger in the last decade and tend to overlap with the overall effect. For this reason, the average elasticity estimated over the whole period is used in our quantitative exercise. 3.2. Alternative Classification Algorithm As our exercise relies on the k-means++ algorithm, it is important to check whether our results are robust to alternative classification algorithms. Changing algorithm may alter the results in two dimensions: the composition of PTAs in each cluster and the estimated elasticity for each cluster in the structural gravity. Both dimensions have implications for the counterfactuals performed afterwards. In order to consistently test the different approaches, we first compute the clusters using the alternative methods and then reesti- mate a new set of trade elasticity for each cluster. In this application, three different clustering routines are 18 In fact, the usual clustering procedures provide biased (too narrow) standard errors in such a setting. 19 Multilateral resistance terms and factory-gate prices are indexes subject to a normalization. This imposes choosing a reference country not directly impacted by the counterfactual and restricting the exercise to cross-sectional comparisons across countries. See Yotov et al. (2017, chapter 2) for a detailed discussion. As we have chosen South Africa as numeraire, the change in exports of the Sub-Saharan Africa region is underestimated because we do not change the vector of PTAs for South Africa. 378 Fontagné et al. Table 5. PPML: Gravity Estimations of the Elasticity of Trade to PTAs by Alternative Cluster Definitions Xijt /Xjt Dep var: (1) (2) (3) (4) (5) (6) =1 PTAz i j,t 0.525 0.542 0.500 0.525 0.482 0.501 (0.041) (0.053) (0.037) (0.048) (0.038) (0.048) Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 =2 PTAz i j,t 0.186 0.219 0.189 0.220 0.185 0.215 (0.031) (0.039) (0.029) (0.036) (0.029) (0.036) =3 PTAz i j,t 0.112 0.127 0.088 0.095 0.105 0.117 (0.038) (0.046) (0.042) (0.049) (0.040) (0.047) Transitory_PTAsi j,t 0.198 0.224 0.207 0.237 0.200 0.228 (0.037) (0.046) (0.039) (0.048) (0.038) (0.047) INTL_BRDR ∗ 1980 −1.077 −1.063 −1.076 −1.061 −1.078 −1.064 (0.048) (0.076) (0.048) (0.076) (0.048) (0.076) INTL_BRDR ∗ 1990 −0.793 −0.789 −0.790 −0.785 −0.794 −0.790 (0.038) (0.056) (0.037) (0.055) (0.037) (0.055) INTL_BRDR ∗ 2000 −0.262 −0.257 −0.265 −0.261 −0.266 −0.262 (0.028) (0.043) (0.028) (0.043) (0.028) (0.043) Cluster definition PAM Hierarchical Reclassify “borderline” PTAs Period 1978–2018 1978–2018 1978–2018 1978–2018 1978–2018 1978–2018 N. Country ID 136 136 136 136 136 136 Data Raw Raw Raw Raw Raw Raw Observations 338,685 338,685 338,685 338,685 338,685 338,685 FEs it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij Three-way Three-way Three-way correction correction correction Source: Authors’ calculation. Note: Exporter-time (it), importer-time (jt), and exporter-importer (ij) fixed effects are always included. Standard errors in parentheses are clustered by country pair. In columns (2), (4), and (6), both standard errors and point estimates are corrected using the Weidner and Zylkin (2021) procedure, implemented in Stata with ppml_fe_bias. The genetic algorithm in columns (1) and (2) classifies 24 PTAs in cluster #1, 113 PTAs in cluster #2, and the remaining 141 PTAs in cluster #3. The hierarchical algorithm in column (3) and (4) classifies 33 PTAs in cluster #1, 103 PTAs in cluster #2, and the remaining 142 PTAs in cluster #3. The manual reclassification in columns (5) and (6) implies 34 PTAs in cluster #1, 94 PTAs in cluster #2, and the remaining 150 PTAs in cluster #3. In columns (2), (4), and (6), both standard errors and point estimates are corrected using the Weidner and Zylkin (2021) procedure, implemented in Stata with ppml_fe_bias. PTA stands for preferential trade agreements. PPML: Poisson pseudo maximum likelihood. INTL_BDR: dummy taking the value 1 for international trade flows. considered:20 an alternative iterative algorithm (the partitioning around medoids, PAM), a non-iterative hierarchical clustering procedure, and last, a manual reclassification of “borderline PTAs.”21 The main differences are as follows. With the genetic algorithm, group 1 is restricted to 24 PTAs, compared to 29 with our preferred algorithm, 33 with the hierarchical one, and 34 with the reclassification of “borderline PTAs” in columns (5) and (6) (see the note for table 5). All in all, the k-means++ algorithm sounds a good compromise that properly selects the 29 most ambitious PTAs and populates the least ambitious group (153 PTAs).22 Estimations are robust to these alternative classification algorithms, as shown in table 5 showing the results of the PPML in share with raw data. These results, to be compared with columns (5) and (7) in table 4, are not statistically different for deep PTAs in cluster #1. This is of first-order importance since our main counterfactuals exploit the intensive margin of regional integration by shifting existing to the “deep” cluster: as agreements with the higher trade effects (i.e. cluster #1) largely overlap with those 20 The list of PTAs in each of the three clusters is provided in the Mendeley repository associated with this article. 21 The reclassification is based on a visual inspection of the cluster space obtained with the k-means++ algorithm, reported in fig. 3. 22 The five additional PTAs in cluster #1, compared to the genetic algorithm, are EFTA-Serbia, EFTA-Ukraine, EFTA- Montenegro, EFTA-Bosnia and Herzegovina, Peru–Mexico. The World Bank Economic Review 379 Figure 3. Cluster Space 4 2 Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 PC1 (11.5 % ) -2 -4 -6 0 -6 -4 -2 0 2 4 PC2 (65.2 %) Cluster 1 Cluster 2 Cluster 3 Source: Authors’ calculation. Note: Spatial representation of the three clusters. Each point represents a trade agreement. The x-axis and y-axis are defined using the first two principal components of the 18 features used by the clustering algorithm, centered around zero. identified by the preferred routine, hence the magnitude of the trade premium associated with the most ambitious group is hardly affected. We lastly recalculate the estimated effects after reallocating the few “borderline” agreements to the closest alternative clusters in columns (5) and (6) of table 5. “Borderline” agreements are manually re- classified as visualized in fig. 3 reporting the position of the 278 agreements over the cluster space. The coordinates represent the first two principal components extracted from the 18 features used in the clus- tering algorithm. The color and shape of each point represent the different clusters. While the separation between clusters is clear-cut with very few agreements at the “border” of their partition, and agreements in cluster #1 (the “deep” ones) standing apart from the rest of the sample, there are a couple of instances where the separation is less clear-cut. These are the reclassified PTAs. Overall, results prove again to be robust to this reclassification. 4. Addressing the Endogeneity of the Content of PTAs As discussed in the Introduction, Blanga-Gubbay, Conconi, and Parenti (2020) find evidence that big firms in the United States tend to lobby for specific provisions to be included in PTAs. This implies controlling also for the intensity of bilateral (time-varying) economic interests that may encourage lobbing activity. Baier and Bergstrand (2007) show that dyadic fixed effects are required to control for self-selection of country pairs into PTAs. These fixed effects have already been introduced in the estimated equations; this section goes beyond this approach to investigate endogeneity concerns. As cross-country data on political economy incentives to lobbying—generally used as instrumental variables—are not available, the section proceeds in three steps. 380 Fontagné et al. Figure 4. Distribution of Estimated Coefficients for Randomly Defined Clusters (1,000 Replications) +σ Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 β RTA σ- Cluster #1 Cluster #2 Cluster #3 1000 replications. β: 0.273 σ: 0.040 Source: Authors’ calculation. Note: The boxes plot the distribution of the estimated partial trade effects of randomly defined clusters. In each of the 1,000 replications, country pairs with a PTA were randomly assigned to a given cluster. First, country pairs are randomly reallocated into groups of PTAs, keeping the number of groups and the number of pairs in each group constant. Second, a control for the bilateral intensity of GVC income by country pair is introduced. This measure traces the income of domestic factors (capital and labor) generated through foreign production chains, thus capturing the intensity of bilateral economic interests (Johnson 2018). Importantly, there is also a potential issue of endogeneity in such a metric, as bilateral trade flows are partly determined by GVCs. This is why we rely on the extraction method to filter the contribution of direct trade to ensure only “indirect” GVC income flows are used.23 Third, we control for the “dependency” of the importing country vis-a-vis its partner country by nor- malizing trade flows by destination country total absorption. 4.1. Randomization of the Grouping of Country Pairs Figure 4 reports the distribution of the estimated coefficients when the country pairs sharing a trade agreement in the estimation sample are randomly assigned to a given cluster. After 1,000 replications, the estimated coefficient for each cluster is not statistically different from the main PTAs coefficient reported in column (3) of table 4. If the different impact of PTAs grouped in different clusters is not an artefact, then randomly allocating PTAs to three clusters (and keeping the number of PTAs in each cluster unchanged) should end up in all clusters having the same impact on trade. 4.2. Controlling for Indirect GVC Participation and Alternative Timing In order to control for the intensity of country-pair production linkages, which may encourage lobbying for selected provisions, we include as a control variable bilateral (indirect) GVC income flows. As the 23 See Borin and Mancini (2023) for a presentation of the extraction method as well as other insightful decompositions. The World Bank Economic Review 381 Table 6. Elasticity of Trade to PTAs by Cluster Controlling for GVCIndirect Income Flows Xijt /Xjt Dep var: (1) (2) (3) (4) (5) (6) =1 PTAz i j,t 0.447 0.447 0.446 0.446 0.440 0.454 (0.039) (0.039) (0.039) (0.039) (0.039) (0.048) Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 =2 PTAz i j,t 0.139 0.139 0.138 0.139 0.135 0.152 (0.028) (0.028) (0.028) (0.028) (0.029) (0.033) =3 PTAz i j,t 0.050 0.050 0.050 0.050 0.050 0.073 (0.048) (0.048) (0.048) (0.048) (0.048) (0.058) 99pc GVCi j,t −3 — −0.002 — — — — (0.035) 90pc GVCi j,t −3 — — 0.032 — — — (0.027) 75pc GVCi j,t −3 — — — 0.042 — — (0.031) IHS(GVCij,t−3 ) — — — — 0.026 0.034 (0.017) (0.024) Period 1993–2018 1993–2018 1993–2018 1993–2018 1993–2018 1993–2018 N. country ID 120 120 120 120 120 120 Observations 230,059 230,059 230,059 230,059 230,059 230,059 FEs it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij Three-way correction Source: Authors’ calculation. Note: Exporter-time (it), importer-time (jt), and exporter-importer (ij) FEs, International Border and Transitory PTAs always included. Standard errors in parentheses clustered by country pair. The variable GVCij,t−3 measures the income generated through indirect supply chain linkages between countries i and j, excluding any direct effect of trade in intermediate goods between the country pair by the extraction method. In columns (2)–(4), indirect GVCij,t−3 income is included as a dummy variable if bilateral flows are above the 95th, 90th, or 75th percentile respectively. In columns (5) and (6), IHS(GVCij,t−3 ) refers to the inverse hyperbolic sine function (IHS). In column (6), both standard errors and point estimates are corrected using the Weidner and Zylkin (2021) procedure, implemented in Stata with ppml_fe_bias. PTA stands for preferential trade agreements. GVC: global value chains. INTL_BDR: dummy taking the value 1 for international trade flows. measure is built with an extraction method, it excludes income generated through bilateral trade flows between countries i and j and captures only the bilateral income generated throughout production linkages with third countries. The EORA MRIO database helps trace income flows through production chains; as the data cover the period 1990–2015, we restrict the estimation sample accordingly to 1993–2018, leaving a three-year lag between GVC participation and trade variables.24 Results of this new specification are shown in table 6. Column (1) replicates the baseline specification (i.e. column (6) of table 4), but on a shorter period, due to the availability of the multiregion input–output database needed to compute bilateral GVC income flows. In columns (2) to (4) we include, as a control for the intensity of bilateral production linkages, a dummy variable taking the value 1 if GVC income flows are above the 99th, 90th, and 75th percentile respectively. Finally, column (5) introduces in the estimated equation a hyperbolic transformation of the indirect GVC income variable instead of a dummy, while column (6) replicates the last estimation using the correction for the incidental parameter problem from Weidner and Zylkin (2021). In table 7, the robustness of our main finding is tested across periods: preferential trade agreements in cluster #1 are more effective in promoting trade between members. To proceed we first report in col- umn (1) the same specification as in column (3) of table 6, and then replicate the same specification over 24 The EORA database provides sectoral input–output (IO) tables for 189 countries (and a Rest of the World aggregate): https://worldmrio.com/eora26/. 382 Fontagné et al. Table 7. Elasticity of Trade to PTAs by Cluster and Period, Controlling for GVCIndirect Income Flows Xijt /Xjt Dep var: (1) (2) (3) (4) (5) =1 PTAz i j,t 0.446 0.388 0.341 0.396 0.387 (0.039) (0.043) (0.046) (0.100) (0.117) Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 =2 PTAz i j,t 0.138 0.142 0.135 0.122 0.137 (0.028) (0.030) (0.029) (0.033) (0.037) =3 PTAz i j,t 0.050 −0.047 0.046 0.082 0.075 (0.048) (0.050) (0.039) (0.064) (0.070) Period 1993–2018 1998–2018 2003–2018 2008–2018 2008–2018 N. country ID 120 111 110 99 99 Observations 230,059 177,745 127,580 77,849 77,849 FEs it, jt, ij it, jt, ij it, jt, ij it, jt, ij it, jt, ij Three-way correction Source: Authors’ calculation. Note: Exporter-time (it), importer-time (jt), and exporter-importer (ij) FEs, International Border and Transitory PTAs always included. Standard errors in parentheses 90pc 90 pc clustered by country pair. The variable GVCi j,t −3 is included in all the regressions. The variable GVCi j,t −3 is a dummy variable taking the value 1 if bilateral GVC income flows is higher than the 90th percentile. GVC measures the income generated through indirect supply chain linkages between countries i and j, excluding any direct effect of trade in intermediate goods between the country pair by the extraction method. PTA stands for preferential trade agreements. GVC: global value chains. shorter and more recent time periods: 1998–2018 (column 2), 2003–2018 (column 3), and 2008–2018 (column 4). As in the previous tables, column (5) of table 7 replicates the last estimation using the correc- tion for the incidental parameter problem. Reassuringly, neither the inclusion of the indirect GVC intensity nor the change in time horizons affects the relative magnitude of the estimated coefficients for the different clusters. 5. General-Equilibrium Gravity and Counterfactuals Relying on the analysis of the impact of different clusters of signed PTAs, this section assesses the economic consequences of deepening existing trade agreements. 5.1. Background A large body of literature focusing on PTAs and trade relies on the assessment of the partial impact of agreements on trade within the countries that are signatories of such agreements (see Limão (2016) for a survey). However, PTAs affect the global matrix of relative trade costs between country pairs and the prices faced by exporters and importers in any country through general-equilibrium effects. Thus, quantifying the trade effects of PTAs of uneven ambition can be done in calibrated computable general-equilibrium (CGE) models or in estimated structural gravity models. Notwithstanding the drawback of relying on elasticities estimated outside the model, CGE models offer a flexible tool to assess the sectoral impact of detailed tariff shocks. However, when the information on trade-cost reduction is not sector specific (as it would be with tariffs or with the sector-level estimation of the trade impeding impact of regulations), but origin-destination specific, with no sectoral dimension, as in the case of the used database on PTAs, the advantage of the large sectoral decomposition of these models vanishes, making the structural gravity approach more appealing. Against this background, a recent strand of literature is using estimated models (or a combination of estimation and calibration) inspired by the structural gravity literature initiated by Anderson and Van Wincoop (2003), to assess the GE effects of shocks to the matrix of trade costs (see Yotov et al. (2017) The World Bank Economic Review 383 for a didactic presentation). A first intrinsic advantage of these models is to have the trade elasticity estimated with the data used for the counterfactual exercise. The second advantage is to be rather agnostic in terms of the trade effects of provisions of PTAs going beyond the phasing out of tariffs among signatory countries.25 A natural extension of such a modeling approach, in line with the spirit of this paper, is to assess the uneven impacts of PTAs (Baier, Yotov, and Zylkin 2019), provided that the “ambition” of signed agreements differs strongly. Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 5.2. Quantification Strategy This section relies on a general-equilibrium gravity model for an endowment economy and quantifies the intensive margin of regional integration, whereby the countries increase the ambition of existing agree- ments instead of signing new PTAs.26 The first consequence of deepening agreements, as evidenced above by our gravity estimation, is to modify the overall structure of trade costs. The bilateral trade adjustment, induced by the change in the policy, manifests through two channels: (a) a direct effect driven by the esti- mated parameters β z , z = (1, 2, 3) in equation (1); (b) an indirect (general-equilibrium) effect induced by third countries’ adjustments. In our case, the typical example is the impact on trade between the United States and Brazil of a simulated deepening of MERCOSUR. The usual trade diversion effect will show up, which here depends on the content of the agreement. The multilateral resistance terms (MRT hereafter) à la Anderson and Van Wincoop (2003) act as general-equilibrium trade-cost indices transmitting local policy shocks to the overall matrix of trade frictions (these effects are formally described using the stan- dard gravity system of equations reported in supplementary online appendix S1). The inward MRT Pj on the importer side accounts for the impact on consumers and the outward MRT i for the impact on producers in the exporter country. Ultimately, the effects also spill over on the price of the exported variety (by the representative producer) and on the expenditure in the importing country. This corresponds to the general-equilibrium effects for an endowment economy (Head and Mayer 2014). After switching the PTA dummies to their new values reflecting the design of the counterfactual, in each simulation Xijt is predicted using the new matrix of PTAs, while constraining the coefficients β z , β T , and the μij of equation (1) to their initial values, to obtain counterfactual values for the MRTs and eventually solve for the associated general-equilibrium effects in an endowment economy.27 Starting from the baseline trade-cost matrix, tij,t , 3 ti1j− σ ,t = exp μi j + βz PTAz i j,t + β Transitory_PTAsi j,t z=1 + βT INTL_BRDRi j ∗ T + ln(Xi j,t /Xi j,t ) , 25 The last generation of these models relies on the properties of the PPML estimator (Silva and Tenreyro 2006) demon- strated by Fally (2015): the solution of the GE system of equations derived from a gravity model can indifferently be estimated (Anderson, Larch, and Yotov 2018; Fontagné and Santoni 2021) or computed with a solver (Head and Mayer 2014). And when the error term is in a multiplicative form (Anderson, Larch, and Yotov 2018), this is equivalent to the so-called “hat algebra” resolution, in line for instance with the approach coined as “trade theory with numbers” (Arkolakis, Costinot, and Rodríguez-Clare 2012)—and thus not fundamentally different from what a resolution of a CGE implies. 26 For comparison purposes, we provide in the six first columns of table S2.4 a sense of the impact of the extensive margin of regional integration, whereby EAP countries sign new PTAs of a different level of ambition with all countries with which they do not yet have an agreement. 27 In the year 2018, 0.06 percent of dyadic fixed effects cannot be identified due to separation: this is equivalent to 80 observations out of 12,544 (i.e. 112*112). As a value for the dyadic fixed effect for the general-equilibrium exercise is needed, as a measure of trade easiness between country pairs, we assume that the bilateral trade easiness value is equal to 0. 384 Fontagné et al. where the inclusion of the ratio between observed and predicted trade from equation (1), ln(Xi j,t /Xi j,t ), ensures a perfect fit for the observed trade flows; the equilibrium in each counterfactual derives directly from the adjustments in our system of equations induced by the change in the trade-cost vector, tij,t . In order to trace these effects we follow Yotov et al. (2017) and Fontagné and Santoni (2021), using the following notation:28 Yi the value of production in the exporting country, Ej the expenditure at desti- nation, and Y the value of world output. The variable Qi is the endowment (the quantity produced) by the Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 exporter country, pi the factory-gate price of the exporter, and φ i is related to the trade balance. The direct effect of a change in trade costs on trade flows between exporter i and importer j, Xij , can be inferred from the estimated coefficients of the structural gravity equation, holding the MRTs i and Pj constant. In turn, MRTs are impacted by the change in trade costs implied by our counterfactuals, because deepening an existing PTA or signing a new one between countries i and j will affect the overall matrix of trade costs and thus the structure of relative prices. These indirect effects, as well as their feedbacks on exporter and importer countries’ relative prices, concur in determining the final GE effects for an endowment economy. As the series of bilateral estimated fixed effects are the counterparts of the MRTs when relying on a PPML estimator (Fally 2015), we follow Yotov et al. (2017) and Anderson, Larch, and Yotov (2018) and solve our system of equations accordingly. The same approach pertains to our counterfactuals, whereby the system is solved with the alternative trade frictions derived from signing missing PTAs among countries in a region or alternatively deepening the already signed ones. We first recover β z (the average trade- cost elasticity over the period considered) and μij (the bilateral fixed effects) from the baseline gravity equation (1) including both intra-national and international yearly trade flows covering the period 1978– 2018; then we solve the counterfactual gravity system and compute the associated general-equilibrium indices. We solve the model using the “estibration” procedure (Yotov et al. 2017; Anderson, Larch, and Yotov 2018), which gives a solution identical to the “exact hat” algebra (Dekle, Eaton, and Kortum 2007). 5.3. Deepening Existing or New PTAs This section proceeds in four steps. Results are given in terms of relative variation, i.e. in percentage deviation in the counterfactual relative to the baseline scenario. Computations are executed on trade flows for the year 2018.29 In a first counterfactual, we switch all the existing agreements from their current level of ambition to the highest level of ambition and we compute the change in countries’ total exports and GDP. Individual effects are weighted by country GDP or export and add up to a 3.9 percent increase in exports and a 0.9 percent increase in GDP for the world economy as a whole. The results by country of such a generalized shift in deep trade agreements, hence at the intensive margin, are shown in fig. 5. Few countries are negatively affected by the global deepening of trade cooperation, but overall gains are significantly positive for most countries, including for many developing economies. In a second counterfactual, we shift region by region all the existing agreements of the region from their current level of ambition to the highest level of ambition and again compute the change in total exports and GDP for the countries in each region separately, before aggregating these results at the region level. Results for the different regions are reported in columns (1) and (2) of table 8. As an illustration, deepening all existing agreements among countries in the East Asia and Pacific region would lead to, on average, a 5.99 percent increase in their exports and a 1.05 percent increase in their GDP, compared to a baseline where these agreements have the current level of depth. This average increase is done in general equilibrium but at the country level before aggregation, and the detailed effects at the country level are shown in the first two columns in table S2.3 for this region. 28 See appendix S1 for a more detailed presentation 29 In general, reaching a deep trade agreement entails economic benefits but certainly also costs, particularly related to the heterogeneity of collective preferences across countries and regions. Our work aims to quantify the benefits by leaving out the assessment of political costs. The World Bank Economic Review 385 Figure 5. General-Equilibrium Effects of Deepening Existing PTAs Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Source: Note: The map shows the contribution of each country to the “world” effect as reported in columns (1) and (2) of table 8. Five countries report a moderate or insignificant contraction in GDP following a global deepening of trade cooperation: Senegal (0.3 percent), Palestine (0.5 percent), Russia (0.042 percent), Latvia (0.015 percent). For both GDP and exports, changes smaller than ±0.015 percent have been rounded to 0. 386 Fontagné et al. Table 8. General-Equilibrium Effects of Deeper Trade Integration by Region, Year 2018 All PTADeep Within region PTADeep With RoW PTADeep Export GDP Export GDP Export GDP Country name Iso3 (1) (2) (3) (4) (5) (6) East Asia & Pacific EAP 5.99 1.05 4.66 0.86 1.24 0.05 Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Europe & Central Asia ECA 1.66 0.72 1.14 0.55 0.50 0.08 Latin America & Caribbean LAC 8.74 3.83 2.79 0.73 4.68 1.45 Middle East & North Africa MENA 8.13 0.37 0.87 0.04 4.76 0.17 North America NA 6.35 1.06 2.91 0.65 2.15 0.23 South Asia SA 15.61 1.68 1.96 0.17 5.05 0.45 Sub-Saharan Africa SSA 3.83 0.80 0.39 0.10 1.93 0.37 Source: Authors’ calculation. Note: Percentage change compared to the baseline in total exports and GDP of participating countries. We assume σ = 5. The reference country for the normalization is South Africa. The list of PTAs for the reference country remains at the baseline. For each region, all the existing agreements of the region are shifted from their current level of ambition to the highest level of ambition. The variation in total exports and GDP for the countries in each region were computed separately and aggregated using GDP or exports weights. PTA stands for preferential trade agreement. RoW: Rest of the World. The third step addresses a different question: Is it preferable for a given region to deepen the PTAs signed within other countries in the region, or alternatively to deepen the PTAs signed with countries external to the region? Specifically, columns (3) and (4) of table 8 report the results for trade and GDP when all the existing agreements within each region of the world economy are switched from their current level of ambition to the highest level of ambition, keeping the content of all current extra-regional PTAs constant. In contrast, columns (5) and (6) present the results when all the existing agreements of each region with partners outside the region are switched from their current level to the highest level of am- bition, while keeping unaltered the content of PTAs with other regional partners. As before, we compute the change in total exports and GDP of each region to illustrate the results and we report the detail for the EAP region in columns (3) to (6) of table S2.3. This decomposition provides interesting new insights and helps explain the mechanisms at play. The economic effect of deepening existing trade agreements within regions is driven by two factors: first, the actual level of ambition of existing agreements within the different regions, as opposed to between regions, and second, the economic size of trading partners with whom the deepening of an existing agreement is envisaged. Given the geography and depth of the existing agreements, the economic effects are very different across regions. At one extreme, most of the potential gains come from deepening intra-regional PTAs, as in the East-Asia and Pacific case where agreements in the baseline are relatively shallow and the economic size of partners is significant. At the other extreme, all the gains come from deepening agreements with partners outside the region. This is the case for a region like the Middle East and North Africa, where larger markets are predominantly extra-regional. In between these situations, Latin America, Africa, South Asia, and Europe and Central Asia would have similar gains in deepening trade agreements within or outside the region. This section ultimately turns to a different question. Instead of deepening existing agreements, new agreements with a different level of ambition are signed, allowing the impact of the extensive margin of regional integration to be addressed in a fourth step. Columns (1) to (6) in table S2.4 show the trade and GDP impact of EAP countries signing all missing agreements within or without the region. The results are detailed by country in the region and aggregated in the penultimate row of the table. The last row shows the impact on countries out of the region. On the extensive margin as well, signing a deep agreement makes a big difference: the average impact on exports of EAP countries is 17.8 percent and the impact on GDP is 1.8 percent, to be compared with respectively 4.7 percent and 0.4 percent should new shallow The World Bank Economic Review 387 agreements be signed. A side result is illustrated by comparing columns (3) and (4) with columns (7) and (8) respectively. 6. Conclusion This paper uses new data on the content of trade agreements and a structural gravity general-equilibrium Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 model to quantitatively assess the economic impacts of deepening trade agreements. Based on a clustering of 278 PTAs, comprising 910 provisions grouped in 18 policy areas, it shows that PTAs of higher ambition are associated with a much larger trade elasticity to PTAs. This finding confirms that using an average effect of PTAs, disregarding the depth and content of trade agreements, is misleading. The simulation of a series of full general-equilibrium counterfactual situations for endowment economies reveals the economic impacts of deepening all existing PTAs, and of deepening trade agreements within regions and between regions. These exercises focusing on the intensive margin of PTAs suggest that deepening trade agreements can have significant effects on trade and GDP. The effects of deepening trade agreements within and between regions vary across countries and world regions, depending on the initial depth of trade agreements and on the size of regional markets, but are largely positive. Potential differences in collective preferences between regions, which make negotiating deep trade agreements more difficult, suggest that for most countries the low-hanging fruits may be in deepening trade agreements within regions. Data Availability Lionel Fontagné thanks the support of the EUR grant ANR-17-EURE-0001. References Anderson, J. E., M. Larch, and Y. V. Yotov. 2018. “GEPPML: General Equilibrium Analysis with PPML.” World Economy 41(10): 2750–82. Anderson, J. E., and E. Van Wincoop. 2003. “Gravity with Gravitas: A Solution to the Border Puzzle.” American Economic Review 93(1): 170–92. Anderson, J. E., and Y. V. Yotov. 2016. “Terms of Trade and Global Efficiency Effects of Free Trade Agreements, 1990–2002.” Journal of International Economics 99: 279–98. Arkolakis, C., A. Costinot, and A. Rodríguez-Clare. 2012. “New Trade Models, Same Old Gains?” American Eco- nomic Review 102(1): 94–130. Arthur, D., and S. Vassilvitskii. 2007. “k-Means++: The Advantages of Careful Seeding.” In Proceedings of the Eigh- teenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’07, USA, pp. 1027–35. Society for Indus- trial and Applied Mathematics. Baier, S. L., and J. H. Bergstrand. 2004. “Economic Determinants of Free Trade Agreements.” Journal of International Economics 64(1): 29–63. ———. 2007. “Do Free Trade Agreements Actually Increase Members’ International Trade?” Journal of International Economics 71(1): 72–95. Baier, S. L., Y. V. Yotov, and T. Zylkin. 2019. “On the Widely Differing Effects of Free Trade Agreements: Lessons from Twenty Years of Trade Integration.” Journal of International Economics 116: 206–26. Blanchard, E. J., C. P. Bown, and R. C. Johnson. 2016. “Global Supply Chains and Trade Policy.” Technical report, Working Paper NBER 21883. Blanga-Gubbay, M., P. Conconi, and M. Parenti. 2020. “Lobbying for Globalization.” CEPR Discussion Paper 14597. Borin, A., and M. Mancini. 2023. “Measuring What Matters in Value-Added Trade,” Economic Systems Research, DOI: 10.1080/09535314.2022.2153221. Bown, C. P., P. Conconi, A. Erbahar, and L. Trimarchi. 2021. “Trade Protection along Supply Chains.” mimeo Bown, C. P., A. Erbahar, and M. Zanardi. 2021. “Global Value Chains and the Removal of Trade Protection.” Euro- pean Economic Review 140: 103937. 388 Fontagné et al. Dekle, R., J. Eaton, and S. Kortum. 2007. “Unbalanced Trade.” American Economic Review 97(2): 351–55. Eaton, J., S. Kortum, and S. Sotelo. 2013. International Trade: Linking Micro and Macro, Volume 2 of Econometric Society Monographs, pp. 329–70. Cambridge University Press. Egger, P. H., M. Larch, and Y. Yotov. 2022. “Gravity Estimations with Interval Data: Revisiting the Impact of Free Trade Agreements.” Economica 89(353): 44–61. Egger, P. H., and S. Nigai. 2015. “Structural Gravity with Dummies Only: Constrained Anova-Type Estimation of Downloaded from https://academic.oup.com/wber/article/37/3/366/7068456 by World Bank and IMF user on 14 September 2023 Gravity Models.” Journal of International Economics 97(1): 86–99. Fally, T. 2015. “Structural Gravity and Fixed Effects.” Journal of International Economics 97(1): 76–85. Fontagné, L., H. Guimbard, and G. Orefice. 2022. “Tariff-Based Product-Level Trade Elasticities.” Journal of Interna- tional Economics 137: 103593. Fontagné, L., and G. Santoni. 2021. “GVCs and the Endogenous Geography of RTAs.” European Economic Review 132: 103656. Head, K., and T. Mayer. 2014. Gravity Equations: Workhorse, Toolkit, and Cookbook, Volume 4 of Handbook of International Economics, Chapter 3, pp. 131–95. Elsevier. ———. 2021. “The United States of Europe: A Gravity Model Evaluation of the Four Freedoms.” Journal of Economic Perspectives 35(2): 23–48. Johnson, R. C. 2018. “Measuring Global Value Chains.” Annual Review of Economics 10(1): 207–36. Kharel, P. 2019. “The Effect of Free Trade Agreements Revisited: Does Residual Trade Cost Bias Matter?” Review of International Economics 27(1): 367–89. Limão, N. 2016. “Preferential Trade Agreements.” In Handbook of Commercial Policy, 279–367. Elsevier. Mattoo, A., A. Mulabdic, and M. Ruta. 2022. “Trade Creation and Trade Diversion in Deep Agreements.” Canadian Journal of Economics/Revue canadienne d’économique 55(3): 1598–1637. Mattoo, A., N. Rocha, and M. Ruta. 2020. Handbook of Deep Trade Agreements. Number 34055 in World Bank Publications. The World Bank. Rousseeuw, P. J. 1987. “Silhouettes: A Graphical Aid to the Interpretation and Validation of Cluster Analysis.” Journal of Computational and Applied Mathematics 20: 53–65. Silva, J. S., and S. Tenreyro. 2006. “The Log of Gravity.” Review of Economics and Statistics 88(4): 641–58. Sotelo, S. 2019. “Practical Aspects of Implementing the MultinomialPML Estimator.” mimeo. Weidner, M., and T. Zylkin. 2021. “Bias and Consistency in Three-Way Gravity Models.” Journal of International Economics 132: 103513. Yotov, Y. V. 2012. “A Simple Solution to the Distance Puzzle in International Trade.” Economics Letters 117(3): 794–98. ———. 2022. “On the Role of Domestic Trade Flows for Estimating the Gravity Model of Trade.” Contemporary Economic Policy 40(3): 526–40. Yotov, Y. V., R. Piermartini, J.-A. Monteiro, and M. Larch. 2017. An Advanced Guide to Trade Policy Analysis: The Structural Gravity Model. United Nations Publication. World Trade Organization.