The euro effect on trade.

(1)

Universiteit van Amsterdam Faculty of Economics and Business

The euro effect on trade

Student: Regina Kiss Student number: 5925142 MSc thesis

Supervisors:

Dr. Maurice Bun Dr. Franc Klaassen

(2)

1. Introduction

Amid the unfolding eurozone debt crisis, the costs and benefits of the common currency got into the forefront of policy discussions again. Being a member of a monetary union has its pros and cons. The main disadvantage is the loss of monetary policy as a national stabilisation tool. The expected advantage is the enhancement of investment and trade, in line with the Optimal Currency Area (OCA) theory of Mundell (1961).

This study focuses on the aggregate trade effects of the euro. Following the completion of the Single Market project in 1992, which removed the remaining administrative and technical barriers to integration, the introduction of the common currency in 1999 was expected to foster further intra-eurozone trade. Lower transaction costs, greater price transparency and reduced exchange rate volatility were identified as the main channels through which the common currency could potentially enhance intra-eurozone trade.

The literature on the aggregate effect of a common currency on trade has been launched by the influential paper of Rose (2000). The main finding of the study was that currency union members traded with each other three times as much as with similar trading partners. The paper was motivated by the introduction of the euro in 1999, but the estimates were based on existing data for much smaller countries that had adopted currency unions in the past. The applied methodology was the gravity model of international trade including a dummy for the currency union. In a few years enough data had accumulated to allow for an analysis of the trade effects of the euro and since then numerous papers have been published, mostly also applying variants of the gravity model.

Havranek (2010) performed a meta-analysis of the eurozone strand of the Rosean literature and concluded that it has almost completed the economic research cycle. The results are getting close to those ’before Rose’, finding that the euro had only a small effect on intra-eurozone trade. The aim of this study is to examine whether the aggregate trade effect of the euro changes when recent developments are taken into account. The gravity specification of Bun and Klaassen (2007) is estimated using an updated dataset, because it had an important role in forming the general consensus of the literature that when the gravity specification includes only country-pair and time fixed effects, there are omitted time-varying variables (the so-called ’Anderson - van Wincoop misinterpretation’). In order to account for time-varying unobserved heterogeneity, Bun and Klaassen (2007) includes country-pair specific trends in the gravity model. Several recent papers (e.g. Gengenbach (2010), Camarero et al. (2012)) consider their specification as a benchmark and estimate it by using new generation panel techniques.

In the followings the theoretical foundations of the trade gravity model are presented briefly (Section 2), the empirical literature is reviewed (Section 3), the estimation is carried out (Section 4) and finally the conclusion is drawn (Section 5).

(4)

2. Theoretical foundations of the trade gravity model

The analogy with the law of gravity in physics was applied for the first time in economics in the 19th century by Ravenstein (Anderson, 2010), who used it to describe migration patterns. Tinbergen (1962) was the first to explain trade flows with the concept of gravity.

According to the traditional trade gravity specification, the mass of goods produced at origin i (Y_i) is proportional to a mass of demand for goods at destination j (E_j), but the potential flow is reduced by distance between them (dij) and increased by the so-called gravitational constant

(G). Therefore the predicted movement of goods between countries iand j (X_ij) equals:

X_ij = GYiEj

d_ij . (1)

As an empirical specification the traditional gravity had a high explanatory power, especially when other proxies of trade frictions had been added as well (e.g. political borders, common language etc.). However there were two critical points of the model. First it was lacking strong theoretical foundations. Second, prices were not included in the model, based on the assumption that they are identical across countries.1

Regarding the first shortcoming, microfoundations have been provided within different theoretical frameworks since the end of the 1970s. Anderson (1979) was the first to provide clear microfoun-dations by deriving gravity from a CES (Constant Elasticity of Substitution) preference structure and assuming that goods were differentiated by the country of origin (the so-called Armington assumption). Later the new trade theory provided microfoundations: Bergstrand (1985, 1989) used a monopolistic trade model with goods being differentiated by firms instead of countries. Subsequently Deardorff (1998) derived a gravity specification from the Heckscher-Ohlin model, which explains trade based on relative differences in factor endowments. Eaton and Kortum (2002) used a Ricardian type of model, where trade arises based on technological differences across countries. Chaney (2008) introduces firm heterogeneity to the gravity model and shows that contrary to the model with identical firms; in this case the impact of trade barriers on trade flows is decreased by the elasticity of substitution and not magnified.

The second problem, the issue of identical prices has been somewhat corrected by the above-mentioned models with microfoundations, as they necessarily allow for differing prices due to trade barriers. However they account for the difference in prices due to trade costs only bilaterally, ignoring the fact that any bilateral sale interacts with all others (the so-called resistance of country i to export / import from others than a given country j). This general equilibrium problem has been solved by the structural gravity model of Anderson and van Wincoop (2003), which includes

1

The traditional gravity literature recently developed remoteness indexes of each country’s ’average’ effective

distance to or from its partners (P

idij/Yij). However as it is shown later, this still means accounting only for

(5)

the comparative statics of multilateral resistance.In the following the structural gravity model of Anderson and van Wincoop (2003) is briefly discussed because the econometrics of the euro trade effect literature is mostly based on their paper.

The first building block of the model is the Armington / CES preference structure. Note that the Armington assumption (i.e. each nation produces a unique good) implies that all adjustment happens in the price of the exported good. The preferences are homothetic (i.e. shares are invari-ant to income) and identical, therefore the same parameters characterize expenditure behaviour in all locations. Ifc_ij is the consumption of goods produced in countryiby consumers of country j, then consumers in country j maximize the following CES expenditure function:

max cij X i β 1−σ σ i c σ−1 σ ij _σ−1σ , (2)

subject to their budget constraint:

p_ijc_ij = y_j, (3)

whereβ_i is a distribution parameter,c_ij is consumption of the good exported fromitoj,σis the elasticity of substitution (σ > 1),p_ij is the consumer price of the good exported from ito j and yj is the nominal income of the country j residents.

Note that the ’distribution parameter’β_ican have several interpretations. Here it is an exogenous taste parameter, however for instance in the monopolistic competition framework (e.g. Bergstrand (1989)) and recent interpretations of Anderson and van Wincoop (2003) (e.g. Baldwin (2006) and (2008)) it is proportional to the number of firms exporting distinct varieties from country i to j (n_ij). As mentioned earlier, Anderson and van Wincoop (2003) assumes that each nation produces only one good and all adjustment happens in the price of the traded good. However in the monopolistic competition models, all adjustment happens in the number of varieties that each nation has to offer. On the whole, in both variations of the gravity model, the larger the exporting country, the more it has to offer. In the case of the Anderson and van Wincoop (2003) version the large country exports more because it has lower prices, and in the monopolistic model it sells more products because it has more firms offering distinct product varieties.

Returning back to the Anderson and van Wincoop (2003) derivation, the consumer price of the good exported from country i toj is defined as follows:

p_ij = p_it_ij, (4)

wherepi is the exporter’s supply price (the so-called factory gate price) andtij is the trade cost

factor between origin iand destination j.

However it is also important to note that the consumer price of the exported good has a broader definition in other interpretations of the gravity model. For instance Baldwin and Taglioni (2006) define the consumer price as:

(6)

p_ij = t_ijµ_ije_ijmc_i, (5)

whereµij is the bilateral mark-up, mci is the marginal cost of production and eij is the nominal

bilateral exchange rate. The inclusion of the bilateral mark-up underpins theoretically that the euro may increase trade not only through lower transaction costs (these show up in the trade cost factor), but also by enhancing greater price transparency (this affects the bilateral mark-up channel) and by decreasing exchange rate volatility. The marginal cost of production is included in the Anderson and van Wincoop (2003) specification (4) as well through the exporter’s supply price (pi).

Continuing the Anderson and van Wincoop (2003) derivation - and wrapping up the first building block of the model, the Armington / CES preference structure - the maximization of (2) subject to (3) provides the nominal demand for country i goods by countryj consumers:

xij =

β_ip_it_ij P_j

1−σ

yj, (6)

whereP_j is the price level of goods in countryj, given:

P_j = X

i

(β_ip_it_ij)1−σ1/1−σ. (7)

The second building block of the model – in line with its general economic equilibrium structure – imposes market clearance. The total nominal income in countryihave to be equal to the nominal value of total exports to allj destinations:

yi=

X

j

xij. (8)

Dividing both sides of (8) byy_jand summing overjyields a solution for the scaled prices(β_ip_i)1−σ:

(β_ip_i)1−σ =_P yi

j(tij/Pj)1−σyj

. (9)

Substituting the scaled prices into the demand equation (6) and defining the world income y_w= P jyj yields: x_ij = yiyj y_w t_ij Π_iP_j 1−σ , (10) where

(7)

Π1−σ_i =X j t_ij Pj 1−σ y_j yw , (11)

is the exporting nation’s ’market access’ or ’market potential’. By substituting the scaled prices into (7) we get: P_j1−σ =X i t_ij Π_i 1−σ y_i y_w, (12)

whereΠ_iandP_j are the outward and inward multilateral resistances. These are called multilateral resistances because each of them depends on bilateral trade barriers with all partners (t_ij). The ’outward’ or ’inward’ refers to whether it is the resistance of the exporter or the importer country. Equations (10)-(11)-(12) constitute the structural gravity model. The key implication (equation (10)) is that after controlling for size, trade between countries depends on the bilateral barrier between them (t_ij) relative to the average trade barrier the two countries face with all their trading partners (Π_iP_j). Regarding all three equations, it can be seen that the effect of trade costs on the nominal value of bilateral trade is complex, as it shows up both directly and indirectly through the multilateral resistance terms.

(8)

3. Empirical literature

As mentioned earlier, the theory and the econometrics of the euro trade effect literature is mostly based on the structural gravity model of Anderson and van Wincoop (2003). Most of the papers in the literature (e.g. Micco et al. (2003), Flam and Nordstrom (2003), Berger and Nitsch (2005), Baldwin and Taglioni (2006)) estimated the variants of the following equation by taking logs and adding time subscript to (10):

ln x_ijt= ln y_ity_jt+ (1 − σ) ln t_ijt− (1 − σ) ln Π_it− (1 − σ) ln P_jt, (13)

where trade costs are assumed to be related to bilateral distance (D_ij), common euro usage (EM Uijt) and other factors (which potentially vary over time and across partners (Vijt)), therefore

the log oft_ijt equals:

ln tijt= β1ln Dij+ β2EM Uijt+ β3ln Vijt. (14)

The majority of the papers in the literature augmented (13) with further dummies (e.g. EU, EFTA dummies etc.) in order to account for trade integration among the members of the currency union. An important variable, which is not included in this empirical specification - but it is part of the theoretical model’s pricing equation (5) - is the real exchange rate variable. The use of the real exchange rate variable is fairly rare in empirical gravity equations, because usually bilateral trade is used as a dependent variable. In this case changes in the bilateral real exchange rate will have offsetting effects in bilateral trade. The inclusion of exchange rates is necessary when the dependent variable is unidirectional (e.g. exports are used).

The gravity equation (13) was estimated by Anderson and van Wincoop (2003) using cross-section data, similarly to earlier papers analysing the effect of currency unions on trade. A Least Squares Dummy Variable (LSDV) methodology was applied, where the country-pair dummies control for the unobserved time-constant characteristics of the given countries, which may influence the propensity of the countries to trade with each other. This methodology provides a consistent estimate whether countries trade more if they are in a currency union relative to non-members. However in the case of the EMU, a more relevant question is whether countries trade more when they become members of a currency union, therefore research on the euro’s trade effect mostly used panel data methodologies. Time dimension is important also because the introduction of the euro had been preceded by other stages of economic integration (customs union, European Monetary System, Single Market).

(9)

In the following the methodology of those studies is overviewed, which use panel techniques and aggregate data, as the current analysis updates the results of Bun and Klaassen (2007) which belongs to this strand of the literature. At the end of the literature review those studies will be also briefly summarised, that apply panel methodologies on micro data.

3.1 Panel studies using aggregate data

3.1.1 Static panel estimations

The early papers within this group applied directly the methodology of Anderson and van Win-coop (2003) to panel data, thus they used pair fixed effects to control for time-constant trade frictions (e.g. distance, cultural factors) and also time-fixed effects to control for developments common for all countries (e.g. global economic shocks).

However as shown in (13) the multilateral resistance terms: the exporting nation’s market access (Πi) and the relative prices in the importing nation (Pj) are time-varying. Additionally as it can

be seen in (14) trade costs are related to country-(pair) specific factors which potentially vary over time (V_ijt)2_{. When these time-varying variables are omitted, the residuals will be correlated with}

the EMU dummy, thus the coefficient for the euro effect on trade will be biased. This mistake has later been dubbed as the ’misinterpretation of Anderson and van Wincoop’ or ’partial gold medal mistake’3 _{(Baldwin and Taglioni, 2006).}

Beside the partial gold medal mistake, Baldwin and Taglioni (2006) identify two smaller flaws of the literature as well. The ’silver medal mistake’ is that many papers average the two-way bilateral flows, while the gravity equation is multiplicative, thus the product should be included instead of the sum. The ’bronze medal mistake’ is that while the gravity equation is based on a CES expenditure function (see equation 2) because trade data is collected in value terms, many authors act as it was based on a demand function and deflate the trade flows with a price index (usually US CPI or PPI). Moreover it is hard to get a good price index for bilateral trade. However the latter mistake is eliminated by including time dummies which is a universal practice.

Additionally to the above-mentioned methodological problems, there are also data issues that justify the need for time-varying country(pair)-specific variables. There was a change in the way of trade data collection in 1993: instead of the customs offices, VAT authorities started to report trade data. The problem with this is not only that tax enforcement changes can distort trade flows, but also that tax fraud creates a gap between the export and import statistics, as in the EU the tax is levied on imports and rebated for exports. Another difficulty is the so-called ’Rotterdam effect’: major part of EU imports arrive through large trade conjunctions and if the goods are not subject to TIR, re-import of the good into another eurozone country may increase intra-eurozone trade. A third problem is that around the introduction of the euro, the Pan-European Cumulation System was implemented, with the aim to simplify the bilateral FTAs through imposing uniform rules of origin, which could also affect trade flows. Handling these

2_{Bun and Klaassen (2007) points out that trade depends on such country-specific factors as productivitiy}

parameters, capital/labor ratios and country-pair specific variables such as transport costs, tariffs.

3_{According to Baldwin and Taglioni (2006), the full gold medal mistake applies when fixed effects are not}

used, thus even the cross-section correlation is not removed. This was typical in the traditional gravity literature, pre-Anderson - van Wincoop, when distance, common language etc. had been explicitly included in the regression, leaving unobservable variables omitted.

(10)

features of the data with simple fixes is impossible as the magnitude of these problems vary over time and member state.

Two further circumstances are also an indication that specifications including only country-pair and time fixed effects may not be sufficient for avoiding the omitted variable problem. The period 1992-1998 was particularly intensive from the point of view of integration, which could be easily controlled for if all EU members introduced the Single Market measures at the same time. However trade measures had been implemented at a different pace by the laggards and the fast implementers. Although to a lesser extent, but it can also distort trade flows, that the euro sharply depreciated after its introduction. The latter may cause an expenditure shifting effect in the case of those eurozone members which trade more with non-eurozone countries (e.g. the major part of Ireland’s trade flows is dollar-based).

In the followings papers applying static panel estimations are briefly summarised by taking into consideration the above-mentioned issues. The focus of the empirical review is on accounting for the omitted time-varying variable bias when estimating the euro effect. It is necessary to restrict the scope of the review, while there are several additional, interesting issues (e.g. the question of trade diversion, the distribution of the euro effect among eurozone countries etc.) which are not included in the current analysis.

(11)

Table 1: Static panel estimations

Authors Methodology Sample Main results

Micco et al. (2003)

Fixed effect, EU trend, dep. var.: total nominal trade

EU15 / 22 OECD, 1992-2002

Euro effect 6% for narrow, 4% for large control group. EMU

effect year-by-year estimates significant and increasing after

1998. Flam and

Nordstrom (2003)

Fixed effect, dummies correcting for data issues, EEA year dummies, dep. var.: directions

specific exports

20 OECD, 1989-2002

Euro effect 16%, EMU effect year-by-year estimates also

significant and increasing.

Berger and Nitsch (2005)

Fixed effect, country-specific institutional integration index / EMU specific linear time trend /

country-pair specific trend, dep. var.: total real trade

EU15 / 22 OECD, 1948-2003

Euro effect (large control group): 40% (baseline), 31% (integration index), zero and insignificant (EMU-specific trend) Baldwin and Taglioni (2006)

Fixed effect, time-varying nation and pair dummies, dep. var.:

total nominal trade

EU15 Euro effect -0.09%, not significant Bun and

Klaassen (2007)

Fixed effect, country-pair specific time trend, dep. var.:

total real trade

EU15, 19 OECD

Euro effect 3%, estimates no longer depend on the sample length in a systematic way Baldwin

and Taglioni

(2008)

Fixed effect, time-varying exporter and importer dummies,

EZ dummy interacted with Mongelli sub-index, dep. var.:

total nominal trade

EU15, 1996-2006

Euro effect 2%, coefficient of Single Market proxy small

(around zero)

Santos et al. (2010)

PML estimator, time-varying exporter and importer dummies,

dep. var.: total nominal trade

EU15, EEA, OECD93, 1993-2007

Euro effect 0.25-0.5%

Micco et al. (2003) is a seminal paper in the first wave of static estimations. They applied directly the methodology of Anderson and van Wincoop (2003) to panel data by including only country-pair and time fixed effects. There are two novelties in their paper. On the one hand they try to control for the varying degree of integration over time by including a trend for the EU, however they still disregard the heterogeneity across countries regarding the pace of implementation of reforms. On the other hand they examine yearly euro effects. However the partial gold medal mistake is present in their paper, because they do not control for time-varying country-pair specific variables.

As a starting point they show that the traditional way of including dummies / variables for common border, distance, language etc. in the gravity equation leads to omitted variable bias. The estimated euro effect is significantly lower when country-pair fixed effects are included in the model instead of specific variables. Their baseline specification is the following:

(12)

ln Tijt= αij+β1ln YitYjt+β2ln yityjt+β3F T Aijt+β4EUijt+β5EU trendijt+β6EM Uijt+γt+ijt, (15)

where α_ij represent country-pair fixed effects,γ_t represent year fixed effects, Y denotes GDP, y represent GDP per capita andEM Uijt,EUijt,F T Aijt are the euro- and the integration dummies.

A trend variable for the EU (EU trend_ijt) is also included in the estimation, in order to account for the varying degree of integration over time for the EU as a whole. However as mentioned earlier, there still remains a heterogeneity across countries regarding the pace of implementation of EU-wide reform (laggards vs. implementers). Other time-varying country-pair specific variables (denoted by V_ijt in (14)) are also not controlled for (partial gold medal mistake).

Dataset features partially mitigate the above-mentioned issues. The length of the sample is rather short (1992-2002), however this is beneficial for two reasons. First, as a result of the omitted time-varying country-pair specific variables, there must be serial correlation left in the residuals, which bias is moderated by the relatively short sample (1992-2002). Second, limiting the dataset to post 1992 data is also useful because the method of collecting EU trade data changed in 1993. The estimations are carried out both with developed and EU countries. The EU sample may provide better results because the control group is as similar to the treatment group as possible and this decreases the bias stemming from the lack of controlling for the different pace of implementing the Single Market measures.

The authors also examined the euro effect over time by interacting the EMU dummy4 _{with year}

dummies. The euro effect already appears in 1998, which is puzzling. According to the authors the reason is that already in May 1998 the irrevocable conversion rates were set, although other papers later in the literature claim that there were still some uncertainties at the time.

Flam and Nordstrom (2003) examines several additional aspects compared to Micco et al. (2003). They use unidirectional trade flows, which makes it possible to decompose the total euro effect into effect on export flows among eurozone members and among eurozone and non-eurozone members in both directions. They use dummies in order to control for the change in EU trade data collection in 1993 and the Rotterdam effect. They include time-varying EU dummies beside time-varying EMU dummies. However similarly to Micco et al. (2003) they also do not control for the differences in the pace of integration across countries and the more general shortcoming of failing to account for omitted time-varying factors (the partial gold medal mistake) is also present in their paper.

The authors estimated two alternative specifications, one with time-varying EU and EMU dum-mies and one with constant ones. The first specification, where the EU (EUt) and EMU dummies

(emu..t) are allowed to vary the whole period is the following:

ln X_ij = α_ij+ γ_t+ β₁ln(RY_i) + β₂ln(RY_j) + β₃ln(REXR_ij) + β₄ln(REXRc_j) + β₅eunew + β₆U R + β₇nomexr + β₈wbnt1993 + β₉weu10t1993 + β₁₀bneu12t1993 + β₁₁wbnt1995

+ β₁₂weu3t1995 + β₁₃bneu3t1995 + δtEUt+ ρ1temu11t+ ρ2temu12t+ ρ3temu21t, (16)

4_{In this case the EMU dummy is 1 for the whole period in case both countries were members of the EMU after}

(13)

where X_ij is exports from country i toj,α_ij and γt are country-pair and time fixed effects, RYi

and RY_j are the real GDPs of the exporting and the importing country, REXR_ij and REXRc_j are the real exchange rate between the exporting and importing country as well as the importing country and third countries, eunew is a dummy for the new EU members entering in 1995, U R is a dummy for the Uruguay Round liberalization, nomexr is the volatility of nominal exchange rate between the exporting and the importing country. The six dummies (wbnt,weu,bneu,wbnt, weu3t, bneu3t) control for two issues mentioned earlier: the changes in the way of trade data collection (1993) and the Rotterdam effect. The EMU effect is decomposed into effects on export flows among eurozone countries (emu11), from eurozone to non-eurozone countries (emu12) and from non-eurozone countries to eurozone countries (emu21)5_.

The time-varying EMU dummies show the difference in the level of exports of these three groups compared to the level of exports between non-eurozone countries each year. When the significance of the differences between years is tested, it can be seen that there is a break in 1998 and after that year intra-eurozone exports and less robustly exports from eurozone to non-eurozone countries increased. The authors argue similarly to Micco et al. (2003) that the trade effect appears already in 1998 because it was possible to hedge against exchange rate movements after the exchange rates were irrevocably fixed in May 1998. However it is important to note that in parallel with the increase in the euro effect, the effectiveness of the Single Market (the coefficient of the EU dummy) decreases, therefore the EMU dummy may pick up the effects of non-monetary integration as well. In the second specification the EMU dummies are kept constant for the euro period (1998-2002) and these results show that the euro effect is around 16%, which is still a relatively large estimate. This result also supports the earlier assumption that the EMU dummy may pick up an increasing trend which is caused by omitted time-varying variables.

The issue of omitted time-varying variables got into the focus of subsequent studies6_{. Berger and}

Nitsch (2005) show that estimating the specification of Micco et al. (2003) for a longer period yields a larger EMU trade effect. Their value added is including country(pair)-specific measures (e.g. institutional integration index, country-pair specific trends) of non-monetary integration dynamics instead of general trends used in earlier papers (e.g. EU trend in Micco et al. (2003)). Their specification is the following:

ln T_ijt= α + φ₁ln Y_itY_jt+ φ₂ln y_ity_jt+

n

X

k=1

β_kX_ijkt+ γEM U_t+ ϕF T A_t+ δIN T_t+ φ_ij+ η_t+ _ijt, (17)

where T_ijt is the volume of trade between country i and j in real US dollars, Y_itY_jt and y_ity_jt are real GDP and GDP per capita, X_ijkt is a set of n conditioning variables that are typically found to affect bilateral trade flows (distance, common language, common border, landlocked and island),EM Ut is the euro,F T Atis the free trade dummy andIN Ttis a pair-wise average of

country-specific measures of institutional integration.

When the index for institutional integration is included, the point estimate of the EMU effect decreases only slightly (from0.34to0.27for the developed country sample). However the overall

5_{Using this decomposition they are able to examine the trade diversion effects of the euro, as mentioned earlier}

this issue is out of the scope of the current analysis.

6_{Bun and Klaassen (2007) adds a country-pair specific trend to the gravity equation and get a euro trade effect}

(14)

long-term effect of European integration7 _{is double of the EMU’s trade effect, which seems more}

reasonable, taking into consideration that European integration has been going on for decades. Instead of the institutional integration index, the authors also try to include a simple linear trend, which is less refined than the index but provides flexibility for the model. In this case the euro effect becomes insignificant. The result is the same when country-pair specific trends are included. Therefore the authors conclude that the EMU is a continuation of a longer term trend. Baldwin and Taglioni (2006) also focus on the issue of omitted time-varying variables. The value added of their paper is that they add time-varying nation dummies to the gravity specification following the approach of Baltagi et al. (2003) in order to control for the multilateral resistance terms. Besides they have also identified the econometric flaws of the existing literature (dubbing them gold, silver and bronze medal mistake) by examining the sign and the size of these biases. However when the authors estimate their theoretically preferred specification, not only the euro’s but also the EU membership’s effect on bilateral trade becomes insignificant.

In their subsequent paper Baldwin and Taglioni (2008)) argue that the time-invariant, digital EU and EMU dummies are not able to capture the time-varying policy changes. Therefore they use the Single Market sub-index of the Mongelli index to control for the impact of EU membership and they improve the euro dummy by interacting it with the Monetary and Financial integration sub-index. They add these ’new’ EU and EMU dummies to the preferred regression8 _{of their}

earlier paper (Baldwin and Taglioni (2006)):

ln X_ijt= α ln y_ity_jt+

n

X

k=1

β_kX_ijkt+ γEM Ut+ ξEUt+ φijt + ηit+ γjt+ ijt, (18)

where X_ijt is a unidirectional nominal trade flow, y_ity_jt is the product of nominal GDPs, EMU and EU are the modified ’new’ integration variables, X_ijkt is a set of n conditioning variables that are typically found to affect bilateral trade flows (distance, common language, contiguity, landlocked),φ_ijtis the country-pair specific trend,η_itandγ_jt are the unrestricted country-specific trends. According to the results with the 1990-2006 sample, the euro effect is 2% and significant. However the coefficient of the Single Market proxy is quite small, which shows that these proxies could be still improved further.

After the need for including country(pair)-specific time-varying variables became a general con-sensus in the literature, the focus of the papers shifted to other aspects of estimating the gravity equation.

Santos et al. (2010) examine the issue of log-linearization based on their earlier paper (Santos et al. 2006). Log-linearization of the gravity equation is problematic for two reasons. On the one hand whenever the variance of the pair fixed effect (ηij) depends on regressors (e.g. sizes of the

countries captured by GDP), the conditional expectation of ln η_ij also depends on the regressors, which violates the condition for consistency of OLS. On the other hand when the specification is log-linearized countries with zero trade flows have to be dropped, thus the sample is reduced. Both problems can be addressed by estimating the gravity equation in multiplicative form by the pseudo-maximum likelihood (PML estimator):

7

The total impact of EU membership is obtained by adding the coefficients of the institutional measure, the EU and the FTA dummies.

8

(15)

X_ijt= exp(ξ_it+ µ_jt+ δ ln dist_ij+ (1 − σ)βZ_ij+ λCU_ij+ φEU RO12_ij)_ij (19)

_ij = exp((1 − σ)u_ij), (20)

whereξ_itandµ_jt are time-varying exporter and importer fixed effects,dist_ij is distance computed by using the great circle distance algorithm9_, _Z

ij is a vector whose elements are dummies for

common language, colonial ties, contiguity and sharing an FTA. The CU dummy is 1 only for those years when both countries are members of the eurozone, while theEU RO12dummy equals 1 irrespective of the year if both countries are in the eurozone.

The authors check whether the model specifications are adequate by using the RESET test10. While the coefficient on the EU dummy is small and not significant over the three samples, the EU RO12dummy point estimates are between 0.22-0.50% and significant. As a robustness check, the authors exclude the CU dummy and interact theEU RO12dummy with dummies for different years, which yield similar results. Therefore the authors conclude that after controlling for the fact that the eurozone countries traded intensively in the past, the magnitude of the euro effect is small.

Although the methodologies are different, the results of Santos et al. (2010) reinforce those of Berger and Nitsch (2005), Bun and Klaassen (2007) and Baldwin and Taglioni (2008) on the importance of accounting for heterogeneous time trends in the integration process when the euro’s trade effect is estimated with a static fixed effects model.

3.1.2 Dynamic panel estimations

The second group of papers are those carrying out dynamic estimations. As seen earlier when the gravity equation is estimated with only time and pair fixed effects, the omitted time-varying variables show up in the residual (Anderson and van Wincoop misinterpretation). The dynamic estimations consider the serial correlation of the residuals as an evidence of the dynamic nature of trade. Theoretically trade is persistent because of the so-called ’beachhead costs’: sunk market-entry costs resulting from the need for setting up a distribution and service network in the partner country by the exporters. In order to control for the persistency of trade which causes the serial correlation in the residuals, these papers use different dynamic panel techniques (e.g. ADL model estimated by LSDV, DOLS, GMM and Arellano Bond estimators).

9_{The authors allow the distance to have a more flexible effect by dividing it into its four quartiles and the log}

of disrtance is interacted with a dummy for each quartile.

10_{This test also checks whether the conditional variance of Tij is proportional to the conditional mean of Tij,}

(16)

Table 2: Dynamic panel estimations

Authors Methodology Sample Main results

Bun and Klaassen (2002)

LSDV, dep. var: total real trade 19 OECD, 1988-2001

Euro effect: 38.8% (long run effect), half-time in 2006 Faruquee

(2004) DOLS, dep. var: total real trade

22 OECD, 1992-2002 Euro effect 7-8% De Nardis and Vicarelli (2007) Blundell-Bond System GMM, dep. var: total real trade

23 OECD,

1988-2003 Euro effect 4-5%

Bun and Klaassen (2002) estimate an autoregressive distributed lag (ADL) model to capture the euro’s trade effect, based on three main theoretical considerations.

• The applied theoretical framework is the two-country model of Goldstein and Khan (1985) which implies that domestic exports and foreign goods are imperfect substitutes. Therefore real exports depend on real foreign income and the bilateral exchange rate.

• The traders are forward looking. At time t-r the expected value and the variance of the real exchange rate at time t (E_t−r{RERijt} = RERij,t−r and Vt−r{RERijt} = RERV OLij,t−r

respectively) as well as the expected orders betweent − rand tare determinants of the real bilateral trade (EXP ORT_ijt) at timet. In the case of the expected orders perfect foresight is assumed and it is represented in the model by GDP_j,t−q (q < r), the real income of the foreign country.

• Trade is inherently dynamic. Setting up distribution networks is costly (sunk costs) and foreign customers have got accustomed to domestic goods (habit formation). Therefore lags of bilateral export flows are also included.

The specification of Bun and Klaassen (2002) is the following:

EXP ORTijt= α + 2

X

p=1

γpEXP ORTij,t−p+ 2 X q=1 β1qGDPj,t−q+ 2 X r=1

β2rRERij,t−r+ β3rRERV OLij,t−r

+ δ₁EM U_ijt+ δ₂F T AEU R_ijt + δ₃F T AAm_ijt + δ₄CU_ijt+ η_ij+ λt+ ijt, (21)

assuming that there is a stable dynamic relationship between the dependent variable and the regressors (γ1+ γ2 < 1). Country-pair and time fixed effects (ηij andλt) are used because applying

random effects would require that regressors are exogenous, however country-pair fixed effects are correlated with lagged exports and time fixed effects are correlated with GDP.

The specification is estimated by LSDV, which is inconsistent when T is finite, but as T here is sufficiently large (1965-2001) the bias is small. The two channels through which euro affects trade in this framework are the real exchange rate volatility and the additional effects of EMU (e.g. decrease in transaction costs, integration of capital markets) captured by the EMU dummy.

(17)

Beside the standard estimation results the authors include the economically more meaningful cumulative percentage changes in exports caused by both channels as well as long-run and half-time estimates. According to the results, the long-run effect of the real exchange rate channel is rather small (-0.7%), while the additional effect of the euro is large (38.8%). The standard errors are large, because there are only a few years of the EMU period in the dataset, it has been tested that including additional years decreases the standard errors. The gains from EMU may be realised quite fast as the estimated half-time is 200611_.

While Bun and Klaassen (2002) account for the dynamic nature of trade by estimating an ADL model, Faruquee (2004) takes into account also the endogeneity of the income variables12and the non-stationarity of the gravity variables.

First Faruquee (2004) determines the order of integration of the trade and income variables by using different panel unit root tests. As a second step panel cointegration tests are used in order to test whether the linear combination of the integrated variables is stationary. Finally in order to obtain valid statistical inference with non-stationary panel data, the gravity equation is estimated by DOLS that corrects for the endogeneity bias and also the serial correlation of the residuals. Faruquee (2004) estimates the following specification as a cross-check for results with standard OLS:

ln(T radeijt) = γij+ τt+ βZijt+ m

X

k=−m

γ_k∆Z_ijt+k+ αF T Aijt+ δtEUijt+ λtEM Uijt+ ijt, (22)

where γ_ij and τ_t are country-pair and time fixed effects, Z_ijt refers to the vector of integrated regressors consisting of partners GDP and GDP per capita,F T A,EU andEM Uare the integration dummies. The time-varying coefficients of theEU and EM U variables indicate that the author -similarly to others mentioned earlier - views both the EU and the EMU as an integration process and tries to account for the varying degree of economic and monetary integration. Both terms can be viewed as an alternative to the sum of the given dummy with constant coefficient and a group specific trend. According to the estimation results, the EMU raised intra-area trade by 7-8% relative to trade among industrial countries.

De Nardis and Vicarelli (2007) account for the dynamic nature of trade in a third way. The authors emphasize that the dynamic nature of trade is related to investment in export-oriented infrastructure and invisible assets such as political, cultural and geographical factors, which result in sticky behaviour of exporting firms.

The authors estimate the following equation by Blundell-Bond (1998) system GMM:

ln Exp_ijt= b₁ln(Exp_ijt−n)+b₂ln(SumV A_ijt)+b₃ln Dist_ij+b₄vol_ijt+b₅euro_ijt+b₆EU_ijt+b₇α_i+b₈β_j+b₉τ (23)

11_{This implies that 2006 is the first year when the cumulative trade effect is at least half of the long-run effect.}

12

According to the endogenous OCA theory (e.g. Frankel and Rose (2000) sharing a common currency may enhance processes increasing the integration of countries, which implies that trade may have an impact on income, while the gravity equation captures the causality in the contrary direction.

(18)

whereExp_ijt is the volume of exports from country ito country j,SumV A_ijt is the sum of value added of the exporting and importing countries,Dist_ij is bilateral distance between capital cities, volijtis the nominal exchange rate volatility, euroijtand EUijt are the regular dummies for EMU

and EU membership, α_i is the exporting country dummy, β_j is the importing country dummy and τ is the year dummy.

Estimating the above equation by the Blundell-Bond system GMM implies that the authors run the Hansen two-step GMM estimator on the differenced model using lagged levels as instruments and the level model using lagged differences as instruments. The authors prefer to include the level model as well, because in the case of the Arellano-Bond estimator – which includes only the differenced model – the lagged level of time series with near unit root properties are weak instruments for the first differences. An additional argument is that including only the first differenced model implies that differencing removes not only the fixed effects but also the time invariant regressors.

The authors also examine the serial correlation in the disturbance of the first differenced equation. According to the AR(1) and AR(2) test results the first differenced equation contains first-order, but no second-order serial correlation of the residuals, therefore OLS would be inconsistent and GMM is a consistent estimator. The Hansen test shows that all moment restrictions are satisfied for the dynamic specifications.

De Nardis and Vicarelli (2007) using the above methodology estimates a euro effect of 4%, which is in line with the dynamic estimate of Faruquee (2004) and its magnitude is also similar to those static estimates that account for time-heterogeneity of the country(pair)-specific variables.

3.1.3 New generation panel techniques

The third group of studies are the most recent ones and they apply new generation panel tech-niques in order to account for cross-sectional correlation between the elements of the panel. As Camarero et al. (2012) points out cross-sectional dependence may be present in the gravity model for several reasons: other economies’ business cycles may influence foreign trade, the gravity model includes bilateral trade flows together with aggregate national variables and finally the model also implies spatial dependence in the data due to the hypothesized effect of distance on trade. In addition to cross-sectional dependence, the methodology applied by Camarero et al. (2012) allows also for structural breaks.

(19)

Table 3: New generation panel techniques Authors Methodology Sample (country, period) Main results Gengenbach (2010)

CUP, CCEP, CCEP for ECM estimator, dep. var: total real

trade

19 OECD, 1967-2002

Euro effect ranges between 1%-6.3%

Camarero et al. (2012)

CUP estimator, allowing for structural breaks in the time domain, dep. var: total nominal

trade

26 OECD, 1967-2008

Euro effect between -6.8 and -8.6%.

Gengenbach (2010) first tests the non-stationarity of the series by applying several unit root tests. The Bai and Ng (2004) unit root test is reviewed here, which assumes the following general model:

Y_i,t= γ_i,tFt+ Ei,t, (24)

where Ft is a k-vector common factor and Ei,t is an idiosyncratic component. In order to test

for unit roots the first difference of model (24) is taken and the method of principal components is applied. In the case of the idiosyncratic components, individual unit roots (ADF-test) and pooled unit roots (Fisher-type test) are tested as well. Regarding the common factors, ADF test is used if a single common factor is estimated. If more than one common factors are estimated, an iterative procedure similar to the Johansen test of cointegration is used to select the number of independent stochastic trends (H₀ : k₁ = m against H₁ : k₁ < m is tested for m starting from ˆ

k). The majority of the test statistics accept the null of unit root, therefore the gravity variables prove to be non-stationary.

As a second step, the author tests for panel cointegration with two of his earlier methods, the test based on Gengenbach (2006) is reviewed here. For each cross-sectional unit a 1 + mvector Z_it= (Y_it, X_it0 )0 is observed and bothY_itandX_ithas a Bai and Ng (2004) type factor representation similar to model (24). Cointegration between the estimated idiosyncratic components Eˆ_i,tY and

ˆ

E_i,tX is tested using the panel no cointegration test of Pedroni (1999). Cointegration between the estimated common factors _FˆX

t and FˆtY is tested applying the Johansen trace test. The majority

of the test statistics reject the null of no cointegration, therefore trade, GDP and GDP per capita are cointegrated.

Gengenbach (2010) estimates the same gravity specification (on the same database) as Bun and Klaassen (2007):

T RADE_ijt= β₁GDP_ijt+ β₂GDP CAP_ijt+ δ₁EU RO_ijt+ δ₂F T A_ijt+ η_ij+ τ_ijt + λ_t+ _ijt, (25)

The variables entering the gravity equation prove to be non-stationary and cointegrated, therefore equation (25) describes a long-run equilibrium relationship between trade, GDP and GDP per

(20)

capita. The parameters of the above static long-run model are estimated by the continuously updated (CUP) estimator of Bai et al. (2009) and the CCEP estimator of Pesaran (2006). The author also estimate a dynamic error correction model (ECM) with a CCEP estimator. Here only the CUP estimator is reviewed, as later it is suggested for the estimation of a model which takes into account structural breaks in the deterministic components. Structural breaks are relevant in the case of the current analysis because the updated period includes the 2008-09 crisis.

The CUP estimator of Bai et al. (2009) is an iterative procedure for estimating the cointegrating vector in a cointegrated panel data model with non-stationary and unobserved common factors.

Y_i,t= βX_i,t+ γ_iFt+ ei,t, (26)

where Ft is a k-vector of common factors, γi is the corresponding vector of factor loadings and

e_i,t is an idiosyncratic error term. Within the iterative procedureβ is estimated by least-squares given Fˆt (an initial estimate of Ft) and then Ft is estimated given βˆ. The iteration through the

CUP estimator goes until convergence. The objective function of the procedure is:

min S_{N T}(β, F, Γ) = N X i=1 T X t=1 (Y_i,t− βX_i,t− γ_iFt)2 (27) s.t. T−2 T X t=1 FtFt0= Ik, (28)

where ΓΓ0 is positive definite and Γ = (γ₁0, . . . , γ_n0). The procedure requires that the number of common factors is known; therefore it is estimated using theBIC3 criterion of Bai and Ng (2002).

Then the fixed effects are concentrated out of model (25) and the long-run parametersβ₁ andβ₂ as well as the coefficients of the two dummy variables EU RO_ijt and F T A_ijt are estimated which minimize (27).

The static and dynamic estimates of Gengenbach (2010) are compared to those of Bun and Klaassen (2007) and (2002) respectively. The results prove to be similar. There is a strong drop in the estimated coefficients when country-pair specific trends are included and the euro has a positive effect, which is rather small ranging between 1% and 6.3%.

The methodology of Camarero et al. (2012) is similar to Gengenbach (2010), but an additional feature of the study that it accounts for structural breaks in the time domain. They use a dataset containing annual data from 26 OECD countries that covers the period 1967-200813.

First the authors test for cross-sectional dependence by using the test of Pesaran (2004) which is robust to structural breaks. The test statistic strongly rejects the null-hypothesis of independence; therefore they conclude that cross-section dependence has to be considered when computing the panel data statistics. As a next step, similarly to Gengenbach (2010), the Pesaran (2007) and Bai and Ng (2004) panel unit root tests are applied and all variables prove to be non-stationary.

13

However their data covers only the period 1967-2008. ’Although the number of years available was higher, we have opted by restrict our sample to this period to exclude the effects of the financial crisis that started in 2008’ (p. 12).

(21)

As the gravity variables proved to be non-stationary, cointegration is tested by using the method-ology of Banerjee and Carron-i-Silvestre (2010), which similarly to Gengenbach (2006) tackle cross-section dependence by using factor models. However this test also allows for structural breaks both in the deterministic component and the cointegrating vector. The model is the following: y_it= D_it+ x0_itδ_it+ u_it (29) u_it= γ_iF_t+ e_it (30) D_it= µ_i+ β_it + mi X j=1 Θ_ijDU_ijt+ mi X j=1 γ_ijDT_ijt, (31)

where Y_it = (y_it, x0_it) is an (m × 1) vector of non-stationary stochastic processes whose elements are individually I(1). F_t is ak-vector of common factors,γ_i is the corresponding vector of factor loadings ande_i,t is an idiosyncratic error term. D_it is the deterministic term whereDU_ijt= 1 and DT_ijt = (t − T_itb) for t T_i,tb and 0 otherwise. T_itb = λb_i,jT denotes the timing of the j-th break, j = 1, . . . , mi for the i-th unit14 i = 1, . . . , N,λb_i,jT ∈ Λ, beingΛ a closed subset of (0, 1).

The authors test six different model specifications: either a constant or a changing cointegrating vector (δ_it) is combined with the cases when there is no linear trend (β_i = γ_ij = 0), there is a stable trend (β_i6= 0,γ_ij = 0) and there are changes in the trend (β_i6= 0,γ_ij 6= 0). For each model the first differences are computed, then orthogonal projections are taken and the common factors and factor loadings are estimated using principal components following the methodology of Bai and Ng (2004).

In the above specifications Banerjee and Carrion-i-Silvestre (2010) recover the idiosyncratic dis-turbance terms (e˜_i,t) through cumulation of the estimated residuals and H₀ : no cointegration against H₁ : cointegration with a break is tested using an ADF statistic. The null hypothesis of non-cointegration can be rejected. Using the BIC information criterion, the model containing a constant, a trend and a structural break that affects them simultaneously proved to be the best specification (without change in the cointegrating vector), if the variables are constructed following Baldwin’s critiques. The time of the break is found in 1989.

The authors filter the gravity variables of the deterministic components specified above. Then similarly to Gengenbach (2010), the long-run parameters and the coefficients of the dummy variables are estimated which minimize the objective function of the CUP estimator. The euro effect is small, which is in line with those papers that account for trends in the integration process (e.g. Berger and Nitsch (2005), Bun and Klaassen (2007), Gengenbach (2010)).

On the whole both papers using new generation panel techniques draw a similar conclusion as other strands of the literature. The increase of trade within the euro area may be a continuation of a long-run trend, probably linked to the broader set of EU’s economic integration policies and institutional changes.

14

(22)

3.2 Panel studies using sectoral data

While the focus of the current analysis is on estimations using aggregate data, a large number of papers use sectoral data. Using micro data makes the estimates more precise and it allows distinguishing effects on different types of goods. An example for the latter is the study of Flam and Nordstrom (2003), which concludes that the euro significantly increased the trade of goods that are relatively differentiated and processed through the reduction of the costs of vertical specialization.

Micro data also provides an opportunity to separate the euro’s effects on the intensive and the extensive margin of trade, for instance Flam and Nordstrom (2006) concludes that changes in the extensive margin are estimated to be proportionally greater than effects on the intensive margin. This finding for the eurozone as a whole is reinforced by studies focusing on a given country and using more detailed micro data (De Nardis et al. (2008) for Italy, Berthou and Fontagn (2008) for France).

The finding that a substantial portion of the euro trade effect operates at the extensive margin is reconfirmed by Bergin and Lin (2010), however they have a different approach to the mechanism through which this effect operates. By estimating panel regressions, the authors find that the extensive margin began to rise four years before the introduction of the euro, but its effect was mitigated by the negative intensive margin until about one year before the introduction of the euro, when the overall trade effect also became positive. While earlier studies assumed that trade is sluggish, here the authors argue that the firms entered the market because they anticipated the future trade opportunities created by the EMU. The authors simulate this mechanism with a DSGE trade model, which contains forward looking behaviour of firms in response to news about the future.

(23)

4. Empirical analysis

The literature review highlighted that static panel estimations focus on dealing with unobserved time-varying heterogeneity, while dynamic panel estimations consider the serial correlation of the residuals as an evidence of the dynamic nature of trade. Recent papers focus on the non-stationary nature of data and they estimate such static models, which already account for unobserved time-varying heterogeneity by using new generation panel techniques. Both papers reviewed here (Gengenbach (2010) and Camarero et al. (2012)) use the specification of Bun and Klaassen (2007), which accounts for unobserved time-varying heterogeneity by including country-pair spe-cific trends. As the aim of the current analysis is to examine the aggregate trade effect of the euro when recent developments are taken into account, estimating the specification of Bun and Klaassen (2007) using an updated dataset appears to be a suitable choice.

4.1 Data

The construction of the dataset and the sources of the data are the same as in Bun and Klaassen (2007). Data on 19 countries are included, namely all EU countries prior to the May 2004 expansion, Norway, Switzerland, Canada, Japan and the US. Data is combined for Belgium and Luxembourg because prior to 1997 data is available only for the union of the two countries (BLEU). This gives 171 country-pairs and the data covers the period 1967-2012. The total number of observations is 7866, which means 1710 additional observations compared to the original dataset for the period 1967-2002.

Using the definitions given by Bun and Klaassen (2007), the following variables are constructed. T RADE_ijt is the log of real bilateral trade between countries i and j at time t, where real bilateral trade is measured as the average of nominal exports and imports in millions of U.S. dollars deflated by the U.S. producer price index. The source of trade data is the IMF Direction of Trade Statistics (DOTS) and the US producer price index comes from the OECD Main Economic Indicators. GDP_ijt is the log of product of countries’ real GDP, where real GDP is in millions of U.S. dollars divided by the U.S. producer price index. Data on GDP comes from the OECD Economic Outlook. GDP CAP_ijt measures the log of product of countries’ real GDP per capita. Population data are from the US Bureau of Census website. Two dummies are also constructed. EU RO_ijt is1 if both countries have adopted the euro at time t and F T A_ijt is 1 if both countries have a free trade agreement at time t.

4.2 Model specification

(24)

T RADE_ijt= β1GDPijt+ β2GDP CAPijt+ δ1EU ROijt+ δ2F T Aijt+ ηij + τijt + λt+ ijt (32)

where_ijtis allowed to be heteroskedastic (across country-pairs and time), serially correlated and cross-sectionally correlated. As mentioned earlier serial correlation and cross-sectional correlation of the standard errors are relevant in the case of the gravity model. Serial correlation of the stan-dard errors may come from sunk costs and habit formation, while cross-sectional correlation may stem from regional trade shocks or such nation-specific shocks that affect trading partners as well. The widely used White standard errors are robust only for arbitrary heteroskedasticity over time and country-pairs, therefore Bun and Klaassen (2007) reports Driscoll-Kraay-Newey-West stan-dard errors as well, which are robust for serial correlation as well as lagged and contemporaneous cross-sectional correlation.

It is also assumed that ijt is stationary and all regressors are treated as strictly exogenous

with respect to _ijt. The sensitivity analysis of Bun and Klaassen (2007) presents evidence that treating cointegration and endogeneity does not alter significantly the euro’s trade effect, therefore these assumptions are not important regarding the main conclusion of their study. In the current analysis the cointegration estimations are carried out and in the case of the euro effect no significant difference is found.

In most cases the gravity equation with fixed effects is estimated by the standard within trans-formation. However as Bun and Klaassen (2007) points out the standard within transformation - which subtracts country-pair specific means over time from each variable - does not wipe out country-pair specific trends. Therefore here an extended within transformation is applied. Fol-lowing Bun and Klaassen (2007) the so-called projection argument is used in which all variables in the model are projected on the null-space of the matrix of dummy/time variables corresponding to all µ_ij,τ_ijt and λ_t.

4.3 Estimation results

The aim of the empirical analysis is to examine whether the aggregate trade effect of the euro changes when recent developments are taken into account. The empirical strategy of Bun and Klaassen (2007) is applied, using an updated dataset. The results are compared with those of the original paper, therefore two datasets are used: the downloadable dataset of Bun and Klassen (2007) and an updated dataset which is constructed along the lines of the original. At most four alternatives are reported: the original results from Bun and Klassen (2007) (denoted by ’B-K’), own estimations with the Bun and Klaassen (2007) dataset (denoted by ’own (B-K data)’), finally estimations with the updated dataset (’own data’) for both the original (1967-2002) and the updated (1967-2012) period. The empirical analysis is especially interesting because the updated sample includes several events (e.g. the entry of ten new members to the EU in 2004, the global recession of 2008-09), which may alter the original estimation results.

(25)

4.3.1 Estimation without country-pair specific trends

First a specification with only country-pair and time fixed effects is considered (Table 4). Time fixed effects control for a common trend for all country-pairs (e.g. global economic shocks), but do not account for time-heterogeneity of the country-pair specific variables.

Along with the coefficient estimates, two types of standard errors are reported in the tables, i.e. White and Driscoll-Kraay standard errors. As mentioned earlier, in the case of gravity models, cross-sectional dependence is often present. In this case the standard error estimates of commonly applied covariance matrix estimation techniques (e.g. OLS, White, Rogers, clustered standard errors) are biased. By relying on large-T asymptotics, Driscoll and Kraay demonstrate that the standard non-parametric time-series covariance matrix estimator can be modified such that it is robust to general forms of cross-sectional as well as temporal dependence (Hoechle (2007))15. The robust Driscoll-Kraay standard errors are in all cases larger than the White standard errors (Table 4); therefore cross-sectional dependence may be present in the datasets.

Table 4: Estimation results for trade model (32) without country-pair specific trends

No trends

1967-2002 1967-2012

B-K own (B-K data) own data own data

EU RO_ijt 0.410 0.413 0.405 0.377 {0.028} {0.029} {0.028} {0.015} (0.075) (0.069) (0.070) (0.067) F T A_ijt 0.41 0.39 0.40 0.44 {0.02} {0.02} {0.02} {0.02} (0.09) (0.09) (0.09) (0.09) GDP_ijt 1.41 1.35 1.48 1.72 {0.10} {0.10} {0.10} {0.08} (0.39) (0.37) (0.37) (0.21) GDP CAPijt -0.68 -0.63 -0.75 -1.03 {0.09} {0.10} {0.09} {0.07} (0.37) (0.35) (0.34) (0.20) No. observations 6156 6156 6156 7866

No. fixed effects 206 206 206 216

White standard errors are in braces, Driscoll-Kraay standard errors in parentheses.

The coefficients of the euro dummy are quite stable across datasets and samples. The euro effect corresponds to a relative change of trade between 46% and 51% respectively16.

Bun and Klaassen (2007) argue that if trends with cross-sectional variation are omitted, the euro effect estimates may pick up their effect. In order to support this idea, they estimate the benchmark specification including only country-pair and time fixed effects both without and with the euro dummy and plot the residuals. The residual series are averaged separately for the 55 country-pairs involving two euro countries and for the other 116 country-pairs.

15

The procedure consists of applying a Newey-West type correction to the sequence of cross-sectional averages of the moment conditions.

16

(26)

Figure 1 shows that the residuals for the euro country-pairs have an upward trend until 2002 and stagnate afterwards. This would suggest that omitted variables increasingly explained trade growth in the case of the euro country-pairs mainly during the 1990s and their effect stopped to increase further afterwards. When the euro dummy is included in the specification it takes up this effect from 1999 onwards.

Figure 1: Residuals of trade model (32) without country-pair specific trends

EUROijt not in model (32) EUROijt in model (32)

-0.2 -0.1 0 0.1 0.2 0.3 1967 1969 1971 1973 1975 1977 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011

euro country-pairs B-K euro country-pairs own other country-pairs B-K other country-pairs own

-0.2 -0.1 0 0.1 0.2 0.3 1967 1969 1971 1973 1975 1977 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011

4.3.2 Estimation with country-pair specific trends

The problem of the potential omitted variables is treated in different ways in the literature. Some authors include indices of institutional integration (e.g. Berger and Nitsch (2005), Baldwin (2008)) or time-varying exporter and importer fixed effects (e.g. Baltagi et al. (2003)). Bun and Klaassen (2007) introduce country-pair specific time trends mainly for two reasons. On the one hand general trends are already accounted for by including trending variables (GDP) and time fixed effects. On the other hand trade data is country-pair oriented, which supports the inclusion of country-pair specific trends rather than country-specific ones.

The specification of Bun and Klaassen (2007) including country-pair specific trends is first esti-mated for the same period (1967-2002), with similar results (Table 5). As a result of accounting for the omitted time-varying variables, the coefficients of the main variables (euro, FTA, GDP) decrease compared to the specification without country-pair specific trends. The euro effect is lower when estimated with the new dataset compared to the results with the B-K dataset and it becomes insignificant as well. However, the volatility of the euro estimates across samples may be reasonable. The euro effect is estimated using data only for three years (1999-2002); therefore data revisions may have a major effect on both the magnitude and the significance of the esti-mated coefficients. The effect of GDP per capita is insignificant in the case of both the B-K and the new datasets.

When the specification including country-pair specific trends is estimated for the updated period (1967-2012) the euro effect decreases further and it remains insignificant. The Driscoll-Kraay standard errors of the euro coefficient are considerably larger in the case of the updated period than for the shorter sample. The magnitude and the sign of the coefficients of the GDP and

(27)

GDPCAP variables change substantially, while their standard errors increase and the coefficient of GDP becomes insignificant.

Table 5: Estimation results for trade model (32) with country-pair specific trends

Trends

1967-2002 1967-2012

B-K own (B-K data) own data own data EU RO_ijt 0.032 0.030 0.006 -0.065 {0.014} {0.015} {0.015} {0.014} (0.016) (0.017) (0.016) (0.047) F T A_ijt 0.06 0.04 0.04 0.09 {0.01} {0.01} {0.01} {0.01} (0.03) (0.03) (0.03) (0.04) GDP_ijt 0.70 0.68 0.65 0.08 {0.15} {0.15} {0.16} {0.14} (0.36) (0.43) (0.44) (0.55) GDP CAP_ijt -0.23 -0.22 -0.19 0.53 {0.15} {0.15} {0.15} {0.13} (0.35) (0.41) (0.43) (0.54) No. observations 6156 6156 6156 7866

The pattern of the residuals also seems to change in the case of the updated period (Figure 2). Up to the end of the 1990s, the residuals are similar to those of Bun and Klaassen (2007), the inclusion of the country-pair specific trends wipes out trends in the residuals. However during the 2000s there are trends left in the residuals.

Figure 2: Residuals of trade model (32) with country-pair specific trends

EURO_ijt not in model (32) EURO_ijt in model (32)

-0.2 -0.1 0 0.1 0.2 0.3 1967 1969 1971 1973 1975 1977 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011

(28)

The downward trend of the residuals at the end of the sample may suggest that the linear incidental trends of model (32) do not fully capture all time-varying unobserved heterogeneity. As it will be seen in the next section, one of the sensitivity analyses adds non-linear (quadratic) trends to model (32). The latter specification is estimated already here in order to cross-check both the coefficient estimates and the residuals when possible non-linearity is accounted for.

Table 6: Adding quadratic country-pair specific trends and omitting GDP CAP from model (32)

Full specification GDP CAP omitted Linear t. Quadratic t. Quadratic t.

1967-2012 1967-2002 1967-2012 EU RO_ijt -0.065 0.050 -0.018 0.048 {0.014} {0.016} {0.015} {0.016} (0.047) (0.030) (0.026) (0.042) F T A_ijt 0.09 0.07 0.07 0.05 {0.01} {0.01} {0.01} {0.01} (0.04) (0.03) (0.02) (0.03) GDP_ijt 0.08 -1.44 0.50 0.45 {0.14} {0.12} {0.02} {0.02} (0.55) (0.17) (0.03) (0.03) GDP CAP_ijt 0.53 1.86 - -{0.13} {0.12} - -(0.54) (0.16) - -No. observations 7866 7866 6156 7866

The euro effect is positive for the updated period when quadratic trends are added to model (32) (Table 6). The coefficient of the euro dummy is now closer to being singificant than in the case of model (32) 17_{. The income coefficients estimated for the updated period differ from those of}

Bun and Klaassen (2007) both in the case of model (32) and when quadratic trends are added. In the gravity framework GDP measures countries’ economic size, while GDP per capita is gen-erally considered (e.g. Bergstrand (1989)) as a proxy for the purchasing power of the importing country and as a proxy for the factor endowment (capital/labour ratio) of the exporting country. As a sensitivity check GDPCAP is left out of the model (Table (6)), however the results are similar when GDP is left out. Dropping one of the income variables is reasonable, because the high variance inflation factors (VIFs) of the income variables indicate a high degree of collinearity. Although in the presence of country-pair specific trends, the identification of such country-specific variables as GDP is possible, a high degree of collinearity may occur. When GDPCAP is dropped, the magnitude of the euro (and the FTA) effect practically does not change. The coefficient of GDP is now more homogenous across samples and the VIF of GDP is substantially lower, although it is still above the common cut-off criterion of 10 (Kutner et al. (2004)).

The question arises that how does adding quadratic trends to model (32) influence the residuals. Figure 3 shows the residuals of trade model (32) both when quadratic trends are not added (left

17_{The p-value of the euro coefficient (using Driscoll-Kraay standard errors) is now 0.107, while it is 0.171 in the}

The euro effect on trade.

The euro effect on trade

Contents

1.

Introduction

2.

Theoretical foundations of the trade gravity model

3.

Empirical literature

3.1

Panel studies using aggregate data

3.2

Panel studies using sectoral data

4.

Empirical analysis

4.1

Data

4.2

Model specification

4.3

Estimation results