Consequences of FTA Withdrawal: Evidence from EU Enlargement

Transcription

1 Consequences of FTA Withdrawal: Evidence from EU Enlargement PRELIMINARY VERSION Hinnerk Gnutzmann a a Institute for Macroeconomics, Leibniz University Hannover, Germany Abstract Political opposition to trade agreements has been on the rise, suggesting that some existing agreements may be severed in the future. But so far very little is known about the potential effects of withdrawal from trade agreements, because economic integration has been largely a one way street in recent decades. This paper studies a notable exception: Estonia withdrew from its FTA with Ukraine in the course of its EU accession. Through a triple difference in differences strategy, I estimate that Ukrainian imports to fell by more 20%. The result is robust to different specifications and obtains in an environment where preference margins were low. This suggests that loss of FTAs can be very costly for trade. Keywords: free trade agreement, withdrawal, European Union, Estonia, Ukraine JEL: F13, F14 address: gnutzmann@mak.uni-hannover.de (Hinnerk Gnutzmann) 1

2 1. Introduction Trade agreements are facing their first serious political backlash in decades. Some of the most ambitious free trade agreements (FTAs) of recent years the TPP and TTIP now look unlikely to ever be ratified and implemented by the United States. The United Kingdom has started the process of leaving the European Union and its customs union (CU). Meanwhile, the new US Administration put withdrawal from the North American Free Trade Agreement (NAFTA) on the table if renegotiations are not successful. After the rapid growth of preferential trade agreements in past decades, it now seems possible that some agreements may be scaled back or undone entirely. Understanding the effects of withdrawal from trade agreements has thus suddenly become an important research question for international economics, who have so far focused almost entirely on the formation rather than dissolution of trade agreements. This paper is the first to estimate the trade effects of withdrawal from a Free Trade Agreement (FTA), drawing upon a unique quasi experiment. With the Eastern Enlargement of the European Union in 2004, 1 the new member states also joined a customs union with centralized competence for trade agreements; in other words, member states replaced their earlier national trade agreements with those of the Union on the day of accession. This caused little upset to existing trade agreements by the new members, since their trade agreements had been negotiated with EU accession in mind. 2 However, the is a single exception: Estonia, one of the new EU members, had an FTA in force with Ukraine since Because the EU did not have an FTA with Ukraine at the time, Estonia had to withdraw from this agreement as part of its EU accession process. This withdrawal, which as I argue below was driven by considerations exogenous to Ukraine Estonian bilateral effects, makes the case study possible. To identify the causal impact of FTA withdrawal on trade, I use a triple difference in differences estimator which exploits cross country variation in trade agreement shocks upon EU accession. Robustly, I my point estimate of the FTA withdrawal effect is a loss of more than 20% in bilateral trade. This result is statistically significant at the 1% level throughout, and clearly economically significant. Indeed, Baier et al. (2016, table 1) find an average FTA trade creation effect of 26% 3. Thus, the finding is consistent with essentially all FTA trade gains becoming undone after FTA withdrawal and no trading capital being built up that outlasts the agreement (as in Head et al. (2010)). Second, the result appears not to be driven by tariff 1 See Gateva (2016, ch. 2) for a discussion of the A10 Enlargement Process. 2 For example, the A10 states had an FTA among each other; upon accession, these agreements were upgraded to the EU Common Market. The A10 also had an FTA with the European Free Trade Association (EFTA); after EU accession, the countries continued to enjoy an FTA with the EFTA countries since the EU also had an FTA with them in place. 3 Table 1, weighted average by inverse, quoted as 0.236; hence exp(0.236) 1 26% 2

3 preferences. Prior to EU accession, Estonia was close to pursuing unilateral free trade 4 ; however, the Estonia-Ukraine FTA did grant tariff-free access for all tariff lines. This suggests that particularly comprehensive agreements may have extra trade-promoting effects. Third, while there is clear evidence for a trade increase prior to FTA withdrawal (anticipation effect), there is no evidence of a delay in trade reduction afterwards: a Chow tests fails to find evidence for a gradual fall in trade. These findings are potentially important in the context of future UK trade policy after Brexit. One possible policy discussed in a Parliamentary Select Committee on 29 November in unilateral tariff elimination. In this context, it has been argued that the country does not need free trade agreements (Minford). The evidence from Estonia suggests that, even when MFN tariffs are close to zero, a comprehensive FTA can give a substantial boost to trade. Second, some researchers have referred to reduced form estimates of FTA effects to gauge potential Brexit trade consequences 6. This implicitly assumes symmetry between accession and withdrawal effects; our findings suggest this assumption may be tenable. This paper proceeds by describing the Estonia Ukraine FTA in section 2; the estimation strategy is outlined in the following section, while 4 turns to results. Finally, 5 concludes. 2. The Estonia Ukraine Free Trade Agreement Estonia is a small open economy bordered by the Baltic Sea, Russia and Latvia. After the fall of the Soviet Union, Estonia transitioned rapidly to a market economy and is considered one of the most successful post-socialist economies (Norkus, 2007). Prior to joining the European Union in 2004, Estonia also had a very liberal trade policy: according to the World Trade Organization, its average MFN tariff was only 1.68% in 2002 and for 93% of tariff lines, Estonia granted tariff free access on an MFN basis. In other words, Estonian tariffs were unusually low by international standards. Additionally, Estonia had free trade agreements in place with the European Union, the EFTA countries and Ukraine. These agreements were unusually comprehensive, since Estonia granted tariff free access on all goods to each FTA partner. Upon EU accession, Estonian trade policy underwent a discontinuity. Estonia, together with seven other Eastern European countries (collectively known as A8 ), 7 joined the European Union on 1st May 2004, which before consisted of 15 countries (the EU15 ). This meant that its trade policy changed overnight: trade with other EU countries was now governed by the European Single Market. With all non EU countries, Estonia now implements 4 Mean MFN tariffs were very low, as discussed above, below 2%. See also Feldmann and Sally (2002) for an in-depth analysis. 5 See international-trade-committee/news-parliament-2015/uk-trade-options-beyond-2019-first-evidence-16-17/ for further information and a video recording 6 For example, Dhingra et al. (2017, p. 24) refers to the FTA accession effect study of Baier et al. (2008) 7 Besides the A8, Cyprus and Malta joined the EU on the same day. 3

4 the common trade policy of the European Union. As a consequence, the Estonian MFN tariff more than doubled (reaching 4.18% by 2005). Moreover, Estonia withdrew from its existing trade agreements and switched to applying the agreements of the Union. Figure 1(a) shows the resulting changes to Estonia s preferential trade regime. In particular, note that Ukraine lost its FTA status, because there was no EU Ukraine FTA in place at the time. 8. Instead, Estonia now applied some preferential tariff reductions for Ukrainian imports based on the EU s Generalized Scheme of Preferences. Russia, earlier treated as a third country on an MFN basis, likewise now became eligible for GSP preferences in Estonia. Both Estonia and the EU had longstanding FTAs in place with the European Free Trade Association, comprising Norway, Iceland, Switzerland and Liechtenstein. Hence, the status of the EFTA countries did not change after Estonian EU accession. The Estonia Ukraine dyad was the only one to suffer a downgrade of its trade relations. Overall, the EU Enlargement process shows careful sequencing to avoid disruption of existing trade relations. All accession countries had FTAs with the EU15 and EFTA countries already in place; these trade links either were upgraded to the Single Market or stayed in place as before. For Estonia, forgoing the Ukraine FTA was an acceptable loss in economic terms: trade amounted to 87m per year on average from , or for 1.7% of total imports. It was also unavoidable: for the Union to allow an exception to the common trade policy for this FTA would have been legally and administratively challenging, 9 and an EU Ukraine FTA was not on the political agenda at the time. Because of these factors, one can think of this case of FTA withdrawal as a quasi-experiment, which occurred for reasons entirely unrelated to any bilateral Estonian/Ukrainian shocks. At first pass, the trade impact of FTA withdrawal seems very large. Before EU accession, Estonia traded with Russia on an MFN basis and an FTA was in place in Ukraine; afterwards, both partners had access on GSP terms only. This suggests a difference in differences estimate, which exploits variation in relative import penetration by Ukraine and Russia prior to and after EU accession. From the gravity model of trade flows, we know that the ratio of imports from any two countries can be written as the product of relative country effects, reflecting relative competitiveness and growth, and the ratio of relative dyad effects, capturing changes in trade agreements. Assuming for the moment that relative country effects of Ukraine and Russia did not change over the period, this suggests the following calculation: on average, Ukraine imports were 16% the level of Russian imports in Estonia while the agreement was in force. After Estonian EU accession, the relative important penetration of Ukraine (as compared to Russia) fell to 10%, or a drop in 38%. This is is clearly economically highly 8 Interestingly, the new Deep and Comprehensive FTA between the EU and Ukraine, in effect since 1st January 2015, is less comprehensive in terms of tariff elimination than the earlier Estonia Ukraine FTA. According to the WTO, various lines are exempted! 9 Contrary to theory, there are instances where members of a Customs Union can still have different FTAs. For example, Turkey and EU are in a Customs Union but have some non-overlapping FTAs (World Bank, 2014) 4

5 FIGURE 1 Trade Policy and Trends in the Baltics (a) Structure of Preferential Trade Regimes Ukraine Russia EFTA EU15 & A8 Ukraine Russia EFTA Estonia Latvia EU15 & A GSP FTA EU (b) Trends in Import Penetration by Ukraine, Relative to Russia Ratio: Imports from Ukraine / Russia Estonia Latvia Notes: The vertical dashed line indicates the timing of the EU Eastern Enlargement in May 2004, the accession of inter alia 8 Eastern European countries ( A8 ); the A8 include Estonia and Latvia. EU15 are the 15 member states of the EU prior to the enlargement; EFTA is the European Free Trade Association. Trade agreement data based on the EIA September 2015 database (Baier et al., 2014), updated by the author after Besides GSP, the EU and Russia have Partnership and Cooperation Agreement in force since Import penetration figures are the author s calculation based on the Eurostat COMEXT database. 5

6 significant. The calculation is visualized in the top panel of figure 1(b). This initial estimate leaves open a number of questions. First, the assumption of constant relative competitiveness between the two countries is naturally problematic. Fortunately, the structure of EU enlargement allows for this hypothesis to be tested: in Latvia, there was no change in relative preferences (Russia and Ukraine under MFN before EU, GSP after). As the bottom panel of figure 1(b) shows, relative imports hardly changed after EU accession (from 15% to 14%), suggesting that the common trends assumption is reasonable for these two import partners. Second, it is apparent that there is large increase in Ukrainian imports in 2003, followed by a fall in 2004; since the timing of EU accession was known well in advance, 10 this suggests that trade could have been shifted over time in anticipation of the FTA being withdrawn. Unless this effect is controlled for, estimates are likely to be biased. Lastly, we have not addressed issues of robustness, statistical significance and efficient estimation. In the remainder of the paper, we will develop a formal econometric model to address these issues. 3. Data & Specification Data: Bilateral trade flows are sourced from the Eurostat Comext database, covering import flows reported by the Baltic countries of Estonia and Latvia, as well as the aggregate of the EU15 countries. For each reporter, I collect total imports by all partner countries. The frequency is monthly, which is important because the EU enlargement, and associated trade policy treatments, took place on 1st May The sample period runs from January 1999, the first year for which data are available for the Baltic countries, through to December Our measure of bilateral trade relations is based on the September 2015 version of the Economic Integration Agreements database of Baier et al. (2014), updated through to 2014 by the authors. Model: For estimation, I rely on the standard gravity model of trade flows. This model, so widely and successfully used that it can be considered the workhorse (Head and Mayer, 2013) of the empirical trade literature, decomposes the bilateral trade flow X ijt from country i to j into the product of country specific effects relating to the importer and exporter specifically, as well as effects specific to the individual importer-exporter country dyad. Moreover, global time-varying factors (denoted G t ) are allowed for. This yields the general gravity equation: X ijt = G t M im it M ex jt T ij D ijt η ijt (1) Here, M ex it and M im jt contain attributes that vary over time by the respective country; hence, these are commonly referred to as monadic attributes. One possible source of monadic effects is country GDP, reflecting the idea that bilateral trade depends on the combined eco- 10 INSERT DATE WHEN TREATY SIGNED. 6

7 nomic mass of the two countries. However, as the seminal contribution of Anderson and Van Wincoop (2003) points out, monadic effects cannot be reduced to GDP. It is also required to take into account the multilateral resistance, i.e. the trade barriers a country faces with all other countries. 11 To address this problem, I will use an estimator that effectively allows for country month fixed effects. At the dyad level, both time invariant factors (e.g. distance between countries; captured by T ij ) and time varying effects, such as trade agreement status (D ijt ), can play a role. The main interest of this paper lies with estimation of the FTA withdrawal effect. Accordingly, the dyad time effects are parameterized as: ln(d ijt ) = θ ij fta ijt + ω ij ftawithdrawn ijt + γ i gsp ijt (2) where fta is a dummy indicating if the partner countries are currently in an FTA, ftawithdrawn indicates that the same countries were previously in an FTA but are currently not and gsp takes value 1 if the importer grants a preference under the EU GSP program to the given exporter. Importantly, I will allow the GSP coefficient to vary by beneficiary country, since the extent of GSP preferences is tailored to the recipient by the European Commission. In principle, the FTA effect could also vary by dyad, although this will turn out not to matter in my estimation. One very good way to estimate the gravity model is through a triple difference-in-difference estimator (Head and Mayer, 2013, section 3.4). The gravity equation tells us that the relative import penetration of exporters j and k in country i can be written as the ratio of monadic exporter effects (think relative competitiveness of the two exporting countries), multiplied with the ratios of dyadic effects and the error terms: R i[jk]t = X ijt = Mex jt T ij D ijt η ijt X ikt M ex kt T (3) ik D ikt η ikt In other words, this transformation eliminates the global time effect and the monadic importer effect; according to the gravity model, they should affect imports from j and k equally, so they do not change relative import penetration. Variation in this ratio over time is driven by relative dyadic effects, which is what we want capture, and relative competitiveness, which we can t observe and need to control for. To solve this problem, note as did Romalis (2007) that relative import penetration in a different importing country l will be driven by the same unobserved relative monadic effects: R l[jk]t = X ljt = Mex jt T lj D ljt η ljt X lkt M ex kt T (4) lk D lkt η lkt 11 Consider the example given by Adam et al. (2007): There is a reduction in bilateral trade costs between the UK and France, but no change in France UK trade costs. Then, the multilateral resistance of France falls, causing some trade diversion from France Italy to France UK. 7

8 Hence, if we have such a tetrad of four countries, we can use the difference between the difference-in-differences estimates for each importing country to filter out the unobservable monadic exporter factors. This yields our main estimating equation: r [il][jk]t =ln ( Ri[jk]t R l[jk]t ) = {ln(d ijt ) ln(d ikt )} {ln(d }{{} ljt ) ln(d lkt )} +v }{{} [il][jk]t (5) DiD country i DiD country l v [il][jk]t =(η ijt η ikt ) (η ljt η lkt ) (6) Reference Countries: To complete the tetrad, one needs to choose a reference importer to compare with Estonia, and a reference exporter vis-a-vis Ukraine. In both cases, there are obvious choices: Latvia, the southern neighbor of Estonia, shares a rich common history and has simultaneously pursued European integration with Estonia. Regarding trade agreements, the only difference between the two is the Estonia Ukraine FTA. Moreover, the business cycles of these two countries move in sync: GDP growth rates are correlated at 94% over the sample period, as table A1 shows. Thus, Latvia seems to be as good as reference country as one could hope to find. For Ukraine, its eastern neighbor Russia is likewise an obvious choice. Although relations have been of late problematic, GDP growth rates between the two were also very highly correlated throughout the sample period (90%). The two countries differ in their trade relations with third countries (for example, Russia has union state with Belarus and formed the Eurasian Customs Union from 2012 onwards, i.e. towards the end of the sample period), which could affect the relative multilateral resistance of the two. While the estimation controls for this, it is reassuring that world export growth of the two countries remained highly correlated throughout the sample period (91%). Hence, the baseline estimation involves the Estonia-Latvia-Ukraine-Russia tetrad. For robustness, I consider the EFTA countries as an alternative reference exporter. The gravity model tells us that, in principle, it should not matter which reference countries we use to estimate the triple D-iD equation 5, subject to some qualifications. As long as the compound error term is not correlated with changes in trade agreements, the estimates should be unbiased. Choosing similar reference countries does, however, help efficiency because it reduces the variance of the compound error. 12 Clearly, the EFTA countries are not as closely correlated with Ukraine (e.g. GDP growth correlation merely 70%), so some loss of efficiency can be expected. Still, it would be reassuring if results did not depend on the details of country choice, so I also estimate the model for the Estonia-Latvia-Ukraine-EFTA tetrad. Estimation by Tetrad: I first estimate the model separately for each tetrad. After substituting 12 For example, given the high correlation of Ukrainian and Russian export shocks, the difference η i[ua]t η i[ru]t may be very small 8

9 the trade agreement dummies into equation 2, this yields for both dyads: r [EE,LV ][UA,k]t = β + ω EE,UA ftawithdrawn [EE,UA],t + v [EE,LV ][UA,k]t (7) where k {RU,EFTA}. The intercept β contains the relative time invariant dyad fixed effects T ij and the trade effect θ EE,UA of the Estonia Ukraine FTA. Since data for the pre FTA period are not available, this effect cannot be identified. Note that any trade effects of the GSP program are controlled for this specification: each importer granted and revoked GSP preferences to the exporters at the same time, so the tetrad difference is nil in each time period. To this model, I add various time dummy dummies for the period around the EU accession to capture possible anticipation effects on trade. The coefficient on the time dummy, which marks the Eastern EU enlargement, captures the pure effect of FTA withdrawal. Panel Data: Last but not least, I stack up my tetrads into a panel data set. The goal here is to formally test if my estimates of the FTA withdrawal effect depend on the choice of reference country. If they do not, then it is possible to use data from both tetrads jointly, while allowing for tetrad fixed effects of course, to get a more precise coefficient estimate. To make sure the standard errors are reliable, and do not give me a false sense of precision, I use clustering as recommended by Head et al. (2010). Given that we have a single importer exporter pair, this means clustering by month. 4. Results FTA withdrawal leads to a very noticeable fall in trade. Table 1 presents the baseline regression results. Models (1 3) have Russia as the reference exporter. Model (1) reproduces the calculation of section 2; in (2), I add dummies for the years 2003 and 2004 to control for anticipation effects; this gives a point estimate of a e = 24.3% trade reduction due to FTA withdrawal, which is economically important and statistically significant at the 1% level. Adding further dummies for each month from January 2003 to December 2004 effectively causing observations from this period to drop out from OLS estimation has hardly any effect on the coefficient estimate, as column (3) shows. This specification has a minimally higher R 2 than model (2) and a Wald test gives no evidence of improved fit (F 166,22 = ). Annual dummies are enough to capture the anticipation effect for the Russian tetrad. In models (4 6), I repeat the calculations for the case where the EFTA countries are the reference exporters. Reassuringly, the point estimates are very close. As expected, the estimates are slightly less precise when EFTA is used as the reference exporter (standard errors about 1/3 higher), suggesting that Russia has better correlation of monadic shocks vis-a-vis Ukraine. Again, monthly dummies for the transition period do not improve fit compared to the annual dummies. Throughout, the residuals appear to be normally distributed according to the 9

10 TABLE 1 Tetradic Regression Results (1) (2) (3) (4) (5) (6) ftawithdrawn (0.113) (0.089) (0.095) (0.137) (0.116) (0.124) Reference Exporter RU RU RU EFTA EFTA EFTA Controls Years 03, 04 X X X X Months in 03, 04 X X Durbin-Watson Breusch-Pagan Jarque-Bera Observations R Notes: Newey-West HAC standard errors in parentheses. p<0.1; p<0.05; p<0.01. In the year controls specification, indicators are added for the year being 2003 or In the month controls specifications, additional indicators are added for each month from February to December in the years 2003 and For diagnostic tests, p-values are always quoted. Own calculation based on dataset described in the text. Jarque-Bera statistic. Thus, I drop monthly dummies from now on. There is clear evidence for heteroscadasticity in the EFTA model (4), which appears to be driven by more volatile EFTA trade shares in both Estonia and Latvia towards the end of the sample period; also, there is weak evidence for heteroscedasticity in model (2). There is strong evidence for serial correlation across specifications. Visual inspection clearly shows that runs of positive/negative residuals. Figure A1(a) shows the residuals for model (2). There were several months with positive residuals e.g. in 2002 and 2008, and runs with negative residuals in e.g and Likewise, there were runs of residuals with the same sign for the EFTA tetrad (model 4) 13, although these tend to occur at different dates. The visual impression of autocorrelation is clearly confirmed by the highly significant Durbin-Watson statistic, which overwhelmingly rejects absence of autocorrelation for all specifications. Since the data are monthly, one possible source of autocorrelation would be seasonality. However, the correlogram in figure A1(b) shows that autocorrelation is significant only up to 6 months. At month 12, the hypothesis of no autocorrelation cannot be rejected at the 5% level, suggesting that the tetrad design successfully solves any seasonality issues. 14. These findings suggest that serial correlation is driven by short-run adjustment dynamics rather than misspecification of the model. Nevertheless, given the undeniable pres- 13 See online appendix. 14 There is evidence for seasonality in individual trade flows, in particular involving Russia. (ref) 10

11 ence of autocorrelation, OLS may not be the efficient estimator. Thus, I experiment with the Cochrane-Orcutt estimator, which adjusts the model for AR(1) errors, and re-estimate models (2) and (4). The results are models (7 8) in table 2: as promised, standard errors are smaller although the gain in efficiency is minuscule. Moreover, the point estimates of the FTA withdrawal effect are for practical purposes unchanged. Hence I continue estimation with OLS. TABLE 2 Further Regression Results Estimator Cochrane-Orcutt Panel, Fixed Effects (7) (8) (9) (10) (11) (12) ftawithdrawn (0.086) (0.114) (0.100) (0.100) (0.099) (0.108) Interactions ftawithdrawn: Ref. Exporter = EFTA (0.136) Year (0.121) Reference Exporter RU EFTA Both Both Both Both Controls: Years 03, 04 X X X X X X Observations R Notes: In panel estimation, Driscoll-Kraay SE are used, which are robust to serial and cross-sectional dependence. p<0.1; p<0.05; p<0.01 For models (11-12), the sample period is limited from January 1999 to November 2014 only (considered the onset of the Ukrainian crisis ). Panel estimation shows that estimates do not depend on the specific choice of reference exporter; there is very weak evidence of a delayed impact of FTA withdrawal. Models (9 12) in table 2 are estimated using the fixed effects panel model, together with robust standard errors to account for serial correlation and cross-sectional clustering. As model (10)shows, the interaction between FTA withdrawal and the choice of reference exporter is entirely insignificant practically small and imprecisely estimated. This is as it should be, given the identification strategy. Hence, I estimate the FTA withdrawal effect jointly from both tetrads in column (9), which gives a point estimate of e = 22.7%. Lastly, for robustness, I consider a subsample that ends in November 2014, considered the starting point of the Ukrainian crisis. In theory, the Ukrainian crisis is a monadic shock that our identification strategy controls for. In practice, this seems to be the case as the estimate is hardly sensitive to the change in sample period (model 11). Lastly, model (12) adds an interaction for FTA withdrawal with the period until The idea is to split the post FTA period into two halves and test the structural stability of coefficient estimates. They seem to rather stable (interaction term has 11

12 a p-value of 0.25), so any evidence for a delayed impact of FTA withdrawal is weak at best. 5. Conclusion This paper studied the trade effects of FTA withdrawal by drawing on a unique natural experiment. When Estonia joined the European Union, it needed to withdraw from an FTA with Ukraine as part of the acquis of joining the Union. Since Estonia-Ukraine trade was relatively small, compared to the potential gains from EU membership, this was a trade-off worth accepting for Estonia. Hence, the withdrawal is plausibly exogenous and not related to any Estonia-Ukraine specific shocks, such as political turmoil. I estimated the standard gravity model using the tetrad model, with Latvia used as a counterfactual for Estonia. For Ukraine, I consider both Russia and the EFTA bloc as possible counterfactuals; the choice of counterfactuals can be well argued based on the economic and historical fundamentals of the region. Moreover, statistical evidence from e.g. correlations in output growth, support the setting. Finally, in a panel model I cannot reject that results are invariant to the counterfactual assumed. The central finding is that FTA withdrawal reduced bilateral trade by more than 20% (point estimate). This result is significant statistically at the 10% level; clearly, it is also economically significant. There is no evidence of a delayed effect: the trade reduction follows immediately after the withdrawal. However, before the agreement comes out of force, there are substantial anticipation effects. References Adam, C., D. Cobham et al., Modelling multilateral trade resistance in a gravity model with exchange rate regimes, in Centre for dynamic macroeconomic analysis conference papers (2007). Anderson, J. E. and E. Van Wincoop, Gravity with gravitas: a solution to the border puzzle, American Economic Review 93 (2003), Baier, S. L., J. H. Bergstrand, P. Egger and P. A. McLaughlin, Do economic integration agreements actually work? Issues in understanding the causes and consequences of the growth of regionalism, The World Economy 31 (2008), Baier, S. L., J. H. Bergstrand and M. Feng, Economic integration agreements and the margins of international trade, Journal of International Economics 93 (2014), Baier, S. L., Y. Yotov and T. Zylkin, On the widely differing effects of free trade agreements: Lessons from twenty years of trade integration, (2016). Dhingra, S., G. I. Ottaviano, T. Sampson and J. V. Reenen, The consequences of Brexit for UK trade and living standards, Technical Report,

13 Feldmann, M. and R. Sally, From the Soviet Union to the European Union: Estonian Trade Policy, , The World Economy 25 (2002), Gateva, E., European Union Enlargement Conditionality, Palgrave Studies in European Union Politics (Palgrave Macmillan UK, 2016). Head, K. and T. Mayer, Gravity equations: Workhorse, toolkit, and cookbook, (2013). Head, K., T. Mayer and J. Ries, The erosion of colonial trade linkages after independence, Journal of international Economics 81 (2010), Norkus, Z., Why did Estonia perform best? The north south gap in the post-socialist economic transition of the Baltic states, Journal of Baltic Studies 38 (2007), Romalis, J., NAFTA s and CUSFTA s Impact on International Trade, The Review of Economics and Statistics 89 (2007), World Bank, Evaluation of the EU-Turkey Customs Union, Report No TR, World Bank,

14 Appendix FIGURE A1 Regression Diagnostics, Model (2) (a) Regression residuals (b) Residual Autocorrelation Function Lag (Months) Notes: Vertical dashed line in (a) indicates EU accession. Own calculations 14

15 TABLE A1 Pairwise Correlations of Key Aggreggates EST LVA RUS UKR EST LVA RUS UKR 1.00 (a) Change in Log GDP EST LVA RUS UKR EST LVA RUS UKR 1.00 (c) Change in log Total Imports EST LVA RUS UKR EST LVA RUS UKR 1.00 (b) Change in log GDP per Capita EST LVA RUS UKR EST LVA RUS UKR 1.00 (d) Change in log Total Exports Source: GDP data are sourced from the IMF World Economic Outlook, October 2016 release. The series are NGDPD (GDP) and NGDPDPC (GDP per capita). World Import data are sourced from COMTRADE. All tables are based on annual data for the period