Measuring EU Trade Integration within the Gravity Framework

Measuring EU Trade Integration within the Gravity Framework Andrea Molinari INTRODUCTION... 2 CHAPTER I. ECONOMIC HISTORY AND TRADE STYLISED FACTS... 4 CHAPTER II. TRADE INTEGRATION AND GRAVITY MODELS: SURVEY AND SOME THEORETICAL ASPECTS... 5 II.1. GRAVITY MODELS... 5 II.1.1. What do they measure?... 5 II.1.2. On the Theoretical Foundations of Gravity Models... 7 CHAPTER III. THE MODEL... 8 III.2.2. Data and Proxies... 9 III.2.2.1. The Baseline Gravity Variables...10 III.2.2.2. The Integration (institutional) Effects...11 III.2.2.3. Unobservable Time Effects and Country Characteristics...12 III.4. THE RESULTS... 15 III.4.1. SURE Estimation... 16 III.4.1.1. Baseline gravity variables...16 III.4.1.2. The integration effects...20 III.5. MAIN FINDINGS... 23 CHAPTER IV. FURTHER RESEARCH EXTENSIONS AND CONCLUSIONS... 24 APPENDIX A. WELFARE EFFECTS OF REGIONAL TRADE AGREEMENTS... 25 REFERENCES... 27

Introduction This thesis focuses on trade integration among the countries that belong to a regional trade agreement. Increase in intra-bloc trade is one of the most crucial ingredients for prosperous economic integration process such as the one implemented by the European Union (EU). We are especially interested in applying this issue to the EU because of the high economic integration reached by its members since the post-war period. Hence, our aim is to assess the contribution of the EU institutional framework and its evolution to the intra-eu trade deepening process. There is no doubt that intra-eu trade has increased after the creation of the European Economic Community (ECC) in the late 1950s. There are many potential explanations for this growth. On the one hand, EU members have experienced, as have most OECD countries, an increase in their economic size and wealth, which may well have stimulated their exports and imports to each other. In addition, doing business has become easier through the past decades due to the growing globalisation process. Proximity among EU members also operates as an important positive determinant of the countries trade flows, given the lower transport and transaction costs that physical closeness entails. Another determinant of the observed rise in trade flows among EU countries may have been the institutional effect entailed on the creation of this regional trade agreement. The well-known gravity framework provides us with the necessary tool to test whether the institutional creation of the EEC in itself has had a significant impact on intra-eu trade. Gravity models have been widely used in the literature to find the determinants of bilateral trade flows. Their baseline specification includes the size of the trading countries, their common characteristics, and the distance between them. By adding a dummy variable for joint membership to a certain regional trade agreement, it is possible to account for EU s institutional incidence in bilateral trade flows. Given that we try to reflect the evolution through time of the integration process, it is necessary to control for unobservable time effects in the baseline gravity equation. These effects capture further trade determinants that are difficult to measure, such as the globalisation process, and therefore cannot be included into the equation explicitly. Along similar lines, it is also important to capture countries specific characteristics. Thus, with a gravity model that controls for the baseline trade determinants (called natural by Frankel, Stein and Wei, 1993) and for the unobservable time and 2

individual effects we are able to isolate the institutional effects of the EU. This then allows us to assess whether they were important in determining intra-eu trade flows. Although the gravity framework s lack of theoretical foundations provokes criticism, it also provides some flexibility for application to various samples of countries and years. However, the gravity framework has two intrinsic problems. First, it includes time-invariant regressors, such as distance and common characteristics between countries, which can make estimating their effects on trade more difficult. Moreover, explaining trade with the partners economic sizes generates endogeneity problems. This means that a country s exports both depend on and determine its economic size. We estimate our gravity model by three methods. In order to differentiate EU members from a broader sample of countries with similar characteristics, and to analyse the evolution through time of the institutional EU effect, we estimate our gravity specification for a sample of developed OECD countries over the period 1960-1997. To account for the time correlation between the cross-section gravity equations, we first estimate a Seemingly Unrelated Regression (SURE) model. We then consider a panel data General Least Squares (GLS) method to control for unobservable individual effects in the sample. Although this method might provide inconsistent estimators, which are probably due to the mentioned endogeneity problem, GLS is preferable to the within-group, since this sweeps out the effects of time-invariant regressors. Our third method is an Instrumental Variables (IV) GLS method suggested by Hausman and Taylor (1981), a useful way of accounting for both the unobservable individual and time effects and the endogeneity issues. The IV-GLS method uses internal instruments to control for the mass endogeneity and allows for the inclusion of timeinvariant regressors (such as geographical distance and common characteristics). Our main findings indicate that the EU effect decreases when controlling for unobservable individual effects and mass endogeneity. The EU estimate is also sensitive to globalisation and to the three external enlargement effects. In addition, we find a positive long-run external effect for the first and third enlargements on bilateral exports and a negative medium-run external effect for the first enlargement. This is mainly due to a learning process entailed in the accession of new members, UK s joining, and the first oil shock. Finally, our results show a positive globalisation effect on bilateral exports. 3

To conclude, we suggest some ideas to include economic distance measures and to disaggregate the model into economic sectors. Finally, we begin analysing possible ways to solve the lack of labour market mobility among EU countries by exploring the wages-trade link. Our work is organised as follows: Chapter I summarises the most relevant post-war developments for the OECD countries, focussing mainly on the European integration process and on stylised facts for intra-eu trade patterns. Chapter II introduces the gravity model, discussing some of the theoretical aspects involved in adopting this framework, and surveying the main studies that measure EU trade integration. Chapter III derives a simple gravity model from an imperfect competition setup to then specify and estimate our model to measure the evolution of trade integration among the EU members by three different methods (SURE, GLS and IV-GLS). Finally, Chapter IV poses further issues and ideas to be developed in relation to finding an economic distance measure, sectoral gravity equations and labour market integration. Chapter I. Economic History and Trade Stylised Facts This chapter presents a brief historical overview and some trade stylised facts in order to define the general settings of the period analysed in our model and to understand the main results of our estimations. Section I.1. focuses on some general aspects of the economic history of industrial countries. 1 The next section describes the main postwar multilateral and regional agreements reached by European countries. The last part of the chapter shows stylised facts of the intra-european trade as well as trade patterns between Europe and the rest of the world. 2 1 Given that most OECD members are industrial (or developed) countries, these terms will be used interchangeably. 2 This comprises third countries that do not form part of the respective agreement. 4

Chapter II. Trade Integration and Gravity models: Survey and Some Theoretical Aspects We will use the gravity model as a tool to measure trade integration among EU countries. The first part of this chapter explains the gravity framework and highlights some theoretical issues. The second part of the chapter surveys the main studies that measure trade integration. II.1. Gravity models This section introduces the main tool we are going to use for measuring EU trade integration, the gravity model. We first explain the idea behind it and then describe some of the main theoretical issues posed in the literature. II.1.1. What do they measure? Gravity models derive their inspiration from Newton s law of gravity, which recognises that all material particles, and the bodies that are composed of them, have a property called gravitational mass. This property causes any two particles to exert attractive forces upon each other that are directly proportional to the product of the masses and inversely proportional to the square of the distance between the particles. Gravity models are used to explain bilateral trade links between countries as directly proportional to their size and inversely related to the distance between them. Most models also include some common idiosyncratic characteristics of these countries, such as the sharing of a common language or membership in certain preferential trade agreements. They have proved to be empirically robust and consistent with the observed data and offer a systematic framework for measuring bilateral trade patterns around the world. As Eichengreen and Irwin (1998) assert, Few aggregate economic relationships are as robust 3. Baldwin (1994) makes a useful and intuitive analogy of an individual family s pattern of purchases to explain the idea behind the gravity model: A family lives near two shopping areas. Factors influencing how much the family buys at each shopping area may be divided into those that concern the family s characteristics and those that relate to the particular shopping area s traits. For instance, the richer the family becomes per capita, the more they will tend to spend on goods from both shopping 3 Eichengreen and Irwin (1998), page 34. 5

areas. Similarly, holding constant the per capita income of the family but increasing the family s total income, and thereby the size of the family, would increase the amount bought at both sites. The division of purchases between the two shopping areas would depend primarily on the various characteristics of the shopping areas themselves. It is likely that the family would buy relatively more from the area that offered the wider selection of goods. Also, other things being equal, the family will tend to do more of their shopping at the nearby shopping area. 4 In an international trade setup, the richer and bigger the country, the higher its purchases of foreign goods; i.e. a country s imports increase with its per capita and total income. The volume and the variety of goods produced and available resources will also be greater as a country grows in size and becomes richer. In other words, an exporting country s per capita and total GDP (called mass in the gravity framework) should be positively correlated with its exports. Finally, the greater the goods transportation cost between two countries, the smaller the quantity of trade; i.e. distance (or any other determinant of transaction costs) dampens trade. This positive correlation between exports and GDP is possible as long as greater production is evenly distributed across all goods and services. It is possible that a country s economy responds to an increase in GDP only by expanding its non-traded sector. In that case, there would not be such a correlation between trade and size. Moreover, an argument can be made for bigger countries to be self-sufficient. During the 1990s, gravity models became widely used as a tool to explain bilateral trade flows among countries or regions. Some of their applications incorporate trade blocs dummies to test how natural those blocs are (Frankel, Stein and Wei, 1993 and 1998). A second group of studies includes both internal and external trade and distance proxies (Wei, 1996; Helliwell, 1996, 1997 and 1998; Nitsch, 1999) to measure the width of the borders. The gravity setup has also been used to measure the potential trade of developing countries, such as Eastern Europe EU accession (Wang and Winters, 1991 and 1994; Hamilton and Winters, 1992; Baldwin, 1994), or within the Southern Africa region (Foroutan and Pritchett, 1993; Cassim and Hartzenburg, 2000). Due to our main interest in EU trade integration, our survey focuses on the first two lines of study as well as other studies that measure EU trade integration outside the gravity framework. 4 Baldwin (1994), pages 82 and 83. 6

II.1.2. On the Theoretical Foundations of Gravity Models The origins of the gravity model to explore the determinants of bilateral trade flows go back to Linnemann (1966), who proposed to consider the importer s demand, the exporter s supply and the trade costs between them. Since then, the theoretical foundations of gravity models have been questioned. Due to their intuitive specification, gravity models have always been considered to work fairly well in empirical grounds. However, most concerns centre on whether it is possible to derive this model from various theoretical frameworks that adopt different, and sometimes contradictory, assumptions. The critics point out that the gravity framework is compatible both with the perfect competition models, such as Heckscher-Ohlin, and with trade theories that assume imperfect competition. More specifically, Wang and Winters (1991) point out that a simple Cobb-Douglas expenditure system (such as Anderson s, 1979), is not appropriate to derive a gravity specification. As Anderson shows, introducing (...) stochastic errors and/or multiple commodities, (...) the log-linear relationship between aggregates is difficult to support 5. Moving to more comprehensive functional forms, Bergstrand (1989) obtains a gravity equation that explains bilateral trade flows from a general equilibrium model with two differentiated products and two factors. The representative consumer is assumed to maximise a nested Cobb-Douglas-CES-Stone-Geary utility function subject to a budget constraint, whereas the firms produce in a Chamberlinian monopolistic competition setup. Bergstrand shows that, under this framework, demand depends upon relative prices and domestic income. Hence, the gravity equation fits in with both the Heckscher-Ohlin model of inter-industry trade and the Helpman-Krugman- Markusen intra-industry trade models. Deardoff (1995) shows that, starting from a Heckscher-Ohlin model, the gravity model can be derived assuming either frictionless trade or imperfect competition. In the first case, preferences need to be identical and homothetic or demands have to be uncorrelated with supplies. In the context of countries producing differentiated goods, preferences can be either Constant Elasticity of Substitution (CES) or the special case Cobb-Douglas. However, this argument has been discussed by later work. For example, Helliwell (1998) finds that a model of comparative advantage limited by trade barriers would 7

not seem to predict any influence of average incomes on the size of border effects, 6 which he estimates using a gravity model. Moreover, some authors consider this lack of theoretical incompatibility as an attractive feature that gives flexibility to explain (...) bilateral trade flows across a wide variety of countries and periods 7. In sum, some studies have shown that the gravity model can be derived from two opposite theoretical models. However, this only indicates that gravity models cannot be used to test rival trade theories, and hence we believe that the theoretical flexibility of gravity models is an advantage rather than an obstacle for explaining bilateral trade flows. Chapter III. The Model One of the main objectives of an economic union is to increase trade among its members. The trade intensity between countries depends mainly on the explicit and implicit barriers that each imposes on its partners. These barriers generally take the form of transport costs, tariffs, and non-tariff restrictions. Declining costs of transport and communication reduce the economic distance between communities, regardless of which country the communities belong to. These cost reductions are likely to strengthen both domestic and economic linkages, which are necessary for the increased economic integration among the member countries. The stylised facts described in Chapter I show that trade between European Union countries has grown considerably since the creation of the EEC in the late 1950s. However, the simple observation of intra-eu trade patterns does not shed light upon the determinants that caused this significant increase in bilateral exports. These may involve the EU 8 treaties (i.e. institutional effects) or may simply reflect the growth, the wealth and the reduction of countries transport costs attained by trading with their neighbours. A further determinant of intra-eu trade might be the impact of globalisation on trade flows, which reduces international (and national) transaction costs. 5 Wang and Winters (1991), page 7. 6 The border effect measures the extent to which domestic sales of a country are greater than its external trade, after allowing for the effects of economic size, distance, and alternative trading opportunities. 7 Eichengreen and Irwin (1998), pages 33 and 34. 8 For simplicity, we will use the terms EEC, EC and EU interchangeably all throughout this chapter, if this distinction is not fundamental. 8

Hence, we need a method that will control for the so-called natural determinants of trade and at the same time capture the institutional effect of the EU integration process, time and country-specific effects. We accomplish this by estimating a gravity equation. As mentioned in Chapter II, the gravity model is a useful tool to explain bilateral trade as a proportion of the product of both countries masses and inversely related to the distance between them. The baseline gravity model that we consider includes mass of the trading countries, their wealth and dissimilarity, and the distance, adjacency and common cultural linkages between them. It is important to clarify that we do not intend to derive economic welfare results. As Viner (1950) noted, there can be are positive and negative welfare effects when creating a customs union (CU). We are just analysing the bilateral trade determinants of intra-eu trade integration. We add variables to the baseline gravity model that will help us explain the effects of EU and EFTA on bilateral trade flows. As noted in Chapter II, some gravity models introduce internal and external trade and distance proxies to measure the width of the borders, and others include trade bloc dummies to test how natural those blocs are. In this paper, we take a combined approach that tries to measure the width of the borders, which we call institutional effects, by also controlling for the specific time and individual effects of our sample. Our approach adds another dimension to the surveyed studies by estimating a panel data model and uses other techniques in order to account for some of the problems generated by this estimation. More specifically, we estimate an Instrumental Variables General Least Squares (IV-GLS) model developed by Hausman and Taylor (1981). This will allow us to instrument for some endogenous time-varying determinants within a setup that includes time-invariant regressors, which would otherwise be swept out from the estimation. III.2.2. Data and Proxies Our sample contains annual data from 1960 to 1997 for twenty-one developed OECD countries 9 and the RW. We included the latter as the twenty-second country to consider the total trade flows of each country in our sample. As the common feature in gravity models, our dependent variable is bilateral exports between these twenty- 9 The OECD countries are the same used in Chapter I: Australia, Austria, Brussels + Luxembourg, Canada, Denmark, Finland, France, Germany, Greece, Ireland, Israel, Italy, Japan, the Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, the UK, and the US. 9

two countries. Thus, we are working with 462 pairs of countries per year, which without considering missing data becomes a total of 15,178 observations. Bilateral exports and exports price indices are provided by the Directions of Trade Statistics (IMF), and the constant mass measures by the World Development Indicators (World Bank). Pairwise great circle distance measures were taken from Frankel, Stein and Wei (1995) and Wei and Frankel (1995). We created the dummy variables. Our model is based upon four main determinants of bilateral exports: (i) the baseline gravity variables; (ii) the integration effects; (iii) the time effects; (iv) the individual countries characteristics. Each of them comprises a set of regressors. Meanwhile the first two sets capture the main economic and institutional determinants of bilateral exports flows; (iii) and (iv) allow us to explore further the stylised facts described in Chapter I. III.2.2.1. The Baseline Gravity Variables Like most of the studies within the gravity framework, our specification includes the following baseline variables: Mass of both importer and exporter countries. In order to capture the exporter s production we take its GDP, and to proxy the importing country s absorption we use GNP. They are estimated separately to explain exports, as opposed to trade flows, and there is no prior reason to impose a common coefficient between them. Wealth of both exporting and importing countries, as a means of reflecting each country s prosperity. The proxies used are GDP (GNP) per capita for the exporter (importer). 10 Dissimilarity between trading countries, to reflect the relative wealth between them. We measure it as the absolute difference between the wealth of exporting and importing countries. This variable is partly included to control for the potential multicollinearity between the mass and wealth measures. 11 Transport costs, proxied by the great circle physical distance between the capital cities 12 of the trading countries and an adjacency dummy variable. 10 Including global per capita GDP together with GDP is equivalent to taking the latter with population, as Brun, Guillaumont and de Melo (1998) show. 11 One of the ways of solving this problem is to formalise a relationship among these regressors. 12 As will be discussed in the next chapter, a better way of capturing the transport costs involved in trade would be to find a measure of economic distance. 10

Linguistic ties, given that having a language in common may have an impact on transaction costs and increase the transborder contacts and information flows. III.2.2.2. The Integration (institutional) Effects As mentioned in Chapter II, some of the literature uses the gravity equation to measure the width of the borders between two countries. However, this approach needs not only external but also internal bilateral trade and distance measures. For most countries, internal trade and distance proxies are difficult to find, and thus require ad-hoc calculations. Another way of measuring trade integration effects is to estimate a gravity equation for international trade flows, including dummies to account for the determinants of regionalisation among countries. We consider the latter approach more accurate and less sensitive to the calculations adopted, given that it only relies on observed data. In order to test for the EU trade integration process itself, i.e. to distinguish between globalisation and European integration effects, we add the following variables to the baseline gravity model: An EU and an EFTA dummy variables to reflect the time-average effect of the trade between two EU or EFTA member countries. They include the countries joining these blocs at each point in time, and thus vary both through time and across individuals. The idea is to consider the EU (EFTA) as one single country (with the bilateral internal trade given by the individual members), and measure the width of its border with non-eu (EFTA) countries. Some of our panel data specifications include a timeinteracting EU dummy to analyse the evolution of the EU institutional effect over time. For capturing the external effect 13 of the EU enlargements, three types of dummies for exporting countries were included: Enlargement: indicate the long-run external effects of each EU enlargement. New member: to capture the medium-run external effects of a country being a new EU partner. Joining country: to look at the short-run (or first-year impact) of joining EU members. 11

III.2.2.3. Unobservable Time Effects and Country Characteristics As described in Chapter I, along the period 1960-1997 economic conditions have changed considerably. Hence, accounting for unobservable time effects is useful to capture certain trade determinants that could bias our estimates of EU integration. Time effects can be partly due to the globalisation process. Depending on the specification of the model, we proxy globalisation effects with a time trend or time dummies. The former measures the average effects and the latter is useful to analyse the time effects evolution at each point in time and to identify structural time breaks in the sample. Following Brun, Guillaumont and de Melo (1998), we include a linear time trend. We also tried to include a quadratic time trend, 14 but this was not significantly different from zero. Country individual characteristics can be proxied by including n-1 exporting country dummies, taking as base country the RW. Hence, these dummies measure the relative impact of each exporter to the bilateral trade flows with respect to the RW s. The inclusion of these country-specific effects is important to capture any determinant not accounted for in the mass and wealth measures. This not the same as estimating panel data individual effects, since given our individuals are pairs of countries. III.2.3. Parameters of Interest Incorporating the four sets of regressors described in the previous section, the general log linear specification for our gravity model is: (III.10) where: x, m exporting and importing countries, respectively. X xmt natural logarithm of real bilateral exports between x and m in year t. M xt and M mt natural logarithms of real masses of x and m (respectively) in year t. Mpc xt and Mpc mt natural logarithms of real wealth of x and m (respectively) in year t. DISIM xmt natural logarithm of dissimilarity between x and m in year t. 13 Eichengreen and Irwin (1998) add this dummy for only one of the two countries participating in the trade arrangement to test for the external effect of the grouping on trade with nonmembers. 12

DIST xm natural logarithm of distance between x and m. 15 D xm is a partitioned matrix of dummy variables to control for adjacency (ADJ xm) and linguistic ties (LINGTIE xm) between x and m:, and hence EU xmt and EFTA xmt dummy variables for common EU or EFTA membership in t. E xt a partitioned matrix to account for enlargement external effects:, and hence where: ENLARj xt = 1 from the t that x joined the EU onwards, if x joined the EU in the jth. enlargement. NEWj xt = 1 from the t that x was a new member of the EU until the following enlargement, if x joined the EU in the jth. enlargement. JOINj xt = 1 on the year of joining the EU if x joined in the jth. enlargement. for (j = 1, 2, 3) and zero otherwise. t linear time trend. T t set of T-1 time dummies, with 1960 as the base year. C x set of n-1 country dummies, with the RW as the base country. ε xmt error term. Let us first briefly describe the baseline parameters, i.e. the (i) set of regressors. Income (wealth) elasticities of bilateral exports for exporting and importing countries are α 1 and α 2 (α 3 and α 4), respectively. A positive α 5 indicates that two countries trade more with each other the more their wealth differs and could be interpreted as favouring comparative advantage inter-industry trade theory. Conversely, a negative coefficient for DISIM xm, suggesting that two countries with similar endowments trade more, would support an intra-industry trade explanation. Given that we take DIST xm, ADJ xm and LINGTIE xm as proxies for transportation and transaction costs, we expect α 6 14 They expect a concave evolution of the time trend, mainly due to the 1970s oil shocks and the contrashocks of 1985 and late 1990s. 15 Although the Newtonian formula indicates that square distance enters the equation, we follow the usual approach of including a linear distance, since the former was not statistically significant. 13

<0, and the vector α 7 to have positive elements. In other words, the greater the distance between two countries, the smaller their trade (ceteris paribus), and the greater the link between two countries (either geographical or cultural), the greater the amount of bilateral trade between them. Our main interest is in the β coefficients, i.e. the institutional effects of the creation of the EU and EFTA on bilateral exports. These dummies take the value of one if both countries in the trading pair belong to the corresponding trade agreement in year t. For example, the pair France-Italy will have a 1 for the whole 1960-1997 period, whereas France-Spain will only have a one since 1986, when the latter became a member of the EU, and similarly for EFTA. Hence, the base categories for the EU (EFTA) dummy is composed by the pairs of non-eu (EFTA) with non-eu (EFTA) countries and those of EU (EFTA) with non-eu (EFTA) members. The definition of our EU dummy differs with that used in Frankel (1997), where the EC bloc dummy does not vary over time. In a panel data model, the EU effect is composed of β 1 and β 3. The latter is the EU effect at each point in time. If the former were not included, the EU estimate would be biased (i.e., we allow for a non-zero origin of the trended EU coefficient). These semielasticities indicate that two EU members trade more with one another than predicted by their natural trade determinants and the average behaviour of the developed OECD countries. In other words, it suggests an increase in intra-bloc trade. We look at the external effects of the three EU enlargements with the vector of coefficients β 4. To differentiate between the speeds of adjustment of each group of joiners, we define three types of external effects of each enlargement depending on its persistence over time. The long-run (ENLARj xt) measures the effect of joiners since they became EU members; the medium-run effect (NEWj xt) shows the period during which joiners are new members of the EU; and the short-run (JOINj xt) captures the one-year effect of joining the EU. In order to facilitate the interpretation of these enlargement dummies, it is convenient to clarify the base countries of these dummies: The base countries for ENLAR1 xt are all but the UK, Denmark and Ireland for all t, and all countries for all t < 1973. ENLAR2 xt has all countries other than the Mediterranean new members for all t, Greece for all t < 1981, and Portugal and Spain for all t < 1986 as base. For ENLAR3 xt the base are all countries but Austria, Finland and Sweden for all t, and all countries for all t < 1995. 14

The base countries for NEW1 xt are all but the three first joiners for all t, and all countries for all t < 1973 and t 1981. In NEW2 xt, all except the Mediterranean countries for all t, Greece for all t < 1981 and t 1995, and Portugal and Spain for all t < 1986 and t 1995; and for ENLAR3 xt the non-eu15 and the EC12 members for all t, and all countries for all t < 1995. The base countries for JOIN1 xt are all but the first joiners for all t, and all countries for all t 1973; for JOIN2 xt, all countries except the Mediterranean for all t, Greece for all t 1981, and Spain and Portugal for all t 1986; and for ENLAR3 xt all countries except the last joiners for all t, and all countries for all t 1995. Hence, for example, a positive and significant indicates that Denmark, Ireland and the UK increased their bilateral exports relatively more than the other EU and OECD countries. A negative indicates a medium-run decrease in the bilateral exports due to Greece, Portugal and Spain joining the EC. A positive shows a short-run increase in the bilateral exports of Austria, Finland and Sweden after joining the EU. In addition, a positive ϕ indicates that on average, globalisation (among other unobservable time effects) increases bilateral exports. 16 Similarly, a negative γ t for a given period indicates that unobservable time effects decreased bilateral exports compared to the beginning of the period. Finally, a negative coefficient for a particular exporting country (δ x ) suggests that this has a smaller impact on bilateral trade flows than the RW s. We will interchange some of the proxies defined in (ii)-(iv) according to the different dimensions of the alternative methods estimated, without any loss of generality or change in the interpretation of the results obtained. III.4. The Results This section presents the results for the three methods described in III.3. In sum, SURE, accounts for the time correlation between the gravity cross-sections, whereas GLS controls for unobserved time and individual specific effects. Finally, IV-GLS 16 This could also be indicating a positive globalisation effect lower, in absolute value, than the other unobservable time effects. 15

solves the endogeneity problem together with estimating the marginal effects of the time-invariant regressors. III.4.1. SURE Estimation The first step towards estimating our gravity model is to run cross-section estimations. This allows us to test for the baseline model predictions, together with the additional trade determinants included in our model. Following Wei (1996) and Helliwell (1996, 1997), we estimate a system of cross-section equations of the form of (III.11) for each time period of our sample. The Breusch-Pagan test rejects the null hypothesis of a diagonal variance-covariance of the estimated errors matrix, thus confirming the presumed time correlation between the countries cross-sections for different years. This means that it is appropriate to estimate a SURE system of gravity cross-section equations, since it links them through their error terms. These reflect, among other things, the time differences between the cross sections. Although estimating cross sections is not the most efficient way of analysing the effects of different regressors on bilateral trade flows, they allow us to look at the time paths of the estimated export elasticities. Moreover, in this setting we can test for structural change in the coefficients. The overall fit of the model is similar to the ones reported in previous studies. An increasing Adjusted R-squared, with a mean of 86%, shows the good joint significance of the regressors in explaining the bilateral exports. 17 For convenience, we present here the estimated coefficients in graphs. The values for these coefficients and the result of the Breusch-Pagan test can be found in Appendix B. III.4.1.1. Baseline gravity variables Graphs III.1., III.2. and III.3. show the estimated elasticities and semi-elasticities of bilateral exports with respect to the baseline gravity variables. These are the coefficients of the estimated SURE model for the masses of exporting and importing countries, the distance between them, their wealth and their common characteristics. The mass elasticities of the bilateral exports predicted by our model are similar to those found in the literature. We find that on average, a marginal increase in a 17 The number of observations decreases when estimating a SURE, given that availability of observations around the whole time period is imposed. However, estimating a SURE for less years, and then getting more observations per year, gave smaller Adjusted-R 2. 16

country s real size (GDP) or in its real absorption (GNP) generates a significant 18 average increase of 0.7% in both its real 19 exports to and its imports from the other developed OECD countries and the RW. If two countries were twice as apart from each other as from a third country, their trade is on average 0.9% lower. Graph III.1. shows the evolution through time of these elasticities. Graph III.1. Wald tests indicate a significant change over the period for GDP x and for DIST xm. The significantly decreasing relative importance of the exporting country s mass measure, indicates that the size of a country has lost some relevance in explaining its exports. Moreover, size has become a relatively better determinant for the imports of a country, thus indicating a higher preference for imported goods of industrial countries. In order to verify some of the facts described in Chapter I, we performed various Wald tests to account for the significant structural change in the coefficients over time. 20 We found that the break-up of Bretton Woods (1971) has increased significantly the exporter s income elasticity of exports for OECD countries. The drop in this elasticity, which may be partly due to the first oil crisis, is also significant. We also find that the second oil crisis has apparently increased significantly the sensitivity of exports to the importer s income. The relatively stationary pattern around a significantly decreasing trend of the distance elasticity of exports might be due to a shrinking in the transport costs during 18 Unless otherwise stated, this refers to a significantly different from zero at a 1% level. 19 From now on we will refer to real economic variables. 20 The null hypothesis for these tests is that the corresponding coefficient did not change from one year to the following. 17

the last four decades. Moreover, this finding may also be indicating a missspecification problem because of only including a geographical (constant) distance measure, which we tackle in Chapter IV. Graph III.2. shows that the estimated exporting country s wealth effect has been steady until the early 1980s, when it begins to decline. Moreover, this coefficient only becomes non-significantly different from zero towards the end of the period. Throughout the period considered, a one-percentage change in the wealth of a country increases its exports by almost the same amount, whereas decreases its imports by 0.8% on average. Graph III.2. Wald tests indicate a significant change over the period for GDPpc x GDPpc m (at a 5% level). (at a 1% level) and for The negative (and generally significant) elasticity of the bilateral exports with respect to the importer s wealth seems to contradict the common finding of a positive correlation between a country s wealth and its imports. This may be due to a higher domestic income effect, which makes people consume relatively more domestically produced goods the wealthier they are. As mentioned above, to control for multicollinearity of wealth and mass measures, we include DISIM xm. While its mean increases considerably over time, its elasticity is rather small compared to the wealth effects. We find that if two countries become 1% more similar, their trade is on average a 0.1% lower. In addition, the significance of the coefficient varies over time. The positive coefficient of DISIM xm supports the hypothesis of high correlation between GDP per capita differences and differences in factor endowments. This 18

correlation leads us to conclude that smaller differences between countries could reduce trade, especially inter-industry trade driven by comparative advantage. Hence, our results would be more in favour of a Heckscher-Ohlin explanation of trade flows. Moreover, this result opposes the Linder Hypothesis, where trade is higher between countries with similar living standards, given that they share a broader range of goods to trade. Testing for yearly structural breaks in these variables we found that the second oil crisis has increased significantly the sensitivity of exports to the importer s wealth and its dissimilarity with respect to its partner. Graph III.3. shows the path of the adjacency (ADJ xm) and linguistic ties (LINGTIE xm) effects on two countries exports. 21 While a marginal effect of being next to the importing country would make the exports of a country an average of 0.4% greater, sharing a common language would increase its exports by an average of 0.6%. However, these results should be considered with care, given that these coefficients on both dummies are in general not significantly different from zero, even at a 10% level. Graph III.3. Wald tests indicate that neither ADJ xm nor LINGTIE xm have experienced a significant change, even at a 10% level. Although not too significant, the trend seems to indicate that sharing a common language has become less important for doing business during the last four decades. This may be due to the widespread usage of English as the international business language. The fact that having a common border is slightly becoming a better determinant of bilateral trade patterns is perhaps related to a decrease in the transportation costs over the period considered. Given that our time-invariant 19

distance measure does not capture these effects, we would expect to account for them by finding a measure of economic distance that varies through time. III.4.1.2. The integration effects The semi-elasticity of bilateral trade with respect to the EU dummy allows us to measure the so-called border effect of the European Union trade with non-eu countries. Graph III.4. shows the evolution of the implied integration elasticities (or estimated coefficients) for EU and EFTA blocs. 22 We find that on average EU membership has increased bilateral exports by 26%, 23 whereas this has decreased for two members of the EFTA by approximately 16% (although this is not too significantly different from zero). 24 The decreasing path of the EU xm coefficient since the late 1960s may be partly due to the consolidation of the bloc after each enlargement, thus leaving less and less scope to trade with the new joining countries. However, we will see later that unobservable time and individual effects account for much of this trend. Graph III.4. Wald tests indicate that only EFTA xm changed significantly over the period (at a 1% level). Although, as far as we are aware, there are no studies that estimate the evolution of the EU integration, our findings coincide with other results found in the literature, for somewhat different set-ups and models. 21 In this case, we refer to the vector of coefficients in equation (III.11). 22 External effects of exporting EU countries were not significantly different from zero at a 10%, and hence were not included. 23 In this case, it is more convenient to work with 100% changes of the explanatory variables. 24 To a smaller extent, EFTA dummy depends inversely on the enlargements of the EU, since countries like Denmark, the UK and Finland left it as were joining the EU. 20

Nitsch (1999) estimates a SURE system of gravity equations using internal and external measures of trade, masses and distance, for the first twelve EU members 25 over the period 1979-1990. Bearing in mind the different methodologies and data employed, our results do not differ from his findings of a decreasing home bias. 26 Frankel (1997) estimates gravity cross-section equations for every-five years and finds that the EU effect is not significant until 1985, when it attains a 20%, which then rises to 30% in 1990. The positive trend of our EU integration indicator during the 1960s reflects the increase in trade integration of the EEC6 countries. This was mainly favoured by the stability of having their currencies fully convertible at fixed, but adjustable, exchange rates ruled by the Bretton Woods System. Our indicator also shows a positive and significant 27 response to the completion of the customs union in 1968. The benefit of the EU effect slowed in 1970, mainly due to the inflationary pressures on the stable fixed exchange rates system, which drove to the adoption of flexible exchange rates regimes by 1971. De Grauwe (1988) finds that, although with relatively stable exchange rates, EEC6 members experienced a strong decline in their tradeintegration process since the 1970s. In the early 1970s, EU integration seemed to be recovering its upward path. However, the first oil shock of 1973 exacerbated the very strong inflationary pressures and plunged most oil importing countries into massive trade deficits. This led most OECD countries in general to adopt national protectionist measures, thus reducing their trade. Given the relatively high trade integration achieved by the EEC6, this change in policies might have dampened their intra-eec trade even further. Moreover, EU effect s main drops coincide with the three enlargements of the European Union. 28 This finding can be explained by three different motives, concerning the trade orientation of joiners, their size and the external economic circumstances of each period. 25 Given the period of his sample, Nitsch does not consider the third enlargement. 26 Although Nitsch employs a different methodology, this variable is capturing the EU effect, comparable with our EU dummy. 27 We ran Wald tests to check for structural changes on the EU coefficients, and rejected the null of equal coefficients at a 1% level. 28 This is the other reason why Frankel (1997) adopts as time-invariant EU dummy. However, given that our arguments are plausible with the stylised facts described in Chapter I, we will keep a time-varying EU dummy. 21

The trade orientation cause entails a learning process for the trade patterns among the old and the new EU countries when the latter first enter the union. This process partly determines the extent to which the integration coefficient falls and then recuperates. 29 By learning process we mean the adjustment that new members necessarily do in order to catch up with the degree of integration achieved by the old EU countries. For example, in the first enlargement, the three new members had to go through a transitional process to finally approach the degree of integration achieved in fifteen years by the EEC6. Moreover, specially UK s stronger ties with non-eec6 countries also contributed to its learning process. The second and third causes can also be better observed taking the first enlargement. The significant 30 decrease in the EU coefficient may also be picking up the arrival of a big country, the UK, into the EU group. Finally, the temporary increase in the weight of imports from OPEC countries during the 1960s may have contributed to lowering the intra-eec trade share during the early 1970s. In other words, the first oil shock, together with the inclusion of three countries that handled their economic policies differently to the EEC6 s, might also have affected trade patterns between the EEC9. Other results of the yearly-structural change Wald tests on the EU estimated coefficients indicate that neither the establishment of the EMS (1979) nor the inclusion of Greece (1981) significantly affected EU integration. Spain and Portugal EU only had a significant negative effect on EU trade integration one year after joining this bloc during the same year that the SEA was adopted (1987). The late 1980s sought a change in trade patterns in favour of intra-eec trade. The early 1990s significant recovery may have been partly favoured by the repercussion of the SEA (1987) and the Maastricht treaty of 1991. EU trade integration appears to recover a slightly positive trend in 1993, with a smaller but significant learning process negative effect after the third enlargement. Finally, the decrease of the EFTA coefficient since the first EU enlargement, which coincided with the loss of two important member countries, indicates an increase in the relative importance of the EU over the other European free trade agreement. This could be partly explained by the more limited agreements taken by EFTA members, as explained in Chapter I. Conversely, EU countries, by adopting a customs union by 29 The decrease in the significance of this dummy after the mid-1990s does not allow us to observe a reversion of the trend after the third enlargement. 30 Wald test rejected at a 1% level. 22

signing the Treaty of Rome, and then an economic union with the Maastricht Treaty, compromised committed themselves to a higher degree of economic integration. III.5. Main Findings This chapter has focused on three methods to measure EU trade integration. Given that there is a significant time-correlation between the errors of each cross-section gravity equation, we have first estimated a SURE model for a system of thirty-eight cross-sections. We find that being in the EU increases bilateral exports by 26%. Because of the need to control for unobserved individual effects without sweeping out the effects of observable time-invariant regressors, our second chosen method is a panel data GLS. Given the panel data time dimension, we add a globalisation proxy and the high degrees of freedom allow us to incorporate long, medium and short-run external effects of the three EU enlargements. However, GLS seems to derive inconsistent estimators, may be due to an endogeneity problem between masses and exports. Hence, we finally adopt the HT method that allows us to instrument the mass variables by using internal instruments and is able to estimate the marginal effects of time-invariant regressors. Rejecting the equality between the estimated coefficients of HT and GLS methods, we can conclude that the former derives (at least) more consistent estimators. Our main general result is that not accounting for unobservable individual effects and for the endogeneity of mass measures biases upwards the EU effect on bilateral trade. Since HT, unlike SURE, controls for unobserved individual effects and because, converse to both SURE and GLS, it solves the endogeneity problem, we next analyse in greater detail the EU integration effects results from the HT method. EU integration was positively affected by globalisation: not including time effects (model A) gives us an EU coefficient of 12%, whereas doing so (model B) leads to an EU estimate of 20%. In addition, accounting for the external effects of enlargements (model C) take much of the EU effect: it goes down to 3%. Our results are consistent with Frankel s (1997), who finds an EEC12 bloc effect of 16%, and a decrease in the EC bloc when including openness effects. With regard to the significant enlargement effects, our findings coincide with the evolution of the EU effect for the SURE method. The first and third enlargements show positive long-run external effects on bilateral exports of 48% and 32%, 23