Trade Creation and Trade Diversion in Deep Agreements

Preferential trade agreements have boomed in recent years and extended their reach well beyond tariff reduction, to cover policy areas such as investment, competition, and intellectual property rights. This paper uses new information on the content of preferential trade agreements to examine the trade effects of deep agreements and revisit the classic Vinerian question of trade creation and trade diversion. The results indicate that deep agreements lead to more trade creation and less trade diversion than shallow agreements. Furthermore, some provisions of deep agreements have a public good aspect and increase trade also with non-members.


I. Introduction
If a trade economist were abruptly woken up by somebody shouting, "preferential trade agreements" (PTAs), their first thought is likely to be "trade creation and trade diversion". 3 That is a measure of the influence of Jacob Viner's classic book The Customs Union Issue (Viner, 1950) on the profession and the policy debate on the trade effects and, hence, the desirability of preferential arrangements. However, Vinerian analysis was developed in a world where trade agreements were "shallow" and focused only on bilateral tariff liberalization.
Today, PTAs are increasingly "deep" and cover also behind-the-border policy areas, such as competition policy, intellectual property rights and other regulatory issues. 4 In this paper, we empirically investigate how far classic Vinerian logic helps us to understand the trade effects of modern preferential trade arrangements.
Do deep agreements simply lead to more trade creation and more trade diversion than shallow agreements? Intuitively, Vinerian logic does not fully apply to deep agreements because their nature is in part different from shallow PTAs. Shallow agreements are controversial because they are inherently discriminatory. Members grant tariff concessions to each other, leaving tariffs on imports from non-members unconstrained. The resulting tariff preferences are likely to increase trade between members (trade creation), but they can also lead members to substitute imports previously sourced from non-members for within PTA products (trade diversion). 5 Deep agreements can reduce trade costs and discrimination beyond tariff liberalization and hence are expected to lead to even more trade creation. But differently from tariffs, provisions relating to competition policy or subsidies tend to be nondiscriminatory in nature and may reduce trade costs and discrimination also vis-à-vis outsiders, creating a positive spillover effect, or "negative" trade diversion (Baldwin and Low, 2009;Baldwin, 2014). Ultimately, the verdict on what forces dominate is empirical and will crucially depend on the content of the trade agreements. 3 We refer to PTAs as any trade agreement between a subset of countries (two or more). PTAs have been also referred to in the literature as Free Trade Agreements, Regional Trade Agreements, Economic Integration Agreements, etc. As we will further clarify below, we will also use the term Deep Agreements to stress the fact that many of these arrangements have features that go beyond trade policy and are not preferential in nature. 4 The terms "shallow" and "deep" trade agreements were first defined in Lawrence (1996). There is a voluminous literature on the purpose of shallow trade agreements (e.g. Grossman, 2016). The rationale for deep agreements has not received the same attention. Two references that help explain the changing scope of trade agreements include Ederington and Ruta (2016) and Maggi (2016). 5 As it is well known, preferential tariff liberalizations have an ambiguous welfare effect. Trade creation is welfare improving for members. Trade diversion has a negative impact on the welfare of non-members through lower market access as well as on members through reduced tariff revenue. The net welfare effect of shallow PTAs, therefore, depends on which of these two forces dominates.
To empirically address the question of the trade effects of deep agreements, we exploit a new database on the content of PTAs (Hofmann et al., 2017). Since the early 1990s, a large number of trade agreements have entered into force. Focusing on the PTAs still in force in 2015, the number of preferential arrangements increased from 20 in 1990 to 279 at the end of 2015. The content of PTAs too has changed. Newer agreements are "deeper" in the sense that they generally expand the set of policy areas covered by older agreements. Specifically, older PTAs focused on fewer than 10 policy areas, mostly commitments on tariffs on industrial and agricultural goods and other border measures such as export taxes. As agreements become deeper, they increasingly extend their reach first to areas such as trade remedies (i.e. countervailing measures, antidumping duties) and subsidies and then to a broader set of behind the border measures such as intellectual property rights and standards.
To assess the impact of deep agreements on members' and non-members' trade, we augment a standard gravity model, which is widely used in the literature to assess the effects of PTAs on trade flows (see Head and Mayer, 2014;Limão, 2016). We include a variable of depth of agreements between PTA members, and a variable that captures the depth of the agreements of a trading partner with other countries. Using information from the content of PTAs database, we construct different measures of depth based on the policy areas regulated by the agreements and their legal enforceability. As standard in the literature, we include importer and exporter-year fixed effects to control for country-year specific shocks and for the multilateral resistance terms, and we introduce country-pair fixed effects to partially address endogeneity concerns (Baier and Bergstrand, 2007). 6 We also include measures of bilateral tariffs and preference margins (Kee et al. 2008(Kee et al. , 2009Fugazza and Nicita, 2013) to isolate the impact of changes in depth from changes in tariffs. Finally, we include additional controls, such as dummies to identify shallow PTAs, PTAs that are no longer in force, and the presence of other international agreements that may affect bilateral trade. Our sample covers 96 countries, including all major economies, for the period 2002-2014. 7 During this period, the share of country pairs with PTAs increased from 9 to 29 percent, average tariffs were cut by half, while depth (measured as the count of provisions included in the PTA) increased by a factor of three. 6 See Piermartini and Yotov (2016) for a useful guide on estimating trade policy effects with structural gravity models. 7 The country coverage is determined by the availability of comprehensive tariff data for the entire period.
We find that the formation of deep agreements has a meaningful positive impact on the trade flows among members. In particular, we find that trade between country pairs that sign a deep agreement increases by 44 percent. As we control for tariffs and for a PTA dummy, the estimate suggests that deep provisions induce more trade creation than shallow PTAs. When we look at the dynamic effects of deep agreements, we find that future levels of PTA depth are statistically uncorrelated with current levels of trade flows, suggesting that depth of agreements is not determined by the closeness of current trade relations. On average, it takes two years for deep agreements to increase trade flows, consistently with the evidence that reforms of behind the border measures take time to be implemented. This paper contributes to a large body of literature on the trade effects of preferential trade arrangements by including the notion of "depth" in the analysis of PTAs. 8 Previous work in this area suffers from a well-known measurement error problem (Baier and Bergstrand, 2007). Due to lack of data, most studies use dummies to identify the presence of a PTA or distinguish between broad types of trade arrangements (e.g. partial scope agreements, free trade agreements or custom unions, as in Baier et al., 2014). This approach does not adequately capture the variation in the content of preferential trade agreements. Indeed, we show that this variation has important implications for the effects of PTAs both on members' and nonmembers' trade flows. Our analysis also has relevant implications for the longstanding debate on regionalism versus multilateralism (Bhagwati, 1993). A key question in this debate is whether PTAs are building blocks or stumbling blocks of the multilateral trade system. Both formal models and empirical studies in this literature assume that PTAs are mostly about tariff liberalization. 9 The positive impact of deep PTAs on members' and non-members' trade that we find in our analysis supports the view that deep provisions in trade agreements can complement rather than undermine the world trading system (WTO, 2011).
The rest of the paper is organized as follows. The next section discusses the database on the content of trade agreements and the other data used in the analysis. Section 3 presents the empirical strategy. Econometric results are presented in Section 4. Concluding remarks follow.

II. Data
Our measures of depth of preferential trade agreements are based on detailed information on the content of PTAs from a new database (Hofmann et al., 2017). The database covers 279 treaties, which are all the preferential agreements notified to the WTO and in force up to December 2015. 10 Following the methodology proposed by Horn et al. (2014), the focus is on 52 policy areas (see Table A1 in Appendix A), divided into areas that are currently under the mandate of the WTO such as tariffs, antidumping duties and subsidies (referred to as "WTO 8 For recent surveys, see Freund and Ornelas (2010), WTO (2011), Head and Mayer (2014) and Limão (2016). There is a small literature on deep agreements. Osnago, Rocha and Ruta (2017a) look at the impact of deep agreements on countries' participation in global value chains. Mulabdic, Osnago and Ruta (2017) study the effect of Brexit (i.e. the undoing of a deep agreement) on future EU-UK trade relations. Other studies that have looked at the impact of deep agreements based on a more limited database covering around 100 PTAs are Orefice and Rocha (2014) and Osnago, Rocha andRuta (2015 and2017b). 9 Informal arguments on the relationship between deep PTAs and the multilateral trade system have been made in Baldwin andLow (2009) andWTO (2011). 10 The data are freely accessible at http://data.worldbank.org/data-catalog/deep-trade-agreements. +") and areas where the WTO has no comprehensive mandate such as investment and competition policy ("WTO X"   (2017) refer to the expanding scope of PTAs as "horizontal depth". Another dimension of the depth of a trade agreement is "vertical", reflecting the liberalizing content of commitments or the stringency of rules. This information, however, is widely available only for tariffs (see below) and a small subset of policy areas. agreements covering fewer than 10 policy areas dominated up to the late 1990s), an increasing share of PTAs over time has tended to cover a larger number of policy areas, suggesting a deepening of trade agreements.  Table 1 shows that there is an ordering in terms of which provisions are included in trade agreements with different values of depth. Specifically, we divide the agreements into three categories, based on the number of legally enforceable provisions and calculate the share of agreements that include each policy area. We find that policy areas included in shallower agreements ("Less than 10"), are at least as frequent in deeper agreements (cooperation on "statistics" is an exception). The majority of these agreements tend to cover tariffs and other border measures such as export taxes and customs. Competition policy is the only policy area outside the mandate of the WTO appearing in a majority of shallower PTAs. As agreements become deeper ("Between 10-20"), they increasingly extend their reach to a broader set of WTO + areas, including state aid, anti-dumping and countervailing measures. Finally, deeper agreements ("More than 20") tend to cover areas related to intellectual property rights, movement of capital, and standards, in addition to the areas covered by shallower agreements. 14 14 Figure A1 in Appendix A shows that recent agreements signed by the US and the EU include a larger number of areas than earlier agreements. These new areas were often covered in other countries' earlier PTAs, suggesting there may be learning from other countries' PTAs.  1970 1972 1974 1976 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014

Cumulative Number of Agreements Number of Agreements
More than 20 Between 10 and 20 Less than 10 Cumulative Cumulative Based on this evidence, we build several measures of the depth of trade agreements which reflect the extent to which the different policy areas are covered and legally enforceable in a PTA. 15 An area is considered as weakly legally enforceable if the language used is sufficiently precise and binding, but the area has been excluded from dispute settlement procedures under the PTA. While strong legal enforceability refers to areas where the language used is sufficiently precise and binding, and if the area is subject to dispute settlement procedures under the PTA. Using this information, we define alternative measures of depth of an agreement. Specifically, the depth variables are equal to the count of all ("depth all"), weakly legally enforceable ("depth wle"), or legally enforceable ("depth le") provisions included in an agreement. Each measure is normalized between 0 and 1, with 1 indicating the agreement with the highest number of provisions. In characterizing trade agreements, we also consider the policy areas that have been identified in the literature as being more economically relevant ("core" provisions). These core provisions include all WTO + areas and four areas that fall outside the domain of the WTO: competition policy, rules on investment, movements of capital, and intellectual property rights protection. 16 As shown in Hofmann et al. (2017), these policy areas are also the ones that appear more frequently in PTAs.
A useful distinction for our subsequent discussion is between discriminatory and nondiscriminatory policy areas. Here we follow Baldwin and Low (2009) to classify PTA provisions in these two groups. The traditional view of PTAs is that their benefits accrue only to PTA partners. This is indeed the case for traditional trade policies (i.e. tariffs on industrial goods, tariffs on agricultural goods, export taxes, countervailing measures and antidumping duties) that can be implemented on a discriminatory basis based on the origin of the product.
Similarly, government procurement provisions in PTAs tend to open state purchasing to foreign firms on a strictly preferential basis. For other policy areas, however, the traditional view does not appear to hold as PTA provisions may improve the conditions of access in a nondiscriminatory manner (i.e. on a "most-favored-nation," or MFN basis). According to Baldwin and Low (2009), these areas include customs administration, domestic regulation (SPS and TBT measures), competition (state trading enterprises, competition policy), services (GATS), investment (TRIMS and investment rules), property rights (TRIPS and IPR protection), and 15 Given the fact that provisions tend to be highly correlated with each other (see Table A2 of Appendix A), regressions that include individual variables indicating the presence of each provision would suffer from the problem of multicollinearity. 16 Core areas have been identified in Damuri (2012) based on Baldwin (2008).
rules on subsidies and on movements of capital. In some cases, discrimination is simply not possible: if a country limits subsidies to domestic producers or establishes a competition authority in fulfillment of its PTA commitments, these reforms benefit both members and nonmembers of the PTA. In other cases, discrimination is feasible but unlikely for economic or legal reasons: in services, market access is generally granted through reforms of domestic regulation, such as rules on foreign participation or access to essential facilities, which are hard to undertake in a way that grants privileged access.
The Before moving to the econometric analysis, we take a first look at the data. Over the 2002-2014 period, the share of country pairs with PTAs increased from 9 to about 29 percent (Table 2). During the same period, average tariffs (TTRI) were cut by half while depth, irrespective of legal feasibility, increased by a factor of three. As countries reduced bilateral tariffs, the average relative preference margins (RPM) and its standard deviation decreased as well. The two trends together indicate widespread tariff reductions which are less likely to have increased trade diversion. In terms of the content of PTAs, the summary statistics show that there were minor differences (before 2014) between depth constructed using legally enforceable provisions subject to dispute settlement ("Depth LE" or "Depth Legally Enforceable"), and depth constructed on the basis of legally enforceable language ("Depth WLE" or "Depth Weakly Legally Enforceable"). There is also some evidence that the newest agreements tend to be deeper. The average maximum depth ("max Depth LE" or "max Depth Core LE") by importer almost doubled from 2002 to 2014. Part of these increases are due to countries signing agreements for the first time, but this trend is also observed when we restrict the sample to country pairs which already have a PTA.   Figure 2 shows that country-pairs with higher levels of depth trade more on average. The right panel shows that on average, countries export relatively less to partners involved in shallow agreements (i.e. "Low depth") than partners without PTAs.
However, this negative effect is reversed as partners sign deeper agreements ("Medium depth" and "High depth") which are associated with distributions shifted to the right of the "no PTA".
This suggests that deep agreements tend to benefit excluded countries as well, possibly due to the inclusion of provisions that are de jure or de facto MFN.

III. Trade effects of deep agreements: Empirical strategy
In this section, we begin the empirical investigation of the trade impact of deep agreements. A number of policy-related factors contribute to trade costs between countries, which create a gap between the price in the importing country and the export price. Trade agreements allow members to reduce these costs and hence increase bilateral trade. A concern, well understood since Viner (1950), is that this mechanism could also generate trade diversion, that is, a substitution of trade away from non-members. Deep agreements can reduce trade costs among members by eliminating tariffs and by reducing other frictions. Examples of the latter are contingent protection measures like antidumping, countervailing and safeguard actions, and differences in national regulations that create an adaptation cost for foreign producers. Even other provisions of PTAs, such as disciplines on subsidies or strengthened protection of intellectual property rights, reduce the risk of exporting due to policy uncertainty, and hence can be seen as reducing trade costs (Limão and Maggi, 2015). 18 Therefore, we expect PTAs 18 The assumption that trade agreements reduce trade costs helps us to cast the following discussion in the framework of the gravity model. We recognize that other provisions of agreements, such as those relating to labor or environmental standards, do not necessarily lead to a reduction in trade costs. The extent of the aggregate impact of these heterogeneous provisions is, therefore, an empirical question. We come back to this issue below. The impact of a deep agreement on non-members is more complicated. The rules in a deep PTA can be implemented either to reduce costs only for members (e.g. by exempting only them from burdensome regulatory requirements) or also for non-members (e.g. by simplifying customs procedures for all trading partners). If these rules are implemented in a discriminatory way, they inflict a further competitive disadvantage on third countries. Since member countries must now pay neither tariffs nor frictional costs, they can expand sales in their markets, driving down prices and hurting exports of third countries. However, if frictional barriers are eliminated in a non-discriminatory way, third countries also benefit from the reduction in associated costs. 19 In these circumstances, third countries still suffer from the decline in price in destination markets due to preferential access granted to members of PTAs, but the price they actually receive is closer to the destination price because the elimination of the frictional costs reduces the total trade tax they pay. If the decline in trade costs for non-members is sufficiently large relative to the preferences members receive, then we may observe "negative trade diversion" (Baldwin, 2014): third countries see an increase in the export price they receive and expand quantity exported as a result of a deep PTA. 20 Below, we introduce the empirical model and identification strategy used to analyze the effect of deep agreements on members' and non-members' trade. In line with the above discussion, we augment a standard gravity model to include a variable of depth between PTA members and another variable that captures the depth of agreements trading partners conclude with the rest of the world. We also use information on relative tariff preferences (Fugazza and Nicita, 2013) to assess how their impact is affected by existence of deep agreements.

a. Trade creation
Our main specification is based on the gravity model of trade, which is widely used in the literature to assess the effects of policy variables on trade flows (see Head and Mayer, 2014;Limão, 2016). We begin by discussing how the depth of PTAs can be incorporated into the 19 There is some evidence of these positive externalities. Chen and Mattoo (2008) examine the consequences of harmonization and mutual recognition of standards within PTAs. They show that when these agreements are concluded with restrictive rules of origin which deny their benefit to non-members, the latter suffer a decline in exports to PTA countries. However, when the agreements do not have restrictive rules of origin, nonmembers' exports to PTA countries also increase. 20 Appendix B provides a graphical example of these effects. standard gravity framework. As shown in Costinot and Rodríguez-Clare (2013), the following gravity equation emerges from different theoretical frameworks: where is the bilateral trade flow from country to country , is country j's total expenditure, = ∑ is country i's income, is the trade elasticity with respect to variable trade costs , and is a function of structural parameters distinct from .
We can define trade costs as a collection of different components: where is the ad-valorem import tariff imposed by country on goods imported from , are the iceberg trade costs that the exporter incurs to ship to country . Since deep provisions in PTAs could lower the policy frictions that limit international trade, we account for the term in the empirical model by including a measure of the depth of an agreement between country-pairs and .
Taking the log of both sides of equation (1) and using tariffs and depth to proxy for trade costs in equation (2), we obtain the following modified gravity equation which accounts for the depth of trade agreements as a determinant of bilateral trade: where are bilateral exports from country i to country j in year t and ℎ is a measure of the PTA depth between i and j (normalized between 0 and 1). As discussed in Section II, we use different definitions of depth based on the legal enforceability and the economic relevance of the policy areas covered in the agreement. and are importer-year and exporter-year fixed effects, respectively, that control for any country-year specific shocks and also for the theoretically motivated multilateral resistance. As shown in Baldwin and Taglioni (2006), failing to account for the country-specific time-varying multilateral resistance biases downward the effects of PTAs, or in our case the effect of ℎ on trade. Finally, we include several additional controls: dummies to capture the presence of a PTA (i.e. a shallow PTA dummy), of a PTA no longer in force, or of other international agreements that can have an impact on trade flows, such as a Bilateral Investment Treaty (BIT).
An important issue in the estimation of the effects of any policy variable is endogeneity.
In the trade literature, it has been shown that countries are more likely to sign agreements with partners with whom they already trade more intensively because of geography or cultural proximity or other common characteristics. If countries tend to sign trade agreements with their "natural" trading partners (Krugman, 1991), this would bias the effects of trade agreements upwards especially with cross-sectional data. This bias may be even stronger for depth to the extent that countries may be more willing to sign deeper agreements with their natural trading partners. The issue of endogeneity of trade policies is well known since Trefler (1993), but it is hard to address due to the lack of reliable instruments for panel data. To partially address the endogeneity problem, we follow Baier and Bergstrand (2007) and introduce country-pair fixed effects, , to capture country-pair time-invariant factors determining bilateral trade such as distance or common language. This set of fixed effects accounts for unobserved time-invariant heterogeneity among country pairs which can bias estimates in cross-sectional studies, and hence attenuates the endogeneity bias stemming from omitted variables.
A limitation of previous work is that the use of a dummy variable to identify the trade effect of a PTA is generally associated with a negative bias in the variable's coefficient. We improve with respect to earlier studies on the bias due to measurement error of the trade policy variables by following the suggestion outlined in Baier and Bergstrand (2007)  we include a variable for the depth of trade agreements to capture the degree of trade liberalization between PTA partners. Second, we also include the , the tariff trade restrictiveness index, to isolate the effect of changes in tariffs between country i and j (Kee et al. 2008(Kee et al. , 2009Fugazza and Nicita, 2013) from the impact of changes in depth. The index is obtained using the following formula: where is the average product level exports from country i to country j between 1995 and 1997, is the bilateral import elasticity and is the applied tariff rate on product ℎ . We use export weights based on pre-sample data to reduce the potential endogeneity problem of trade to tariff. The obtained index aggregates bilateral product level tariffs to a uniform tariff equivalent that would maintain exports between i and j constant.
As discussed in Section II, the ℎ variable is defined as the count of provisions included in each agreement normalized between 0 and 1. Our baseline specification relies on the count of legally enforceable provisions, i.e. those which have binding language and are subject to dispute settlement ("depth LE"). We also construct alternative measures of depth by counting the areas covered irrespective of their legal enforceability ("depth all") and by including provisions that are more likely to be economically relevant ("depth core"). 21 The coefficient of depth captures the effect of changes in the coverage of areas in a PTA net of changes in tariffs. Given the set of fixed effects, the identification strategy relies on the variation in depth within country-pairs variation to identify the effect on exports.

b. Trade diversion
To capture effects on a trading partner i from country 's trade agreements, we modify the definition of trade cost in equation (2) to where iceberg trade costs are divided into an "MFN" component, when ≠ , which is a destination specific cost common to all exporters, and that can be eliminated between specific country-pairs. Deep agreements affect non-members in two different ways.
First, as in the case of shallow PTAs, they make non-members less competitive in members' countries by reducing bilateral trade costs of members. This effect results both from the preferential reduction in tariffs ( ) and of other trade costs ( ). Second, deep agreements 21 We also construct a depth variable based on the first component of a Principal Component Analysis (PCA) of the provisions (see Orefice and Rocha, 2014). can have a positive impact on non-members to the extent that they reduce the MFN component of trade costs ( ).
To capture the trade effects of deep agreements on non-members we proceed by steps.
First, we augment equation (3) to include the average depth and relative tariffs for each importing partner with respect to the rest of the world. 22 Thus, equation (3) , is the average tariff the rest of the world is facing at the HS product level, which is then aggregated at the country pair level by weighting each product by country i's exports to country j during the 1995-1997 period to avoid endogeneity. Note that we can retain importer-year fixed effects because both RPM and Others Depth vary by origin country i: RPM more obviously because it incorporates the tariff faced by source country i; Others Depth because it is calculated for any ij pair by taking the weighted average of j's depth vis-à-vis all countries except i. Intuitively, if trading partner j gives better market access to countries that export 22 We assume that = ( ℎ ). In particular, to keep the functional form similar to , we proxy for by country j's trade weighted depth with the rest of the world. For a theoretical derivation of the RPM, see Fugazza and Nicita (2013). goods that are important for i we would expect country i's exports to decrease; similarly, if j signs deep agreements with i competitors, this should have an impact on bilateral trade.
As a second step, we decompose the depth of the PTA into its preferential and MFN components, as suggested in the literature (e.g. Baldwin and Low, 2009) Apart from their direct impact on third countries, deep provisions in agreements may also influence the impact on these countries of conventional tariff preferences. How an MFN reduction in the frictional trade tax for all trading partners influences the marginal effect of tariff preferences on third countries is analytically ambiguous. Therefore, it is worth examining the empirical evidence. We test the following equation: where ℎ is interpreted in two different ways. The first is, as before, an average of the depth of the importers' agreements with the rest of the world. The second is the maximum number of provisions that importer j has in its deepest agreement at time t. This captures the idea that MFN provisions, once introduced in a PTA, may have an impact on all partners because of their intrinsic public good nature. Coefficient of the interaction term in equation (3'') identifies the effect of deep agreements on tariff preferences. A negative coefficient would suggest that tariff preferences have a stronger marginal effect once the importing country signs deeper agreements, whereas a positive coefficient would suggest that tariff preferences matter less when trading partners implement deep agreements.

IV. Econometric results
In this section, we present the results of the estimations from the gravity model. The first subsection focuses on the impact of deep agreements on members' trade. We then study how deep agreements affect trade with non-members.

a. Trade creation
This subsection discusses and presents the estimates from equation (3) and its extensions. The objective is to identify the effect of deep trade agreements on member countries' trade flows. Table 3 reports the PPML estimates from the gravity equation ( 24 In a series of robustness checks, we find similar results when controlling for the presence of bilateral investment treaties (BITs) and using alternative definitions of depth based on the legal language. Results for depth We find that controlling for old PTAs (columns 5 to 7), agreements that are no longer in force and on the content of which we have no information, increases the magnitude and statistical significance of the impact of depth on trade. Specifically, in this specification trade between country pairs that sign an agreement with the highest depth increases by 44 percent.
Intuitively, the inclusion of the old PTA variable increases the magnitude and precision of the depth estimates because it allows us to distinguish between country-pairs in the control group that had a PTA at some point in time and those that never had a PTA and for which depth is equal to zero. 25 core LE become insignificant in a specification where we include controls for old PTA, PTA, and bilateral tariffs together, for which there is limited variation within country-pairs. Finally, results are robust to an alternative definition of depth based on the principal component analysis "PCA," as in Orefice and Rocha (2014). Moreover, we find that results are also robust to the exclusion of crises years (i.e., 2008 and 2009) with coefficients around 10 percent higher than those in specifications using the full sample. 25 The old PTA dummy captures the pre-accession agreements for countries that joined the EU after 2002 and a small number of other PTAs no longer in force: the trade agreement between Mexico and the three Northern Triangle countries -El Salvador, Guatemala and Honduras -that was active between 2001 and 2012, the trade agreement between Mexico andNicaragua (1998-2012), and Closer Economic Partnership Arrangement (CEPA) It is important to note that with the inclusion of a PTA dummy in columns 6 and 7 of Table 3, we capture the effect of depth due to variations within country-pairs and within PTAs.
The PTA dummy could be interpreted either as a trade agreement fixed effect or as an interaction variable that captures the effect of an agreement with zero provisions. Therefore, a positive and significant coefficient indicates that country-pairs with deep agreements trade more with respect to those that have shallower agreements. This suggests that results in the first four columns are not merely due to the presence of a PTA. Moreover, results are robust to the inclusion of bilateral tariffs as well, which suggests that the finding that deep trade agreements increase bilateral trade is not driven by tariff liberalization.
Our baseline specification could suffer from several econometric problems. On the one hand, the relatively small effects of ℎ on trade we find in Table 3 compared to the literature may suggest a downward bias in our coefficients. 26 On the other hand, the potential endogeneity of deep agreements and trade could bias our estimates in the opposite direction.
More specifically, a first econometric issue could be that trade flows tend to adjust slowly to trade cost changes and by using annual data without lags we may not capture the full effect of trade agreements. 27 The issue may be particularly relevant for deep agreements as they tend to have longer implementation phases, which could bias downwards the estimates of depth.
Second, coefficients of the anticipation effects of PTAs may also be a confounding factor in our regression analysis. If trade flows increase in anticipation of the agreement even before its entry into force, we would fail to assign these effects to the agreements when using contemporaneous variables. Evidence of anticipatory effects could also raise concerns about the identification strategy and causality because of the difficulty in distinguishing between anticipation effects and pre-existing trends. Finally, a third concern is that the absence of intrabetween China and Hong Kong SAR, China. The coefficient on old PTAs in column 5 is 0.18 which suggests that older PTAs increased bilateral trade by around 20 percent on average during the 2002-2014 period. This impact is equivalent to signing an agreement that includes 30 legally enforceable provisions and reflects the depth of EU pre-accession agreements. 26 According to a meta-analysis of the effect of trade agreements on trade by Head and Mayer (2014), the median coefficient of a PTA dummy is 0.28. In the case of deep trade agreements such as the EU, the coefficient found in the literature is 0.98, which is higher than our estimate for the trade effect of the agreement with highest depth. 27 Trefler (2004) suggests that trade flows adjust slowly to changes in trade costs and criticizes the use of yearly data. Therefore, we use 3-year intervals to allow more time for trade to adjust to changes in depth of trade agreements. We find that results in Table A3, in Appendix A, are qualitatively and quantitatively similar to results obtained using consecutive years, with the exception for core depth which becomes statistically insignificant. In the rest of the paper, we favor yearly data over 3-year intervals because it is more common in the gravity literature. national trade flows limits the identification to the comparison between PTA member countries and country-pairs without PTAs in a way that is not completely consistent with the theoretical basis of the gravity equation (Larch et al., 2017).
To address these concerns, we extend our baseline specification in equation (3) in several directions. First, we use yearly data and include lags and leads of the depth variables to estimate the dynamic effects of PTAs. Figure 3 presents the results on the dynamic effects of deep agreements. Results are based on specifications (1) to (4) in Table 3, modified to include two leads and four lags of the depth variables to accommodate heterogeneous effects over time and to test for anticipatory effects of agreements. The inclusion of the leads thus also provides for an informal test for the "strict exogeneity" of trade agreements (see Bergstrand et al., 2015). 28 The results suggest that both current and future levels of PTA depth are statistically uncorrelated with current levels of trade flows. It takes at least two years for a deep agreement to increase trade flows and the effects are twice as large as the ones we find in specifications without lags and leads (Table 3). As shown in Figure 3, these results are robust to the use of different measures of depth.  Table 2 which is modified to include two leads and four lags of the depth variables. The solid lines depict the cumulative effect and the broken lines the 95% confidence intervals. Results are robust to alternative numbers of lags and leads.
Second, to further test that coefficients on depth are not capturing a positive export trend between PTA country pairs, we test the significance of future levels of depth at different points prior to the entry into force of trade agreements. 29 A positive and significant coefficient could suggest that there is a positive export trend, or that trade flows increase due to expectations of future reductions in trade costs, or that countries sign agreements because of increases in trade flows. All these scenarios would invalidate a causal interpretation of our results. Specifically, we augment specifications 1 to 4 of Table 3 by the following variable where ℎ ( ) is the future level of depth between country i and j, and is an indicator variable equal to one in the year prior to an agreement entering into force and zero otherwise. If results in Table 3  variable two years prior to an agreement taking effect. Table 4 shows that trade flows are not statistically correlated with future levels of depth. Depth estimates obtained in Table 3 are unaffected by the inclusion of these additional variables. This evidence suggests a causal relationship between depth and trade. 30  2002-2014 2002-2014 2002-2014 2002-2014 2002-2014 2002-2014 2002-2014 2002-2014 Note: Robust standard errors, clustered at the country-pair level, are in parentheses.
*** p<0.01, ** p<0.05, * p<0.1 Another concern with our main specification is that the absence of intra-national trade flows limits the identification to the comparison between members of a PTA and country-pairs without PTAs. To address this concern, we follow Bergstrand et al. (2015) and construct intranational trade flows using GDP data from the Penn World Tables. 31 In this specification, the control group comprises country-pairs without trade agreements and countries' trade with themselves, neither of which sees any change in depth. Results in Table 5 suggest that the exclusion of internal flows plays an important role in explaining the relatively small effects of depth we found earlier, as already documented in the trade gravity literature for the PTA dummy (e.g. Dai et al., 2014;Larch et al., 2017). The coefficients of depth on trade are around three times larger than those presented in Table 3. These results are more in line with the literature in which, for instance, the coefficient for a common currency is 0.98 while we find 30 In an additional falsification test, we assign random levels of depth to trade agreements in our sample. Figure A2 of Appendix A shows the distribution of the coefficients of random depth variable obtained from 500 random draws of depth, obtained from a PPML model that includes a PTA dummy, is normally distributed. 31 The main advantage of constructing intra-national flows with GDP data is the extensive time and country coverage compared to gross output data (e.g. CEPII's TradeProd data are available until 2006). The drawback is that GDP is measured as value added which is an imperfect proxy of gross output. in column 6 that the coefficient for the deepest agreement in our sample (the European Union) is 0.97. Additionally, we find the expected negative and significant impact of tariffs on trade which is not captured in regressions with international flows only. Unfortunately, due to data limitations on product level output, we limit our analysis to international trade when we study the effects of trade diversion. To better understand the impact, and to quantify the effect, of additional provisions in trade agreements, we consider three agreements that are characterized by different levels of  Table A4 in the Appendix.
Based on the estimates in column 6 of Tables 3 and 5, we find that the Peru-Chile FTA increased members' bilateral trade between 10 and 30 percent. For the case of KORUS FTA, which includes additional provisions on state trading enterprises, public procurement, and provisions on intellectual property rights, we find a larger effect, ranging between 14 and 40 percent. Finally, we find that the inclusion of all depth core provisions and 25 other provisions spanning from taxation and money laundering to labor market regulation and visa and asylum, increased trade between 44 and 164 percent among EU countries. Table 6 presents the results on the effect of deep trade agreements on excluded countries. Note first that the depth of PTAs (depth LE) continues to have a consistently significant impact on trade between member countries. Even though the coefficients in Table   6 are slightly different from those in Table 4, the difference is not statistically significant. To ease interpretation all Others variables are standardized and the coefficients capture one standard deviation shocks. We find that the importer's average depth, when counting all the 52 areas, has a positive effect on bilateral trade. In column 2 we limit the analysis to core provisions and find that while the magnitude drops, the estimates increase in statistical significance. The results suggest that a standard deviation increase in partner's depth (depth core LE) increases trade by around 19 percent. These positive effects on third-countries could potentially explain the difference between trade creation estimates with and without internal flows. If deep trade agreements benefit all trade partners, then the effect of signing a deep trade agreement (or unilaterally reducing tariffs) would be absorbed by the country-year fixed effects when using international trade flows only.

b. Trade diversion
We find that the positive effect of deep agreements on third countries is driven by the inclusion of MFN provisions, while the inclusion of preferential provisions has a negative but insignificant impact (columns 3 and 6). The negative effect of preferential provisions becomes significant once we account for the presence of old PTA, agreements for which we do not have information on their content, in columns 9 and 12. Results in columns 7 to 12 reveal that old PTA is associated with a negative average effect on third-countries' trade. This finding suggests that early agreements were more trade diverting, which is consistent with the evidence presented in Section II that PTAs have become deeper over time. Both results for depth and older PTAs are robust to the inclusion of relative and bilateral tariff preferences, which are insignificant for different specifications.
A comparison of the effects of trade creation and trade diversion is helpful to put these results in perspective. As before, we focus on three trade agreements with different levels of depth: Peru-Chile, United States-Korea and the EU. Using estimates from column 9, we find that a medium depth agreement such as KORUS FTA increased exports from excluded countries to members by around 4 percent. We find large effects of the European Union (the deepest agreement in our data) for non-member countries. Estimates suggest that exports from non-EU countries would be around 30 percent lower in the absence of the agreement. Finally, we find shallow agreements between smaller countries such as the Peru-Chile agreement increased trade between members but had a negligible impact on non-members trade. In general, the positive impact on non-members' trade flows is driven by the inclusion of MFN provisions, while preferential provisions have a negative effect.  Table 7. We interact the RPM variable with the importer's average depth of trade agreements with the rest of the world and with the importer's maximum value of depth (i.e. its deepest trade agreement) in a given year. Therefore, we test if commitments, for instance, to improve customs efficiency or to reduce subsidies, soften the consequences of trade preferences for excluded countries.  2002-14 2002-14 2002-14 2002-14 2002-14 2002-14 2002-14 2002-14 Note: LE stands for legally enforceable. Robust standard errors, clustered at the country-pair level, are in parentheses. *** p<0.01, ** p<0.05, * p<0.1 We find that the effect of tariff preferences does depend on the depth of trade agreements concluded by an importing country. The interaction of the relative preference margin is significant with measures of both the average depth vis-à-vis the rest of the world (Others Depth LE or Others Core Depth LE) and of maximum level of commitments that importers undertake (Max Depth LE and Max Depth Core LE) (Table 7). Figure 4, based on results in column 6 of Table 7, shows that when maximum depth core is close to zero, a 1 percent increase in RPM decreases bilateral trade by 4 percent. This negative impact of relative preferences on trade is statistically significant for values of depth core lower than 0.3, while it is completely offset when more than 80 percent of depth core provisions are included. This suggests that tariff preferences have a discriminatory effect in countries that have "shallow" agreements, while the effect is reversed when a country undertakes deep commitments. The statistical insignificance of relative tariff preferences we find in some of our specifications may, therefore, be due to pooling across agreements with different levels of depth.

V. Concluding remarks
Most of the work on PTAs in the literature is based on the implicit assumption that trade agreements are about tariff liberalization. In this literature, the impact of preferential trade agreements is captured by the standard Vinerian analysis of trade creation and trade diversion.
Recent data on the content of trade agreements show, however, that PTAs are deepening, in the sense that they include an expanding set of provisions, often covering behind the border policy areas. The evidence presented in this paper confirms the view that Vinerian logic may provide an incomplete guide to the effects of deep agreements. Intuitively, the reason is that deep provisions do not necessarily act as preferential tariffs. In fact, we find that deep agreements create more trade than shallow agreements and that they can have a positive spillover effect on trade with outsiders when they are non-discriminatory in design or implementation. i.e. the legal commitments embedded in different policy areas covered by the agreement, are likely to matter for trade and beyond. Deep provisions on services and competition will influence the ability of countries to integrate in trade markets; investment rules will affect the ability to attract and retain foreign investment; and the protection granted to intellectual property rights will have an impact on the ability to innovate. As new data are collected on the detailed content of PTAs, an exciting research agenda lies ahead.   *** p<0.01, ** p<0.05, * p<0.1 Figure A2: Falsification test with random levels of depth (500 reps)

APPENDIX B -A MOTIVATING EXAMPLE OF DEEP AGREEMENTS
Trade creation and trade diversion in deep agreements can be illustrated using a standard diagram of the impact of PTAs. 33 The diagram assumes that there are three countries symmetric in size (Home, Partner and RoW), each country exports two goods and imports the other. The diagram displays the market for the good imported by Home, showing the export supply curves (XS) and the import demand curve MD ( Figure B1). All countries have a specific import tariff, t, on all imports. In addition, and for simplicity, assume that the frictional barriers created by non-tariff measures have an ad valorem equivalent tariff T. This implies that the gap between Home's domestic price P and the price of the two exporting countries is precisely given by the sum of the tariff and the frictional barrier, so that the export price is P-t-T.
In this framework, the trade impact of a deep relative to a shallow PTA can be easily assessed. While a shallow agreement would only eliminate the tariff between members, a deep agreement eliminates both the tariff and the frictional barriers, resulting in larger trade creation.
In the diagram, the shift to the right of the export supply curve is larger under a deep relative to a shallow agreement and Partner sees a sharper increase in its export price, leading to a larger increase in exports to Home. Now consider the impact of the agreement on non-members. The deep PTA still eliminates tariffs and other trade costs preferentially, but also reduces part of the frictional barriers on an MFN basis (T MFN , Figure B1). The ultimate impact of a deep PTA on RoW's price and export is ambiguous. The figure also shows that the larger is the proportion of T MFN in total trade costs, the greater is the positive impact of PTAs' on third countries' exports and the lower is the trade diverting effect of preferential tariffs. 33 See Chapter 5 in Baldwin and Wyplosz (2012).