Determinants of the dynamics of the European Union integration process: An ordered logit approach

Transcription

1 Unversty Jaume I From the SelectedWorks of Inma Martnez-Zarzoso 2009 Determnants of the dynamcs of the European Unon ntegraton process: An ordered logt approach Inma Martnez-Zarzoso Laura Marquez Celes Suarez Avalable at:

2 Determnants of the dynamcs of the European Unon ntegraton process: An ordered logt approach LAURA MÁRQUEZ-RAMOS* Department of Economcs and Insttute of Internatonal Economcs, Unverstat Jaume I Campus del Ru Sec, 207 Castellón (Span) lmarquez@eco.uj.es INMACULADA MARTÍNEZ-ZARZOSO * Department of Economcs and Insttute of Internatonal Economcs, Unverstat Jaume I Campus del Ru Sec, 207 Castellón (Span) Ibero-Amerca Insttute for Economc Research, Unverstät Göttngen Platz der Goettgen Seben 3, Goettngen (Germany) martne@eco.uj.es CELESTINO SUÁREZ-BURGUET * Department of Economcs and Insttute of Internatonal Economcs, Unverstat Jaume I Campus del Ru Sec, 207 Castellón (Span) celes@eco.uj.es Correspondng author: Laura Márquez-Ramos. Departamento de Economía, Unverstat Jaume I, Campus del Ru Sec, 207 Castellón (Span). Fax: (+34) E-mal address: lmarquez@eco.uj.es

3 Determnants of the dynamcs of the European Unon ntegraton process: An ordered logt approach Abstract Ths research has three man ams: frstly, to emprcally analyse the determnants of dfferent levels of ntegraton by re-examnng the evdence presented by Baer and Bergstrand (2004) n the JIE 64 (); secondly, to analyse the mportance of addtonal factors, n partcular soco-poltcal factors. Fnally, to analyse the dynamcs of the European Unon ntegraton process. The results show that although economc and geographcal factors are the most mportant explanatory factors for the probablty of regonal ntegraton agreement formaton or enhancement, soco-poltcal varables also contrbute to explan the formaton of regonal ntegraton agreements. Democraces and countres wth a hgher level of economc freedom are more lkely to form or enhance RIAs. Keywords: Regonal ntegraton agreements, European Unon, dscrete choce models, trade flows, soco-poltcal factors, natural partners. JEL classfcaton: F, F2, F5. Introducton A major concern n the tradtonal lterature on the formaton of free trade areas (FTAs) has been whether these areas generate welfare gans for the ndvdual countres that engage n these processes. Snce the 950s (Vner, 950), many authors have contrbuted to ths debate, especally n the 990s when studes based on the gravty model prolferated (Frankel, Sten and We FSW-, 995, 996, 998). However, none of ths research has attempted to evaluate the determnants of FTA formaton. Only recently have Baer and Bergstrand (2004) developed the frst theoretcal and emprcal analyss of the economc determnants of FTA formaton. They provde an 2

4 economc benchmark for future poltcal economy models to explan the determnants of FTAs. They fnd evdence showng that pars of countres wll be more lkely to form FTAs f they share the followng characterstcs: a) they are geographcally close to each other, b) they are remote from the rest of the world, c) they are large and of a smlar economc sze, d) the dfference of captal-labour between them s large and e) the dfference of ther captal-labour ratos s small compared to the rest of the world. Baer and Bergstrand (BB) only consder whether or not each par of countres s nvolved n an FTA. Therefore the varable they attempt to explan s bnary and takes the values zero and one. BB (2005) show the mportance of treatng FTAs as endogenous when the determnants of trade flows are analysed. They show that when the endogenety of the FTA varables s taken nto account n gravty models, ther effect on trade flows s quntupled. In ths paper, we extend BB s work n two ways: frstly, we address the mportance of addtonal economc, geographcal and soco-poltcal varables as determnants of regonal ntegraton agreements (RIAs). Secondly, we nvestgate the determnants of fve dfferent levels of ntegraton between pars of countres: Preferental trade agreement (PTA), free trade agreement (FTA), customs unon (CU), sngle market (SM) and monetary unon (MU). We begn by replcatng BB s emprcal work to verfy the robustness of ther results wth an alternatve data set and by addng soco-poltcal varables to the model. We then estmate an ordered logt model (nstead of a bnary probt) wth the same explanatory varables consdered by BB to benchmark our extenson to ther orgnal work. Fnally, the ordered logt s estmated wth addtonal economc, geographcal and soco-poltcal varables. The economc varables we consder are economc sze, ncome dfferences and factor endowment dfferences. Adjacency and landlocked status 3

5 are added to BB s lst of geographcal varables. The soco-poltcal varables are a shared language, poltcal regme, level of economc freedom and trade barrers. We fnd that: () BB s results are farly robust, although the coeffcent sgns are reversed for the K-L dfference varable wth our database; () the addtonal characterstcs consdered have a sgnfcant mpact on the probablty of an RIA beng formed; () soco-poltcal factors are less mportant than economc and geographcal factors, but stll sgnfcant n explanng RIA formaton or enhancement. To our knowledge, only a few authors have studed the determnants of regonal ntegraton who take nto account the degree of ntegraton. Wu (2004) consdered dfferent levels of ntegraton ranked across countres. However, her paper focuses on the role that poltcal and economc uncertanty plays n explanng RIA formaton and her results are not drectly comparable to Baer and Bergstrand snce she ncludes dfferent explanatory varables n her model. Wu shows that countres per capta ncome, democracy and geographcal characterstcs appear to be the best ndcators of the probablty of partcpaton n a certan level of RIA n the perod Surprsngly, Wu (2004) does not consder the dstance varable as a determnant of RIA formaton. Ths omsson may nfluence the results obtaned for other varables snce the model s not well specfed. Endoh (2006) derved a theoretcal framework to explan the ncentves of countres to conclude an RIA. The author stated that the economc and poltcal characterstcs of determnng the exstence or absence of PTAs are qute dfferent from those of FTAs and CUs. Heterogenety among RIAs s taken nto account n the emprcal analyss, n whch two dfferent dependent varables are consdered (FTAs/CUs based on GATT Artcle XXIV and all the PTAs ncludng other types of agreement based on the Enablng Clause). The methodology used to estmate s Endoh, 2006, page

6 a bnary logt model. Fnally, Vcard (2006) relates economc and poltcal ntegraton, and proves that the determnants of regonal ntegraton dffer accordng to the type of regonal ntegraton agreement. The heterogenety n the nature of RIAs s ntroduced by takng nto account two ntegraton levels: shallow RIAs (PTAs and FTAs) and deep RIAs (CUs and CMs). The author runs three dfferent regressons, one for all RIAs, one for shallow RIAs and one for deep RIAs. Then, a bnary probt model s estmated. Unlke these authors, we take on a more dffcult queston: Why deeper ntegraton? The remander of the paper s structured as follows. In Secton 2 stylsed facts n relaton to the reasons why countres decde to engage n deeper economc ntegraton are dscussed. Secton 3 presents the theoretcal framework and the econometrc model. Secton 4 descrbes the data, the varables and the hypothess to be tested. Secton 5 dscusses the estmaton results. The model n Secton 6 s estmated for an addtonal sample, ncludng data for the EU-27 from 999 to 2007, thus enablng dynamc ssues to be also analysed. Fnally, Secton 7 presents the conclusons. 2. Stylsed facts Decsons concernng economc ntegraton are controversal n most cases; there are global benefts, but they are unevenly dstrbuted among wnners and losers. The best real example of deep economc ntegraton s the European ntegraton process. Although the ntal goal was to avod undesrable wars wthn the contnent, a much more ambtous vson was endorsed over the years, that beng one of the man goals: the completon of the European Monetary Unon. Deep ntegraton of ths form has generated clear benefts to European ctzens n terms of welfare and growth. However, snce the recent accesson of ten new member states n 2004 and two more n 2007, the European Unon (EU) has wtnessed an ntense dscusson regardng ts future. The central queston of the debate s featured n the ttle of the report launched 5

7 by the Constructng Europe Network (EU-CONSENT): Wder Europe, deeper ntegraton? A common theoretcal framework. The man am of the EU-CONSENT s to elaborate the scenaros and strateges for the future of European ntegraton and to evaluate the costs and benefts of each of them, based on the trangle of deepenng, wdenng and completng. Over the years, the EU has been consdered a club wth open membershp, but as ntegraton deepens, the entry condtons become more exhaustve. Although unformty was a rule untl recently, the monetary unon as well as other specfc agreements (Schengen agreement on border controls) were restrcted only to some members. The debate concernng deep ntegraton s also open n North Amerca (Campbell, 2005) and Asa (Wyplosz, 2006). In both cases the expected benefts of deeper ntegraton are only seen as uncertan, whereas the poltcal-costs are hgh. 3. Theoretcal framework and econometrc model 3.. The theory Although deep regonal ntegraton can proceed along dfferent lnes, accordng to McKnnon (979) t should start wth domestc goods market lberalsaton, followed by external trade ntegraton, and should proceed wth domestc fnancal market lberalsaton and nternatonal captal ntegraton. We defne the concept of deeper regonal ntegraton n relaton to the level of economc ntegraton stated by Vner (950). Therefore, deeper RIAs are those nvolvng a hgher level of economc ntegraton. Ths paper s related to recent research n regonal ntegraton that nvestgates why countres enter an RIA, although t also focuses on the queston of why countres engage n deeper ntegraton processes. What are the reasons why countres engage n deeper ntegraton? Untl recently the research n ths feld focused on the effects of regonalsm and dsregarded the economc 6

8 and poltcal factors whch explan the presence or absence of free trade agreements between pars of countres. BB (2004) were the frst authors to theoretcally explan the lkelhood of PTAs between pars of countres usng only economc and geographcal factors. Mansfeld, Mlner and Rosendorff (2002) consdered ths problem from a poltcal-economy pont of vew, and demonstrated that more democratc countres had dsplayed a greater lkelhood of concludng PTAs than other countres. In addton, Endoh (2006) derved a theoretcal framework to explan the ncentves of countres to conclude an RIA. The author stated that the economc and poltcal characterstcs of determnng the exstence or absence of PTAs are qute dfferent from those of FTAs and CUs. The author derves seven testable hypotheses, of whch Hypothess 3 states that the possblty of concludng a PTA by a par of countres ncreases as ther qualty of governance amelorates. Four categores of FTA determnants can be nferred from ths theoretcal framework: economc geography factors, ntra-ndustry trade and nter-ndustry trade determnants and soco-poltcal factors. They wll all be consdered n the emprcal analyss Econometrc model Probt and logt models have often been used to model dscrete choce phenomena (Ben- Akva and Lerman, 985). In ths context, a logt model s a dscrete choce system nterpreted as a partcular case of a model, the dependent varable of whch s subject to lmted varablty, s not contnuous and takes a fnte number of values (McFadden and Tran, 2000; Koppelman and Wen, 998). Ths type of system descrbes the behavour of economc agents n terms of probablty. The probablty of a specfc selecton s assgned to a seres of explanatory values. Ths seres of values gathers the characterstcs of decson-makers and/or the attrbutes of the varous choce alternatves. 7

9 Multnomal logt or probt models are used when there are more than two alternatves. However, they fal to account for the ordnal nature of the dependent varable used n ths research. We am to model the choce of sequental bnary decsons, the frst consstng of a par of countres that ether sgn a preferental trade agreement (PTA) or do not. Once a country comes to a blateral agreement, the next decson wll be whether to take a further step and go to a hgher level of ntegraton. Therefore, the model objectve s to take a seres of bnary decsons, each consstng of the decson of whether to accept the current value or to take one more. 2 In ths context, Amemya (975) descrbes a model that apples to ordered dscrete alternatves, such as the number of cars owned by a household. Ths s based on the assumpton of local (as opposed to global) utlty maxmsaton. The decson-maker stops when the frst local optmum s reached. Economc agents must choose between two sequental optons, and ther selecton depends on ther characterstcs and ther envronment. In accordance wth the characterstcs of our dependent varable, an ordered logt model was specfed n our study. The model s bult around a latent regresson n the same way as the bnomal probt model. An observed ordnal varable, Y, s a functon of an unobserved latent varable, Y*, whch represents the dfference n utlty levels from an acton. The contnuous latent varable Y* has a number of threshold ponts, and the value of the observed varable Y depends on whether or not a partcular threshold s crossed. In the present analyss we assume that fve dfferent ntegraton levels can be reached, therefore the number of thresholds s fve, Y = 0 f Y* δ 2 There are nstances n whch the RIAs are morbund, then countres can decde to take one less. Ths s not the case n the data beng looked at. 8

10 Y = f δ Y* δ 2 Y = 2 f δ 2 Y* δ 3 () Y = 3 f δ 3 Y* δ 4 Y = 4 f δ 4 Y* δ 5 Y = 5 f Y* δ 5 where the δs are the unknown parameters to be estmated. Threshold denotes that a par of countres engages n a PTA, threshold 2 denotes an FTA, threshold 3 s a CU, threshold 4 s an SM, and threshold 5 represents an MU. The contnuous latent varable s gven by, Y * = k k= β X + ε = Z + ε (2) k k where X k are the explanatory varables, β k are the coeffcents and ε s the random dsturbance term that s assumed to be ndependent of X and has a logstc dstrbuton. The ordered logt model estmates, Z = k k= β X = E( Y * ) (3) k k Once the β k parameter and the M- δs have been estmated, they can be used to calculate the probablty that Y wll take on a partcular value. For example, when M=6, Pr( Y = 0) = Pr ( Z 0) = + exp ( Z δ ) Pr( Y = ) = Pr( Z δ ) Pr( Z 0) = (4) + exp Pr( Y = 2) = Pr ( Z δ ) Pr( Z δ ) 2 = + exp ( Z δ ) + exp( Z δ ) 2 ( Z δ ) + exp( Z δ ) 3 2 9

11 Pr( Y = 3) = Pr ( Z δ ) Pr( Z δ ) 3 2 = + exp ( Z δ ) + exp( Z δ ) 4 3 Pr( Y = 4) = Pr ( Z δ ) Pr( Z δ ) 4 3 = + exp ( Z δ ) + exp( Z δ ) 5 4 Pr( Y = 5) = Pr ( δ Z ) 5 = + exp ( Z δ ) 5 Hence, usng the estmated value of Z and the assumed logstc dstrbuton of the dsturbance term, the ordered logt model can be used to estmate the probablty that the unobserved varable Y* falls wthn the varous threshold lmts. The unknown coeffcents and the thresholds can be estmated numercally by the maxmum lkelhood method, where the above probabltes are the elements of the lkelhood functon. The probablty that a hgher ntegraton level s chosen ncreases f the βs are postve and the correspondng explanatory varable ncreases. Ths can be seen by calculatng the dervatves of the cumulatve probabltes: Pr( Y X M ) k = β j exp( Z δ ) ( + exp( Z δ )) 2 k k (5) Snce the nterpretaton of the coeffcents of ths knd of model s unclear, a commonly used practce s to calculate the margnal effects assocated wth the probablty of an RIA beng formed or hgher ntegraton stages beng establshed. They are gven by: Pr( Y X = M ) k exp( Z δ k ) exp( Z δ k ) = β j (6) 2 2 ( + exp( Z )) ( + exp( )) δ k Z δ k One advantage of an ordered logt over an ordered probt model s ts smplcty. However, t s subject to the Independence of Irrelevant Alternatves (IIA) property, whch consttutes a tght lmtaton as all alternatves must follow an ndependent choce 0

12 functon. Selecton pars P /P j of alternatve over j are ndependent of whether thrd alternatves exst. The advantage of ths condton s that t enables the ntroducton of new alternatves, such as new ntegraton levels, wthout havng to re-estmate the model. The dfference between the estmated parameters must be the same, regardless of the number of alternatves that the economc agent faces. The dsadvantage of ths property s that alternatves must be perceved as dstnct and ndependent. The evaluaton of ths type of model dffers from tradtonal models n certan ways. Even though the rato of an estmated coeffcent to ts correspondng estmated standard error follows a t-student dstrbuton, the F test s not approprate for these models. The most commonly accepted test s the Pseudo-R 2, a scalar measure of the explanatory power of the model derved from the maxmum lkelhood rato 3. Ths test s defned as: log L = log L 2 u ρ (7) c Where: L u = the lkelhood functon of the model wth explanatory varables. L c = the lkelhood functon of the model wthout explanatory varables and only one constant. ρ 2 les between zero and one, and equals when the model s a perfect predctor: f Y = P = F( X β ) = (8) 0 f Y = 0 P takes value 0 f log L c = log L u, thus ρ 2 ncreases to when log L c rses n relaton to log L u. An alternatve way to evaluate the goodness of ft of an ordered logt s to calculate the exp (log lkelhood / number of observatons) whch s the geometrc average of P (O j / 3 Also known as the lkelhood rato ndex (LRI).

13 X j, estmates), where O j and X j are the outcome and the explanatory varables for observaton j. Ths rato shows the probablty of obtanng a certan outcome condtonal on the estmates. The hgher the rato s, the greater the explanatory power of the model wll be. The nterpretaton of coeffcents n an ordered logt model also dffers explctly from other models. In dscrete choce logt and probt models, the sgn of the coeffcents denotes the drecton of swtch, but ts magntude s dffcult to nterpret. For example, the postve coeffcents correspondng to the characterstcs of the ndvduals n the ordered logt model estmated n ths paper ncrease the probablty that a par of countres wll be observed n a hgher ntegraton category. However, negatve coeffcents ncrease the probablty that a par of countres wll be observed n a lower ntegraton category. 4. Data, hypothess and varables 4.. The data The model s estmated wth the data of 66 countres from 999, representng over 75% of world trade (see Table A., Appendx A). Data on ncome are obtaned from the World Development Indcators (200). Dstances are the great crcle dstances between economc centres. Data on captal labour ratos are obtaned from the Penn World Tables. Data on blateral exports are obtaned from Statstcs Canada (200), and tarff barrers from the World Bank webste. Informaton about geographcal and language dummes s from the CIA (2003). The Economc Freedom Index was obtaned from the Hertage Foundaton, and the poltcal regme, from the Freedom House. Table A.2 n Appendx A presents a more detaled descrpton of data and sources. Fnally, the agreements consdered to buld the dependent varable are lsted n Table A.3 (Appendx A). 2

14 4.2. Hypothess and varables Accordng to the underlyng theory descrbed above, and n the context of the dscrete choce model, our frst hypothess s that a par of countres wll be more lkely to form or enhance an RIA when the dstance between them s small. We specfy the dstance varable as n BB. Ths varable s called natural as t s defned as the logarthm of the nverse of dstance between tradng partners. A second hypothess s that the probablty of RIA formaton or enhancement ncreases as the remoteness of a country or par of countres from the rest of the world rses. For comparatve purposes, we constructed the same remoteness varable used by BB. When a country s relatvely far from ts tradng partners, t tends to trade more blaterally wth ts neghbours, thereby ncreasng the probablty of RIA formaton. The thrd hypothess s that the larger the economc sze of the tradng countres, the greater the probablty of RIA formaton or enhancement wll be. RGDPj measures the sum of the logs of real GDPs of countres and j n The fourth hypothess s that the more smlar the countres economc sze s, the hgher the probablty of RIA formaton or enhancement wll be. DRGDPj s the absolute value of the dfference between the logs of real GDPs of countres and j n 960. The ffth hypothess s that the larger the countres economc sze outsde the RIA s, the lower the probablty of RIA formaton or enhancement wll be. However, the sze of the rest of the world (ROW) measured by the ROW GDP vares only slghtly n a cross-secton of countres and has not been ncluded n the regresson. BB obtaned a non-sgnfcant coeffcent for ths varable. The sxth hypothess s that the probablty that a par of countres wll form or enhance an RIA s hgher f there s a larger dfference n ther relatve factor endowments snce 4 Data are from 960 to avod the problems derved from the endogenety of ncome n the estmated equaton. The same apples to varables DRGDPj and DKLj. 3

15 tradtonal comparatve advantages wll be further exploted. However, f ntercontnental transport costs are low, ths probablty may also decrease at hgh levels of specalsaton. Ths can be modelled by addng a quadratc term to the estmated equaton. We use absolute dfferences n the captal stock per worker rato (DKLj) as a proxy for relatve factor endowment dfferences, as n BB 5. SQDKLj denotes squared DKLj. The seventh hypothess s that more democratc countres dsplay a greater lkelhood of concludng RIAs than other countres, as stated by Mansfeld, Mlner and Rosendorff (2002). The eghth hypothess s that a par of countres s more lkely to form or enhance an RIA than f they have a hgher level of economc freedom and f they speak a common language. The nnth hypothess s that nteror countres (landlocked) as well as adjacency countres wll have a hgher probablty of engagng n an RIA, especally wth coastal countres. However, when a landlocked country trades wth partners located n another contnent (unnatural partner), t wll have hgher transport costs than a coastal country. Fnally, the tenth hypothess s that countres wth hgher levels of protecton (tarffs) wll have more ncentves to create or enhance an RIA wth other countres n order to lower (or elmnate) artfcal trade barrers and to facltate trade. Supplementary economc, geographcal and soco-poltcal varables are added to the lst of varables used by BB as determnants of RIAs (hypotheses 7-0). Landlocked status and adjacency are added to the lst of geographcal varables used by BB. The soco-poltcal varables consdered are: tarff barrers, sharng a common language, the poltcal regme (ths varable takes a value of when the poltcal regme was a 5 Data are for 965 rather than 960, snce data on captal labour ratos s only avalable from 965 onwards n the Penn World Tables data seres. Baer and Bergstrand (2004) use data for

16 democracy n 950) 6, and the level of economc freedom. The economc freedom varable takes a value between -.99 for free countres, for mostly free countres, for mostly non-free countres and for repressed countres. Accordng to the hypotheses above, tarffs, language and democracy are expected to have a postve sgn, and economc freedom s expected to have negatve coeffcents 7. Blateral trade flows were ntally added as an economc varable. Trade flows were expected to have a postve sgn snce more trade between countres ndcates a strong relatonshp and dependence, and a reason to sgn an RIA. However, due to the endogenety problems found for blateral trade, we chose to exclude ths varable from the estmatons. Magee (2003) provdes one of the frst assessments of the hypothess that two countres are more lkely to form a PTA f they are already major tradng partners. He estmates a probt and a non-lnear two-stage least squares model that consders trade flows to be endogenous n the second specfcaton. Magee s results show that greater blateral trade flows sgnfcantly ncrease the lkelhood that countres wll form a preferental trade agreement n every specfcaton of the model. The frst model estmated s a bnary probt; the dependent varable takes the value of one when the countres reach an ntegraton agreement, and zero otherwse, and the ndependent varables are those lsted above. The second model estmated s an ordered logt. Fve dfferent possble levels of ntegraton between pars of countres are consdered to nvestgate the determnants of regonal ntegraton agreements (RIAs). The varables ncluded are the same, but the nterpretaton of the estmated coeffcents s slghtly dfferent. 6 Data for ths varable were only avalable for the years 950 and To avod the problems derved from the endogenety of democracy n the estmated equaton, we used the data from Note that accordng to the defnton of these varables, hgher values mply lower economc freedom. 5

17 5. Estmaton Results 5.. Probt estmaton The results obtaned when a bnary probt s estmated are shown n Tables and 2. The results n Table wll be comparable to those obtaned by BB 8, although the sample of countres consdered s not exactly the same, the year s 999 nstead of 996, and the defnton of the dependent varable also vares slghtly. Table. Probt results for the probablty of RIA formaton. The frst hypothess to be tested s that the smaller the dstance between the two countres, the more lkely ther socal planners wll be to form an RIA, snce the closer the two tradng partners are, the fewer ther trade barrers wll be. The probablty of establshng an RIA ncreases wth dmnshng dstances between the tradng countres. BB obtan a postve coeffcent (.74) n ther equvalent Model. We also obtan a postve coeffcent (0.56), but t s lower n magntude. The second hypothess s tested n Model 2. For a gven dstance between two countres, the more remote the two contnental tradng partners are from the rest of the world, the more lkely they wll be to form an RIA. We calculate ths varable accordng to BB and we obtan a postve coeffcent that s smlar n magntude. In Model 3, the thrd hypothess s that the larger the tradng partners are n economc terms, the greater the probablty of an RIA beng formed wll be. Ths effect s captured by RGDPj, and t s postve and sgnfcant, as expected. However, the coeffcent obtaned n ths paper s lower than that obtaned n BB. 8 Table B. n Appendx B. 6

18 In Model 4, the fourth hypothess s tested. The greater the smlarty between the economc sze of the two countres, the hgher the probablty of an RIA beng establshed wll be. Ths effect s captured by DRGDPj. We obtan the expected negatve sgn and ths varable s sgnfcant. Fnally, the sxth hypothess s tested n Models 5, 6 and 7. Accordng to BB, the larger the dfference between countres relatve factor endowments, the greater the probablty of FTA formaton wll be, although ths may only be true to a lmted extent. Varables DKLj and SQDKLj (DKLj squared) measure ths effect. When we nclude these two varables n the same regresson, they are not sgnfcant. Snce the two varables are hghly correlated, DKLj and SQDKLj are ncluded n Models 5 and 7 respectvely. Both varables are sgnfcant, but they do not have the expected sgns. The negatve sgn obtaned for DKLj ndcates that the larger the dfference between countres relatve factor endowments s, the lower the probablty of an RIA wll be. Ths result ndcates that the socal planners from the two countres tend to form an RIA when they have smlar relatve factor endowments. Accordngly, hgher levels of ntra-ndustry trade wll be desrable f RIAs are to be formed, snce countres wth smlar endowments trade smlar commodtes. Ths does not support the noton of natural tradng partners defned by Schff (999) as beng complementary between partners (one country tends to mport what the other exports). A plausble explanaton from the demand sde may be that because countres wth smlar endowments have smlar tastes and love varety, ther governments wll be more lkely to negotate hgher levels of ntegraton. BB test an addtonal hypothess. The hgher the absolute dfference between the relatve factor endowment of the member countres and the relatve factor endowment of the ROW, the lower the probablty of FTA formaton wll be, due to potental trade 7

19 dverson. We construct DROWKLj accordng to BB, although we aggregate the rato K/L rather than aggregatng both varables separately 9 snce we do not have detached data for captal and labour. We use Equaton (9) to calculate DROWKLj. {log DROWKLj = N N [ K k / Lk ] log[ K L ] + log [ K k / Lk ] log[ K j L j ] k =, k 2 k =, k j } (9) The coeffcent of ths varable s close to zero (0.05) and s not sgnfcant. Dfferences n the two sets of results can be explaned by the dfferent constructons of the dependent varable: BB consder only full FTAs or customs unons (CUs), whereas our dependent varable ncludes all PTAs, FTAs, CUs, SMs and MUs notfed to GATT/WTO under Artcle XXIV and under the Enablng Clause (see Table A.3, Appendx A). Ths varable s broader snce t regards ntegraton agreements as a process wth dfferent levels of ntegraton; t makes sense to estmate an ordered logt wth ths constructon. BB use the Pseudo R2, calculated accordng to Equaton (7) above, as a measure of explanatory power. However, Henen (993) ponts out that although ths ndex s not affected by changes n sample sze, t s affected by the presence of mssng observatons. 0 In ths case, a better alternatve s to calculate McFadden s R2 whch takes the mssng values nto account. Table shows both Pseudo R2 and McFadden s R2. The McFadden statstc consders that there s a dfferent number of observatons n the restrcted and unrestrcted models when there are mssng values for some varables. 9 Baer and Bergstrand (2004) measure DROWKLj as: {log DROWKLj= N K k k=, k j N k k=, k L log N N [ K L ] + log K k L k log[ K j Lj ] 2 k=, k j k=, k } 0 BB dd not have mssng observatons. Therefore n the context of ther emprcal exercse, the pseudo R-squared seems approprate. 8

20 In the estmated Models 3-7, McFadden s R2 s preferred to Pseudo R2 snce there are zero values n some of the explanatory varables. In order to have a clear representaton of the relatve mportance of the addtonal geographcal and soco-poltcal varables, a probt model extended wth those varables was also estmated. The man results are reported n Table 2. Table 2. Probt results for the probablty of RIA formaton: Extended model. In the frst column of Table 2 (Model 5) the same results reported n Table are also ncluded for comparatve purposes. Model 8 to Model 2 n columns 3 to 7 of Table 2 are estmated for dfferent sets of varables grouped as geographcal (Models 8 and 9), economc and geographcal (Model 0), soco-poltcal varables (Model ) and all the varables (Model 2). Ths sequental analyss enables us to fnd out the most mportant factors n promotng RIAs usng the smple probt model. Models 8 and 9 n Table 2 show the results of the geographcal varables. All geographcal varables are sgnfcant at %, and natural, remoteness and adjacency have a postve sgned coeffcent, whle the landlocked varable coeffcent s negatve. In Model 9, the nteracton varable (landlocked*remoteness) s added to nvestgate the reasons why the landlocked varable s negatvely sgned. The estmated coeffcent for the nteracton term s postve, ndcatng that the probablty of jonng an RIA ncreases for more remote contnental tradng partners when one of them s landlocked. Geographcal varables alone explan 6% of the varablty of the dependent varable. Model 0 reports the results of addng economc and geographcal varables. In ths extended model, the dummes landlocked and adjacency are not statstcally sgnfcant and the former shows a postve sgn. Model, n column sx of Table 2, shows that all 9

21 the soco-poltcal varables are sgnfcant: democracy, the level of economc freedom, the common language and hgher tarffs promote RIAs formaton. However, n terms of goodness of ft, Pseudo R2 s very low (0.05). Fnally, Model 2 reports the results wth economc, geographcal and soco-poltcal varables. In ths model an addtonal nteracton varable has been ncluded (natural*language) snce the effect of the common language dummy s only postve and sgnfcant when t s nteracted wth the natural varable. The adjacency varable s not sgnfcant and the language dummy ntally showed an unexpected negatve sgn that reverses when the varable s nteracted wth the natural varable Ordered logt estmaton We estmate an ordered logt model consstng of a system of 5 equatons wth common coeffcents for all the explanatory varables and wth dfferent constant terms. Ths s known as the proportonal odds model. In the frst column of Table 3 (Model 3), an ordered logt s estmated wth the same varables ncluded n Model 5 (probt estmaton). Model 4 to Model 6 n columns 3 to 5 of Table 3 are estmated for dfferent sets of varables grouped as economc, geographcal, soco-poltcal varables, and Model 7 ncludes all the varables. Ths sequental analyss enables us to fnd out the most mportant factors n promotng RIAs. Table 3. Ordered logt results for the probablty of RIA formaton or enhancement. Model 3 shows that the results are smlar n both probt and ordered logt models, although the logt ordered coeffcents are hgher n magntude. In general terms, we can state that the probablty of reachng a hgher level of ntegraton s hgher than the 20

22 probablty of sgnng any type of RIA when no prevous agreement exsts between the tradng countres. However, as stated above, there s no consensus on the nterpretaton of the magntude of the coeffcents estmated n dscrete choce models. Models 4 and 5 n Table 3 show the results of the geographcal varables. All geographcal varables are sgnfcant at %, and natural, remoteness and adjacency have a postve sgned coeffcent, whle the landlocked varable coeffcent s negatve. In Model 5 the nteracton varable (landlocked*remoteness) s added to consder the ambguous sgn expected for the landlocked varable. The estmated coeffcent shows a postve sgn, ndcatng that the probablty of reachng a hgher level of ntegraton ncreases for more remote contnental tradng partners when one of them s landlocked. Model 6, n column sx of Table 3, shows that all the soco-poltcal varables are sgnfcant: democracy, the level of economc freedom and the common language promote RIA enhancement. However, n terms of goodness-of-ft, Pseudo R2 s very low (0.04). The coeffcent on tarffs s postve, thus showng that a hgher level of protecton ncreases the probablty that a country par wll be observed n a hgher category. Fnally, Model 7 ncludes economc, geographcal and soco-poltcal varables. Some nteracton terms were also added to allow for the possblty that the effect of some varables, namely remoteness and language, could be dfferent for natural and unnatural patterns. In ths model, remoteness presents a negatve sgn, ndcatng that remote countres have a lower probablty of reachng hgher levels of ntegraton, whle the varables adjacency, language and tarffs are not statstcally sgnfcant. The Akake Info Crteron (AIC) shows that the best specfcaton s that estmated n Model 7, where all the varables are consdered. For the specfcaton where only geographcal varables are consdered, the AIC s lower (.542) than that obtaned n 2

23 regressons ncludng only soco-poltcal factors (.68). Ths appears to ndcate that geographcal varables are mportant determnants of RIA formaton. As stated above, the nterpretaton of the coeffcents n an ordered logt does not nform of the magntude of swtch snce we can only state that postve coeffcents ncrease the lkelhood that the country pars wll be observed n a hgher category, and negatve coeffcents ncrease the lkelhood that the country pars wll be observed n a lower category. A preferable nterpretaton of the ordered logt coeffcents s n terms of the odd ratos. The exponentated coeffcents n the logt model, shown n Table 4, can be nterpreted as odds ratos for a -unt change n the correspondng varable. The emphass s on the rato Exp(β), whch s the odds condtonal on x+ dvded by the odds condtonal on x. For example,.9 means that the odds of beng n a hgher ntegraton level ncrease by.9 f RGDP ncreases by. The nterpretaton can also be made n terms of percentages: the exp(.49) obtaned n the natural varable n Model 3 means that the odds ncrease by 346% {[exp(.49)-]*00} f the varable ncreases by, therefore the odds of beng part of the monetary unon versus lower ntegraton levels s 346% hgher for a one-unt ncrease n the natural varable. Table 4 shows that, n Model 3, the most mportant determnant of an RIA s the natural varable, followed by remoteness (.37), real GDP (.9), real GDP dfferences (0.84) and K/L dfferences (0.77). We also calculate sem-standardsed ordered logt coeffcents that control for the metrcs of the ndependent varables to see whether any change occurs n the orderng of effects. The opton of standardsed coeffcents to measure the relatve strength of the effects of the ndependent varables s more approprate n the current emprcal applcaton snce some ndependent varables are measured n dfferent unts. Table 4 shows that when standardsed coeffcents are consdered (e^bstdx), the orderng of the 22

24 effects changes only slghtly. In Model 3, the natural varable s standardsed coeffcent s 3.89, and t s.76 for remoteness,.65 for RGDP, 0.75 for K/L dfferences, and 0.74 for real GDP dfferences. For one standard devaton ncrease n natural, the odds are 3.89 tmes greater (an ncrease of 289%) of countres beng n a hgher ntegraton category when all the other varables are held constant. In Model 7, where soco-poltcal varables are added, the natural varable s stll the most mportant followed by real GDP and democracy. Table 4. Odds ratos for the ordered logt. In order to evaluate the probablty that the dependent varable wll have a partcular value, we use cut-offs terms. From Equaton (), the threshold parameters for Model 3 are gven by: Y = 0 f Y* -3.4 Y = f 3.4 Y* -2.7 Y = 2 f 2.7 Y* -.8 Y = 3 f.8 Y* -.58 Y = 4 f.58 Y* 0.38 Y = 5 f Y* 0.38 For example, when the tradng partners are Argentna and Paraguay, we can calculate the probablty assocated wth ths par of countres by computng Z wth the obtaned coeffcents n Model 3 and the correspondng data: Z = β RGDPj + β DRGDPj + β DKLj + β REMOTE + β NATURAL = 2 3 = ( ) + ( ) + ( ) + (.49 ( 6.94)) + ( ) = 2.28 Pr( Y = 0) = Pr( Z 0) = = exp 2.28 ( 3.4) ( )

25 Pr( Y = ) = Pr( Z δ ) Pr( Z 0) = = exp( 2.28 ( 2.7) ) + exp( 2.28 (3.4)) Pr( Y = 2) = Pr( Z δ 2 ) Pr( Z δ ) = = exp( 2.28 (.8) ) + exp( 2.28 ( 2.7) ) Pr( Y = 3) = Pr( Z δ 3 ) Pr( Z δ 2 ) == = exp( 2.28 (.58) ) + exp( 2.28 (.8) ) Pr( Y = 4) = Pr( Z δ 4 ) Pr( Z δ 3 ) = = exp 2.28 (0.38) + exp 2.28 (.58) Pr( Y = 5) = Pr( δ 5 Z ) = = exp ( 2.28 (0.38)) ( ) ( ) Hence for Argentna and Paraguay, the most lkely outcome s that they wll form a sngle market. In fact, they have been members of Mercosur snce 995. Our second example s Span and France, a par of tradng partners that are members of the European Unon. Our results ndcate that the hghest probablty s that of the establshment of a sngle market. In 999 these countres were already n the thrd phase of the European Monetary Unon (EMU), snce they fulflled the convergence crtera establshed n the Treaty of Maastrcht. However, our results most probably show that they were only n the EMU startng phase. Z = β RGDPj + β DRGDPj + β DKLj + β REMOTE + β NATURAL = 2 3 = ( ) + ( ) + ( ) + (.49 ( 6.96)) + ( ) = 0.4 Pr( Y = 0) = Pr( Z 0) = = exp( 0.4 ( 3.4) ) Pr( Y = ) = Pr( Z δ ) Pr( Z 0) = = exp( 0.4 ( 2.7) ) + exp( 0.4 ( 3.4) ) Pr( Y = 2) = Pr( Z δ 2 ) Pr( Z δ ) = = exp( 0.4 (.8) ) + exp( 0.4 ( 2.7) ) Pr( Y = 3) = Pr( Z δ 3 ) Pr( Z δ 2 ) == = exp( 0.4 (.58) ) + exp( 0.4 (.8) ) Pr( Y = 4) = Pr( Z δ 4 ) Pr( Z δ 3 ) = = exp( 0.4 (0.38)) + exp( 0.4 (.58) ) P( Y = 5) = Pr( δ 5 Z ) = = exp 0.4 (0.38) ( )

26 When the soco-poltcal varables are also consdered (Model 7), then the threshold parameters are gven by: Y = 0 f Y* Y = f Y* -4.0 Y = 2 f -4.0 Y* Y = 3 f Y* -2.9 Y = 4 f -2.9 Y* -0.4 Y = 5 f Y* -0.4 Therefore, for the second example (Span and France), our results ndcate that the hghest probablty s that of the establshment of a monetary unon. Z = β RGDPj + β DRGDPj + β DKLj + β REMOTE + β NATURAL + β ADJj + β LANDj + + β LANDj REMOTE + β NATURAL REMOTE + β β3 TRADE _ BARRIERS + β4 NATURAL LANGj = -. Pr( Y = 0) = Pr( Z 0) = = exp. ( 4.82) Pr( Y = ) = Pr ( Z δ ) Pr( Z 0) 3 ( ) LANGj + β DEM 50j + β FREEDOMj + = = ( Z δ 2 ) Pr( Z δ ) Pr( Y = 2) = Pr + exp (. ( 4.0) ) + exp(. ( 4.82) ) = = ( Z δ 3 ) Pr( Z δ ) Pr( Y = 3) = Pr 2 + exp (. ( 3.23) ) + exp(. ( 4.0) ) == = ( Z δ 4 ) Pr( Z δ ) Pr( Y = 4) = Pr 3 + exp (. ( 2.9) ) + exp(. ( 3.23) ) = = + exp Pr( Y = 5) = Pr( δ 5 Z ) = = exp (. ( 0.4) ) + exp(. ( 2.9) ) (. ( 0.4) ) The calculaton of the predcted probabltes for all the tradng partners shows that 69% of the agreements and 84% of the non-agreements were correctly predcted by the ordered logt model. Of all cases, 7% had excessve blateralsm 2,.e., when the Model 3. 2 Excessve and nsuffcent blateralsm are terms used by BB. 25

27 predcted level of ntegraton was lower than the real level, and we found that blateralsm was nsuffcent for 6.5% of the tradng partners Margnal effects As BB pont out, one complcaton arses n estmatng the partal effects on the response probabltes for the partcular vector of RHS varables, x, n our model by usng mean values for the levels. One of the RHS varables, REMOTE, s the product of a contnuous varable and a bnary varable ( ) the mean value of ths varable s economcally meanngless. 3 As we also use REMOTE, we estmate separately the margnal effects on the response probabltes wth the mean value of REMOTE when the tradng partners are n the same contnent, and when REMOTE takes the value of zero (the tradng partners are not n the same contnent, they are unnatural partners). Margnal effects are calculated for Model 2 (probt) and Model 7 (ordered logt). 4 Our results for the probt estmaton are shown n Table 5 and for the ordered logt estmaton n Table 6. Our results n Table 5 can be compared wth those obtaned by BB, shown n Appendx B (Table B.2), although we nclude two addtonal socopoltcal varables (economc freedom and trade barrers) that were not consdered by BB. Table 5. Response probabltes for natural and unnatural tradng partners n Model 2 (evaluated at the mean level of remote and at remote = 0). Table 5 shows that the response probablty of an RIA beng created s much lower for unnatural partners (5.8%) than for natural partners (88.5%). Moreover, results show that 3 Baer and Bergstrand (2004), page Dummy varables are not ncluded snce the mean values of these varables do not have an economc nterpretaton. 26

28 a untary ncrease n proxmty (natural varable) ncreases ths probablty by 7.3% for two natural partners and.8% for unnatural partners. An ncrease n remoteness from the ROW of two natural partners lessens the response probablty n natural partners. Although ths s an unexpected result, Table 6 shows that the sgn of ths margnal effect changes for dfferent levels of ntegraton when the model estmated s an ordered logt rather than a bnary probt model. Ths sgn s only postve for the frst two ntegraton stages (PTA and FTA). When two countres are n the same contnent and they are relatvely far away from the other countres n ths contnent, then the probablty that they wll reach a customs unon decreases wth the level of remoteness. The results show that economc varables have a lower effect than geographcal and soco-poltcal factors on response probabltes, although dfferences n ncome also play an mportant role n RIA formaton. The response probablty for natural partners s smlar to that obtaned by BB, who fnd 86.7% probablty of an FTA beng establshed between natural partners. However, they only obtan a probablty of.2% for unnatural partners. We obtan a hgher probablty for unnatural partners because we also consdered preferental trade agreements n the constructon of the dependent varable. To compare the effect of the RHS varables across dfferent levels of ntegraton, n Table 6 we estmate the margnal effects for all the ntegraton levels for both natural and unnatural partners. Table 6. Response probabltes for natural and unnatural tradng partners n Model 7 (evaluated at the mean level of remote and at remote = 0). 27

29 Table 6 shows dfferent probabltes dependng on the level of ntegraton. For each level of ntegraton, the probabltes are shown for natural and for unnatural partners. However, for the three last categores (customs unon, sngle market and monetary unon) the probabltes can only be calculated for natural partners snce these ntegraton levels are only reached by countres n the same contnent. These probabltes depend manly on geographcal, soco-poltcal and economc varables, and ther margnal effects dffer across ntegraton levels. On the one hand, the results obtaned for natural partners (countres n the same contnent) ndcate that when remoteness ncreases by %, the probablty of a PTA or an FTA beng establshed ncreases by 337% and 59%, respectvely. However, the probablty of a customs unon or a hgher ntegraton agreement beng establshed decreases wth remoteness. Ths varable, together wth soco-poltcal factors, s the most nfluental factor on the probablty of an RIA beng formed or enhanced between natural partners. Hgher GDP dfferences ncrease the probablty of PTA or FTA formaton for natural partners, although the sgn of the margnal effect for hgher levels of ntegraton s reversed, thus ndcatng that smlarty of ncome, as expected, ncreases the probablty that hgher levels of ntegraton (customs unon, sngle market and monetary unon) wll be reached. The ntegraton theory predcts that the costs of ntegraton are lower when countres have smlar levels of ncome and, consequently, a hgh level of ntra-ndustry trade. For unnatural partners however (countres n a dfferent contnent), the nverse of dstance s the most mportant factor n PTA or FTA formaton, and hgher dfferences n ncome and n factor endowments lower the probablty of a PTA or an FTA beng establshed. 28

30 Fnally, the results show the most lkely outcomes are that natural partners wll establsh a sngle market and unnatural partners wll not reach any agreement. When we order the probabltes for the varous types of ntegraton agreements from the hghest probablty to the lowest probablty for natural partners, we obtan: Pr(SM or 4) = 0.38 Pr(FTA or 2) = 0.9 Pr(PTA or ) = 0.7 Pr(CU or 3) = 0.2 Pr(MU or 5) = 0.03 These fndngs can seem surprsng snce the (conventonally assumed) second most ntegrated type of agreement, a sngle market, s the most lkely type of RIA. An explanaton s that the results obtaned are lkely to be domnated by the European Common Market Senstvty Analyss We performed several robustness tests to valdate our results. Frstly, the ordered logt model s based on the assumpton of parallel slopes but ths may be unrealstc, for example, f geographcal varables are less relevant for hgher ntegraton levels. Therefore, the Brant test of the parallel regresson assumpton s used to valdate the methodology used. The Brant (990) test assesses whether or not the coeffcents are the same for each category of the dependent varable. Ths produces Wald Tests for the null hypothess that the coeffcents n each ndependent varable are constant across categores of the dependent varable. Sgnfcant test statstcs provde evdence that ths assumpton has been volated for most of the varables. Wth the excepton of the captal-labour rato, we cannot accept the equalty of slopes for the dfferent levels of ntegraton (Table A.4). These results ndcate that we should estmate a generalsed 29

31 logt model, and they suggest what varables may be used n determnng the thresholds. We therefore estmated a generalsed ordered logt for all the regressons presented n Table 3. In some cases, the model dd not converge, especally when the varables wth mssng data (K-L dfferences) were ncluded. The results 5 ndcate that the geographcal varables are sgnfcant and show the expected sgns for the lower levels of ntegraton (PTA, FTA), whereas these varables lose sgnfcance and decrease n magntude for the hgher levels. In contrast, the economc and poltcal varables gan mportance n the hgher levels of ntegraton. Secondly, we re-estmated the probt and ordered logt model wth an alternatve data set taken from Magee (2003), whch are avalable for replcatons on hs web ste. Hs dependent varable takes the value of one f the country par has a PTA n 998, and takes zero otherwse. We use the same dependent varable n the probt estmaton, but Magee (2003) consders fewer agreements snce he gnores the General System of Trade Preferences (GSTP), the Protocol related to Trade Negotatons among developng countres (PTN) and the Afrcan Common Market. Addtonally, the varable remoteness s not ncluded as an explanatory varable n Magee (2003). We estmate a bnary probt for 72 countres n 998 and our results confrm the sgn and sgnfcance of the estmated coeffcents for the ncome varables, the relatve factor endowment dfferences and the natural varable (Model 7. n Table ). Contrary to BB, the K-L dfferences varable s negatve and sgnfcant, thus valdatng our evdence. Smlar results are obtaned when an ordered logt s estmated. Thrdly, we also estmated the probt model wth the ncluson of blateral trade as an explanatory varable and usng nstrumental varables to correct endogenety problems. Infrastructure varables were used as nstruments for trade. The results ndcate that 5 Results are avalable upon request from the authors. 30

32 trade s sgnfcant and has a postve sgn when t s added to the lst of explanatory varables n the probt model, confrmng the evdence presented by Magee (2003) wth dfferent data and dfferent model specfcatons. However, as stated above further research s needed to mprove the model specfcaton. Fourthly, the observatons are twce the number of country pars. However, our dependent varable s symmetrc and only trade and tarffs are asymmetrc ( X X ). Therefore, we have re-estmated the model wth only half the observatons to check whether ths would have affected the results. By takng 245 ((66*65)/2) country pars, the results remaned unchanged. 6 Ffthly, an addtonal robustness test has been performed. We checked whether the results were affected by the excluson of an mportant economc bloc, such as the EU. The results excludng the EU countres also remaned unchanged. 7 Fnally, the ordered nature of the dependent varable and the endogenety of trade flows should deally be consdered smultaneously, although ths s beyond the scope of ths research. 6. The dynamcs of the European Unon ntegraton process The EU s the best real example of a successful ntegraton process. However, the fact that the analyss n the prevous sectons focuses on data for 999 covers nether the entrance of 0 countres nto the EU n 2004 nor the adopton of the Euro by Greece. 8 In order to tackle the above-mentoned ssues, the proposed model has been estmated for an addtonal sample, ncludng data for the EU-27 from 999 to A dynamc analyss would also be possble by addng the tme dmenson to the data. j j 6 The results of takng nto account the repetton bas n the 66-country sample are avalable upon request from the authors. 7 These results are avalable upon request from the authors. 8 A referee kndly suggested the ncluson of ths secton n the paper. 3

33 In relaton to the soco-poltcal factors, democracy n 950 was used n Secton 5. Nonetheless, ths varable may have very lttle to do wth the probablty of a country par formng or enhancng an RIA durng the perod n Europe. Although Span and Portugal were dctatorshps n 950, both restored democracy n the md- 970s, and joned the European Communty (EC) n 986. Greece also restored democracy n the md-970s and joned the EC n 98. Hence, these three countres were democraces at the tme they joned the EC. The same apples to the former socalst countres that joned the EU n 2004 and Therefore, unlke the analyss performed n Secton 5, we take nto account the poltcal regme at the tme of entry nto the EC and not the stuaton n 950. Instead of a dummy varable for democracy, the varable polcy s used. 9 Poltcal rghts and cvl lbertes at the tme of entry nto the EC have also been added to the lst of poltcal varables. They are measured on a one-to-seven scale, wth one representng the hghest degree of freedom and seven the lowest. 20 Table 7 shows the results obtaned for the EU-27 sample. In the frst column of Table 7 (Model 8), an ordered logt s estmated wth the same varables ncluded n Model 3 (Table 3). 2 Model 9 to Model 2, n columns 2 to 4 of Table 7 report the results for models wth dfferent sets of varables grouped as geographcal and soco-poltcal varables. Fnally, Model 22 ncludes all the varables, as does Table 3. 9 Annual data for democracy are obtaned from the Polty IV dataset ( The varable POLITY2, whch vares from -0 (strong dctatorshp) to 0 (full democracy), s used n Secton Annual data on poltcal rghts and cvl lbertes are obtaned from The Freedom House (2009): 2 DKL s not ncluded n the analyss for the European ntegraton process snce DKL was not sgnfcant n the deepest ntegraton levels (see Table 6). Remoteness s also calculated for the European country sample as was done n Baer and Bergstrand (2004), however, ths varable s not ncluded n the regressons snce s not consdered as comparable to the one constructed for the 66-country sample whch ncludes unnatural partners. 32

34 Table 7. Ordered logt results for the probablty of RIA formaton or enhancement. The European ntegraton process. Model 8 shows that the sgn of the coeffcents for the EU-27 sample s smlar to the obtaned for the 66-country sample (Model 3), although the coeffcents are lower n magntude. Model 9 shows the results of the geographcal varables. All the geographcal varables are sgnfcant at % and have the expected sgn. Natural and adjacency have a postve-sgned coeffcent, whle the landlocked varable coeffcent s negatve. Model 20 shows that all the soco-poltcal varables are sgnfcant: democracy, the level of economc freedom (property rghts and cvl lbertes) and the common language promote RIA enhancement. Model 2 ncludes an addtonal varable (trade barrers), measurng the blateral weghted tarffs between tradng partners before accessng the EU-27. Unlke the results found n Table 3, the coeffcent of ths varable s negatve, showng that a hgher level of protecton lowers the probablty of a country par beng observed n a hgher category n the European Unon ntegraton process. Model 22 ncludes economc, geographcal and soco-poltcal varables, excludng democracy whch correlates wth the GDP. In ths model, all the varables present the expected sgn and are statstcally sgnfcant. Model 23 ncludes a lagged dependent varable that ndcates the prevous ntegraton level. Ths varable takes nto account the fact that the probablty of reachng an ntegraton level depends on the pont of departure (.e., countres that do not have a prevous agreement do not usually go straght nto a monetary unon). The results show that the probablty of reachng a deeper ntegraton level s hgher f the countres already partcpate n an RIA. Fnally as n BB (2004), the prevous specfcatons assumed that RIAj s ndependent across observatons. Snce ths assumpton s not very realstc and could nfluence the estmaton results, we followed the method proposed by Pesaran (2006) to account for 33

35 nterdependences. Ths method conssts n approxmatng the lnear combnatons of the unobserved factors by cross-secton averages of the explaned and explanatory varables and then runnng standard panel regressons augmented by the cross-secton averages. Ths approach also yelds consstent estmates when the regressors correlate wth the factors. The results are presented n Model 24 and ndcate that nterdependences matter (the added varables are statstcally sgnfcant) but do not alter the sgn of the estmated coeffcents of the varables ncluded n Model 22. As n Secton 5, we evaluate the probablty of the dependent varable havng a partcular value. Then we take the case of Span and France 22 n whch our results for both the 66-country and EU-27 samples ndcate that the hghest probablty s that of the establshment of a monetary unon when soco-poltcal varables were consdered (Model 7 and Model 22, respectvely). Z = β RGDPj + β DRGDPj + β NATURAL + β ADJj + β LANDj β6 LANGj + β7 POLITICAL RIGHTS + β8 CIVIL LIBERTIES = Pr( Y = 0) = Pr( Z 0) = = exp( ( 5.93) ) Pr( Y = 2) = Pr( Z δ ) Pr( Z δ) = + exp( 2.03 ( 2.3) ) + exp( 2.03 ( 5.93) ) Pr( Y = 3) = Pr( Z δ ) Pr( Z δ 2 ) == + exp 2.03 ( 2.00) + exp 2.03 ( 2.3) Pr( Y = 4) = Pr( Z δ ) Pr( Z δ3 ) = + exp( 2.03 (0.25)) + exp( 2.03 ( 2.00) ) Pr( Y = 5) = Pr( δ 5 Z ) = = exp 2.03 (0.25) = ( ) ( ) 3 = 4 = ( ) The results ndcate that the probablty assocated wth level fve s the hghest, whch s consstent wth the fact that France and Span were members of the European Monetary Unon For the EU-27 country sample the probabltes are calculated n the year 999 to be compared to those obtaned n Secton 5 wth the 66-country sample. 34

36 7. Conclusons In ths paper, dscrete choce modellng s used to study the determnants of regonal trade agreements. A bnary probt model and an ordered logt model are estmated, n whch geographcal, economc and soco-poltcal varables are consdered as explanatory varables for RIA formaton. Fve dfferent ntegraton levels are specfed for the dependent varable n the ordered logt estmaton. The results from the probt and ordered logt estmatons show that the probablty of reachng a hgher level of ntegraton ncreases wth ncome level, economc freedom, cultural affntes and remoteness, whereas t decreases wth dstance, protecton levels, ncome dfferences and factor endowment dfferences. Addtonally, although economc and geographcal varables seem to be the most mportant determnants of RIA formaton, the soco-poltcal factors consdered are all statstcally sgnfcant and ther relatve mportance n explanng RIAs enhancement ncreases for hgher ntegraton levels and for natural partners. The margnal effects, calculated for natural and unnatural tradng partners, show that countres n the same contnent (natural partners) wll most probably establsh a sngle market, whereas countres n dfferent contnents (unnatural partners) are most lkely to not sgn any agreement. Ths result s new n the RIA lterature and should be valdated by extendng the sample to nclude more years and countres. The margnal effects also show that some varables, such as remoteness and dfferences n real GDP, have a postve nfluence on the formaton of an RIA, but only for countres n the same contnent and n the early stages of the ntegraton process (PTA, FTA). However, when the categores consdered are hgher ntegraton levels, the effect of these two varables s reversed. The margnal effect of economc freedom s not statstcally sgnfcant for 35

37 unnatural partners n the early stages of the ntegraton process (PTA, FTA). However, t shows that a hgher level of economc freedom has a postve nfluence on the enhancement of a RIA from a customs unon to a sngle market and from a sngle market to a monetary unon. The estmaton of a trade equaton, that consders the formaton of RIAs as an endogenously determned explanatory varable, remans an ssue for further research. Acknowledgments We would lke to thank Jeffrey Bergstrand for hs helpful comments, and also partcpants n the European Trade Study Group conference held n Dubln and n the Atlantc Economc Conference held n New York. Fnancal support from the Spansh Mnstry of Publc Works, Generaltat Valencana (Regonal Valencan Government), and the Spansh Mnstry of Educaton s gratefully acknowledged (P2/08, ACOMP07/02 and SEJ ). References Amemya, T. (975): Qualtatve Response Models. Ann. Econ. Socal Measurement 4, Baer, S.L. and Bergstrand, J.H. (2004): Economc Determnants of Free Trade Agreements. Journal of Internatonal Economcs 64 (), Baer, S.L. and Bergstrand, J.H (2005): Do free trade agreements actually ncrease members nternatonal trade?, Workng Paper , Federal Reserve Bank of Atlanta. Ben-Akva, M., and Lerman, S. (985): Dscrete Choce Analyss. MIT Press, Cambrdge, Massachusetts. Brant, R. (990): Assessng Proportonalty n the Proportonal Odds Model for Ordnal Logstc Regresson. Bometrcs 46,

38 Campbell, B. (2005): The Case Aganst Contnental Deep Integraton. Paper presented to the Centre for Trade Polcy and Law conference Optons for Canada-US Economc Relatons n the 2st Century, November 4, Central Intellgence Agency CIA (2003): The World Factbook. From Endoh, M. (2006): Qualty of Governance and the Formaton of Preferental Trade Agreements, Revew of Internatonal Economcs 4 (5), Frankel, J.A., Sten, E. and We, S.-J. (995): Tradng Blocs and the Amercas: the natural, the unnatural, and the super-natural. Journal of Development Economcs 47 (), Frankel, J.A., Sten, E. and We, S.-J. (996): Regonal tradng arrangements: natural or supernatural. Amercan Economc Revew 86 (2), Frankel, J.A., Sten, E. and We, S.-J. (998): Contnental tradng blocs: are they natural or supernatural. In: Frankel, J.A., Edtor,, 998. The Regonalzaton of the World Economy, Unversty of Chcago Press, Chcago, 9 3. Great crcle dstances (2003): From Henen, T. (993): Dscrete Latent Varable Models. Tlburg Unversty Press. The Netherlands. Koppelman, F.S. and Wen, C. (998): Alternatve nested logt models: Structure, propertes and estmaton. Transportaton Research 32 (3), Magee, C. (2003): Endogenous preferental trade agreements: An emprcal analyss. Contrbutons to Economc Analyss & Polcy 2 (), -7. Mansfeld, E. D., Mlner, H. V. and Rosendorff, B. P. (2002): Why Democraces Cooperate More: Electoral Control and Internatonal Trade Agreements, Internatonal Organzaton 56,

39 McFadden, D. and Tran, K. (2000): Mxed MNL models for dscrete response. Journal of Appled Econometrcs 5 (5), McKnnon, R. (979): Money n Internatonal Exchange: The Convertble Currency System. Oxford. Oxford Unversty Press. Mles, M. A., Feulner Jr., E. J., O Grady M. A. and Eras A. I (2004): Index of Economc Freedom. The Hertage Foundaton. Penn World Tables (2005): From Schff, M. (999), Wll the Real Natural Tradng Partner Please Stand Up? World Bank, Development Research Department Workng Paper. Statstcs Canada (200): World Trade Analyzer. The Internatonal Trade Dvson of Statstcs of Canada. The Freedom House (2005): Democracy s Century. A Survey of Global Poltcal Change n the 20 th Century. From Vcard, V. (2006): "Trade, Conflcts, and Poltcal Integraton: the Regonal Interplays," CESfo Workng Paper Seres CESfo Workng Paper No., CESfo GmbH. Vner, J., (950): The Customs Unon Issue, Carnege Endowment for Internatonal Peace, New York. World Bank (200): World Development Indcators, Washngton. World Bank (2005): From World Trade Organzaton (995): Regonalsm and the World Tradng System. Geneva. World Trade Organzaton (2005): From Wu, J. P.(2004): Measurng and explanng levels of regonal economc ntegraton. Workng Paper B2/2004. Centre for European Integraton Studes. From Wyplosz, C. (2006): Deep Economc Integraton: Is Europe a Blueprnt?, Asan 38

40 Economc Polcy Revew, Table. Probt results for the probablty of RIA formaton. Model Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 7. a Constant 4.07*** 2.07*** 2.3***.63** 2.35* 2.32* 2.34* (7.3) (6.34) (3.42) (2.4) (.88) (.87) (.88) (-.03) NATURAL 0.56*** 0.35*** 0.56*** 0.54*** 0.85*** 0.85*** 0.86***.9*** (20.3) (9.55) (.37) (0.64) (8.5) (8.53) (8.5) (32.63) REMOTE - 0.6*** 0.4*** 0.4*** 0.6*** 0.6*** 0.6*** - (9.35) (6.65) (6.64) (4.40) (4.44) (4.32) RGDP *** 0.06*** 0.09*** 0.09*** 0.09*** 0.*** (3.8) (5.75) (6.4) (6.47) (6.44) (3.87) DRGDP *** -0.5*** -0.5*** -0.5*** *** (-9.47) (-5.75) (-5.83) (-5.7) (-3.47) DKL ** *** - (-2.2) (-.38) (-6.99) SQDKL * - (0.62) (-.7) Pseudo R McFadden s R Log Lkelhood Number of observatons Notes: ***, **, * ndcate sgnfcance at %, 5% and 0%, respectvely. Z-statstcs are n brackets. The dependent varable s a bnary dscrete varable that takes the value of when tradng partners were ntegrated nto a PTA, FTA, CU, SM and MU n 999, and 0 otherwse. The Huber/Whte/sandwch estmator of varance s used nstead of the tradtonal calculaton, therefore the estmaton uses heteroscedastcty-consstent standard errors. a Model 7. was estmated wth an alternatve data set for a cross-secton of 72 countres n 998 from Magee (2003). 39

41 Table 2. Probt results for the probablty of RIA formaton: Extended model. Model 5 Model 8 Model 9 Model 0 Model Model 2 RGDP 0.09*** *** *** (6.40) (4.82) (4.5) DRGDP -0.5*** *** *** (-5.75) (-5.64) (-5.78) DKL -0.09** ** *** (-2.2) (-2.96) (-3.96) NATURAL 0.85*** 0.3*** 0.30*** 0.86*** *** (8.5) (7.99) (7.65) (8.76) (7.70) REMOTE 0.6*** 0.6*** 0.4*** 0.2*** *** (4.40) (9.24) (7.97) (3.3) (5.50) ADJACENCY *** 0.69*** (5.03) (4.98) (0.2) (-0.33) LANDLOCKED *** -0.49*** *** (-5.43) (-5.75) (0.25) (4.34) LANDLOCKED*REMOTE *** 0.7** - - (3.07) (2.7) LANGUAGE *** - (5.47) DEMOCRACY *** 0.65*** (0.96) (5.55) ECONOMIC FREEDOM ** -0.56** (-2.47) (-.97) TRADE BARRIERS *** 0.54*** (5.06) (4.89) NATURAL*LANGUAGE *** (2.96) Constant 2.346*.777***.70*** 2.062* -.43***.047 (.88) (5.) (4.84) (.7) (-7.45) (0.64) McFadden s R Log Lkelhood Number of observatons Notes: ***, **, * ndcate sgnfcance at %, 5% and 0%, respectvely. Z-statstcs are n brackets. The dependent varable s a bnary dscrete varable that takes the value of when tradng partners were ntegrated nto a PTA, FTA, CU, SM and MU n 999, and 0 otherwse. The Huber/Whte/sandwch estmator of varance s used nstead of the tradtonal calculaton; therefore the estmaton uses heteroscedastcty-consstent standard errors. 40

42 Table 3. Ordered logt results for the probablty of RIA formaton or enhancement. Economc varables RGDP DRGDP DKL Model 3 Model 4 Model 5 Model 6 Model 7 0.8*** (6.89) -0.7*** (-3.77) -0.26*** (-3.34) *** (3.45) -0.3*** (-6.4) -0.30*** (-3.25) Geographcal varables NATURAL.49*** 0.84*** 0.83*** 2.3*** - (0.95) (2.83) (2.54) (8.26) REMOTE 0.3*** 0.24*** 0.23*** -2.2** - (6.28) (9.35) (8.22) (-2.40) ADJACENCY *** 0.47*** (2.87) (2.79) (-0.24) LANDLOCKED *** -0.94*** (-5.92) (-5.46) (0.79) LANDLOCKED*REMOTE 0.4** 0.5 (2.36) (.48) NATURAL*REMOTE -0.32*** (-2.89) Soco-poltcal varables LANGUAGE *** -.09 (4.73) (-0.43) DEMOCRACY *** 0.49** (.0) (2.32) ECONOMIC FREEDOM *** -.05** (-2.65) (-2.23) TRADE BARRIERS 0.20** 0.4 (2.46) (0.76) NATURAL*LANGUAGE 0.05 (0.8) Cut Cut Cut Cut Cut McFadden s R Log lkelhood Exp (log lkelhood / observatons) Akake Info Crteron (AIC) Number of observatons Notes: ***, **, * ndcate sgnfcance at %, 5% and 0%, respectvely. Z-statstcs are n brackets. The dependent varable s a dscrete varable that takes the value of, 2, 3, 4 and 5 when tradng partners were ntegrated respectvely nto a PTA, FTA, CU, SM and MU n 999, and 0 otherwse. The Huber/Whte/sandwch estmator of varance s used nstead of the tradtonal calculaton, therefore the estmaton uses heteroscedastcty-consstent standard errors. Blateral trade, exporter s and mporter s trade barrers and economc freedom are shown n natural logarthms. 4

43 Table 4. Odds ratos for the ordered logt. Economc varables RGDP DRGDP DKL Geographcal varables NATURAL REMOTE ADJACENCY LANDLOCKED LANDLOCKED*REMOTE NATURAL*REMOTE Model 3 Model 4 Model 5 Model 6 Model 7 coef 0.8*** *** e^b e^bstdx coef -0.7*** *** e^b e^bstdx coef -0.26*** *** e^b e^bstdx coef.49*** 0.84*** 0.84*** - 2.3*** e^b e^bstdx coef 0.3*** 0.24*** 0.22*** ** e^b e^bstdx coef *** 0.47*** e^b e^bstdx coef *** -0.94*** e^b e^bstdx coef ** e^b e^bstdx coef -0.32*** e^b 0.73 e^bstdx 0.0 Soco-poltcal varables LANGUAGE coef *** -.09 e^b e^bstdx DEMOCRACY coef *** 0.49** e^b e^bstdx ECONOMIC FREEDOM coef *** -.05** e^b e^bstdx TRADE BARRIERS coef ** 0.4 e^b e^bstdx NATURAL*LANGUAGE coef 0.05 e^b.05 e^bstdx.8 Notes: ***, **, * ndcate sgnfcance at %, 5% and 0%, respectvely. Odd ratos are e^b and e^bstdx. e^b = exp(b) = factor change n odds for unt ncrease n X; e^bstdx = exp(b*sd of X) = change n odds for SD ncrease n X. The dependent varable s a dscrete varable that takes the value of, 2, 3, 4 and 5 when tradng partners were ntegrated respectvely nto a PTA, FTA, CU, SM and MU n 999, and 0 otherwse. The Huber/Whte/sandwch estmator of varance s used nstead of the tradtonal calculaton, therefore the estmaton uses heteroscedastcty-consstent standard errors. Exporter s and mporter s trade barrers and economc freedom are shown n natural logarthms. 42

44 Table 5. Response probabltes for natural and unnatural tradng partners n Model 2 (evaluated at the mean level of remote and at remote = 0). Varable Y = Pr (RIA = natural partners) = dy/dx z-statstcs 95% confdence nterval NATURAL 0.073** REMOTE -.598*** RGDP 0.00*** DRGDP *** DKL ECONOMIC FREEDOM *** TRADE BARRIERS 0.2* Y = Pr (RIA = unnatural partners) = dy/dx z-statstcs 95% confdence nterval NATURAL 0.8*** RGDP DRGDP -0.00** DKL -0.03*** ECONOMIC FREEDOM TRADE BARRIERS 0.27*** Notes: ***, **, * ndcate sgnfcance at %, 5% and 0%, respectvely. 43

45 Table 6. Response probabltes for natural and unnatural tradng partners n Model 7 (evaluated at the mean level of remote and at remote = 0). Y = Pr (Preferental Trade Agreement natural partners) = 0.67 Varable dy/dx z-statstcs 95% confdence nterval NATURAL REMOTE 3.373*** RGDP DRGDP 0.028*** DKL ECONOMIC FREEDOM 0.762*** TRADE BARRIERS *** Y = Pr (Preferental Trade Agreement unnatural partners) = 0.03 dy/dx z-statstcs 95% confdence nterval NATURAL 0.07*** RGDP DRGDP ** DKL -0.09*** ECONOMIC FREEDOM TRADE BARRIERS 0.052*** Y = Pr (Free Trade Agreement natural partners) = 0.89 dy/dx z-statstcs 95% confdence nterval NATURAL REMOTE.587*** RGDP DRGDP 0.03*** DKL ECONOMIC FREEDOM 0.359*** TRADE BARRIERS -0.85*** Y = Pr (Free Trade Agreement unnatural partners) = dy/dx z-statstcs 95% confdence nterval NATURAL 0.058*** RGDP DRGDP ** DKL -0.05*** ECONOMIC FREEDOM TRADE BARRIERS 0.043*** Y = Pr (Customs Unon natural partners) = 0.25 dy/dx z-statstcs 95% confdence nterval NATURAL REMOTE RGDP DRGDP DKL ECONOMIC FREEDOM TRADE BARRIERS Y = Pr (Sngle Market natural partners) = 0.38 dy/dx z-statstcs 95% confdence nterval NATURAL REMOTE -7.04*** RGDP

46 DRGDP *** DKL ECONOMIC FREEDOM -.606*** TRADE BARRIERS 0.828*** Y = Pr (Monetary Unon natural partners) = dy/dx z-statstcs 95% confdence nterval NATURAL REMOTE *** RGDP DRGDP *** DKL ECONOMIC FREEDOM -0.94*** TRADE BARRIERS 0.00*** Notes: ***, **, * ndcate sgnfcance at %, 5% and 0%, respectvely. 45

47 Table 7. Ordered logt results for the probablty of RIA formaton or enhancement. The European ntegraton process. Economc varables RGDP DRGDP Geographcal varables NATURAL Model 8 Model 9 Model 20 Model 2 Model 22 Model 23 Model *** (6.53) -0.09*** (-6.3) 0.8*** (6.40) ADJACENCY - LANDLOCKED - Soco-poltcal varables *** (6.38) 0.72*** (7.49) -.48*** (-28.67) LANGUAGE - - DEMOCRACY - - POLITICAL RIGHTS - - CIVIL LIBERTIES TRADE BARRIERS RIAj LAGGED *** (8.90) 0.96 *** (8.6) -.65 *** (-.34) -2.2*** (-26.7).8*** (8.97).24*** (20.22) -.76*** (-9.85) -.47*** (-7.56) -0.50*** (-23.93) 0.0*** (0.52) -0.*** (-7.30) 0.6*** (4.48) 0.6*** (6.3) -0.82*** (-5.48).29*** (7.9) *** (-.64) -2.23*** (-3.97) *** (3.03) -0.0*** (-4.24) 0.2** (2.35) 0.24* (.92 ) 0.2** (2.40) 0.*** (0.50) -0.2*** (-7.62) 0.6*** (4.74) 0.6*** (6.02) -.03*** (-7.87) *** (-.28) (0.55) *** (-3.03) 0.09 (0.99) *** (37.23) -2.2*** (-2.79) -.50*** (-9.0) AVERAGE POLITICAL RIGHTS 2.** (2.26) AVERAGE CIVIL LIBERTIES.53*** (6.76) AVERAGE RIAj.77*** (20.07) Cut Cut Cut Cut McFadden s R Log lkelhood Exp (log lkelhood / observatons) Number of observatons Notes: ***, **, * ndcate sgnfcance at %, 5% and 0%, respectvely. Z-statstcs are n brackets. The dependent varable s a dscrete varable that takes the value of, 2, 3, 4 and 5 when EU-27 tradng partners were ntegrated respectvely nto a PTA, FTA, CU, SM and MU from 999 to 2007, 23 and 0 otherwse. The Huber/Whte/sandwch estmator of varance s used nstead of the tradtonal calculaton; therefore the estmaton uses heteroscedastcty-consstent standard errors. Blateral trade, trade barrers, democracy, cvl lbertes and poltcal rghts are shown n natural logarthms. 23 There are not cases of PTA. 46

48 APPENDIX A Table A.. 66 country sample. Algera Argentna Australa Austra Belgum-Luxembourg Bolva Brazl Bulgara Canada Chle Chna Colomba Costa Rca Croata Cyprus Czech Republc Denmark Domncan Republc Ecuador Egypt Fnland France Germany Ghana Greece Honduras Hong Kong Iceland Inda Ireland Israel Italy Jamaca Japan Kenya Mexco Mozambque Nepal Netherlands Ncaragua Norway Panama Pakstan Paraguay Peru Poland Portugal El Salvador Senegal Sngapore Slovaka South Afrca South Korea Span Sudan Sweden Swtzerland Syran Arab Republc Tanzana Trndad Tobago Turkey UK Uruguay USA Venezuela 47