Predicting the Effects of NAFTA: Now We Can Do It Better!

Transcription

1 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, Predctng the Effects of NAFTA: Now We Can Do It Better! Serge Shkher * Suffolk Unversty Abstract The North Amercan Free Trade Agreement (NAFTA) s one of the most nfluental events n the North Amercan economc hstory, but the economc forecasts regardng ts effects on trade turned out to be vastly ncorrect. Ths paper suggests and evaluates a new model for forecastng the effects of trade lberalzatons that uses the framework to explan ntra-ndustry trade nstead of the usual approach. The performance of the model s evaluated by usng t to forecast the effects of NAFTA from the pont of vew of The predctons of the model are compared wth the post-nafta data and the forecasts of the exstng models. The results show that the new model s able to predct the effects of NAFTA notceably better than the exstng trade models. The paper analyzes the reasons for ths dfference. Keywords: trade lberalzaton, NAFTA, trade forecastng, computable models, heterogeneous producers JEL codes: F1, F13, F15, F17 1. Introducton The goal of NAFTA, whch went nto effect on January 1, 1994 was a gradual reducton and then elmnaton of all tarff and many non-tarff barrers to trade between the U.S., Canada, and Mexco. Beng one of the most sgnfcant events n recent U.S. economc hstory, ts merts were hotly debated, wth many concerned partes partcpatng n the dscusson. Predctons ranged from sgnfcant beneft to the U.S. economy to sgnfcant harm to t. At the end, many predctons turned out to be ncorrect. Unfortunately, many of the forecasts made by economc models also turned out to be ncorrect. Quanttatve economc analyss was extensvely used before the ratfcaton of NAFTA to make predctons about ts effects. The forecasts were made usng computable general equlbrum (CGE) models that reled on the Armngton (1969) assumpton to explan two-way trade between countres and home bas n consumpton. The models were generally smlar, wth the type of competton n the goods market beng the bggest dfference. Ther predctons ponted to lttle effect on trade, output, and employment n the U.S., and moderate effects on Mexco. 1 In actualty, NAFTA had a sgnfcant effect on trade between ts members and a small-tomoderate effect on ther ncomes and employment (Burfsher, Robnson, and Therfelder, 2001; Anderson and van Wncoop, 2002; Romals, 2007). Between 1993 and 2008, the total NAFTA trade relatve to the total NAFTA GDP grew 48%, the fracton of U.S. ncome that was spent on

2 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, Mexcan goods grew 121%, and the fracton of Mexcan goods n total Canadan mports went up 74%. We can now say that the CGE models sgnfcantly underpredcted the effect of NAFTA on trade. In addton, the ndustry-level changes n blateral trade they forecasted turned out to have lttle correlaton wth the actual post-nafta changes. These results cast doubt on the ablty of the exstng models of trade to accurately forecast the effects of trade polces. It s mportant that the models that are used to predct the effects of government polces are transparent and subected to thorough testng and evaluaton. However, t s dffcult to fully evaluate the models of trade used to predct the effects of NAFTA because ther equatons and data are not publcly avalable and, therefore, replcaton s not possble. Ths paper proposes a new model for forecastng the effects of trade lberalzatons, whch are of great mportance to polcy-makers. Instead of usng the Armngton assumpton, ths model employs the framework of Eaton and Kortum (2002) to motvate ntra-ndustry trade. The model s called the HPPC model, whch stands for heterogeneous producers, perfect competton. The equatons and data for the HPPC model are posted onlne and ths paper tests ts ablty to accurately forecast the effects of trade lberalzatons. The performance of the model s evaluated by usng t to forecast the effects of NAFTA from the pont of vew of The predctons are compared wth the post-nafta data and forecasts of other CGE models. The comparsons show that the changes n NAFTA trade predcted by the HPPC model are close to the actual post-nafta changes and much closer to them than the predctons of other CGE models. The HPPC model s also able to better explan the varaton of the predcted changes across countres and ndustres. The paper studes the reasons for better performance of the HPPC model vs. the prevous CGE models. It focuses on the Brown, Deardorff, and Stern (1992a) model as a representatve CGE model. The paper fnds that the BDS model, lke other CGE models, used low Armngton elastctes of substtuton, whch resulted n small forecasted changed n trade overall. The BDS model also suffered from usng naccurate estmates of the polcy barrers to be removed by NAFTA. The paper s organzed as follows. Secton 2 descrbes the general equlbrum model of trade wth heterogeneous producers. Secton 3 evaluates the performance of ths model by comparng ts predctons to the data and predctons of other models. Secton 4 analyzes the results and Secton 5 concludes. 2. Model wth heterogeneous producers Ths secton presents an alternatve to the currently avalable computable models of trade. The new model s formally descrbed n Secton 2.1 and the parametrzaton procedure s explaned n Secton 2.2. The NAFTA smulaton results are presented n Secton 3.

3 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, The model s based on the neoclasscal assumptons of multple ndustres, constant returns to scale, perfectly compettve markets, and several factors that are moble across ndustres. Each ndustry s characterzed by a partcular level of technology, set of factor ntenstes, and a demand functon. Countres dffer n ther factor endowments. 2 In all of these aspects, the model s smlar to the currently avalable computable models of trade. However, whle other models use the Armngton assumpton to explan two-way trade between countres, ths model reles on the Eaton-Kortum (EK) framework at the ndustry level. Wthn each ndustry, there s a contnuum of goods produced wth dfferent productvtes. Producton of each good has constant returns to scale, and goods are prced at margnal cost. Snce the heterogeneous producers and perfect competton are the defnng characterstcs of ths model, t wll be referred to n ths paper as the HPPC model. 3 The use of the Eaton and Kortum (2002) framework nstead of the Armngton (1969) approach has several key mplcatons. The goods are dfferentated by ther features, not by ther country of orgn. The home bas n consumpton and cross-country prce dfferentals are explaned by trade costs rather than the demand-sde parameters. Note that the use of the Eaton-Kortum methodology nstead of the Armngton assumpton by tself does not mprove the forecastng abltes of a model snce the key equatons are the same n both models (as explaned n Eaton and Kortum (2002)). 2.1 Descrpton of the model There are N countres, ndexed by or n, J ndustres, ndexed by or m, and two factors of producton: captal and labor. The ndustry cost functons has a Cobb-Douglas form: c 1 r w, (1) where r s the rate of return on captal, w s the wage, s the prce of ntermedate nputs used n ndustry of country, and 0, 0, and 1 0 are the captal, labor, and ntermedate goods shares, respectvely. The prce of nputs functon of ndustry prces: J s the Cobb-Douglas m, (2) m1 p m J where m 0 s the share of ndustry m goods n the nput of ndustry, such that m 1 m 1,. Intra-ndustry producton, trade, and prces are modeled usng the framework of Eaton and Kortum (2002). In each ndustry, there s a contnuum of goods, wth each good ndexed on the nterval [ 0,1] by l and produced wth ts own productvty. Productvtes z (l) are the result of the R&D process and probablstc, drawn ndependently from the Fréchet dstrbuton wth cdf

4 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, T z F ( z) e, where T 0 and 1 are the parameters. 4 Consumers have CES preferences over the contnuum of goods wthn an ndustry wth the elastcty of substtuton 0. The prce of good l of ndustry produced n country and delvered to country n s p ( l) cd / z ( l), where d s the Samuelson's ceberg transportaton cost of delverng ndustry goods from country to country n. 5 In country n, consumers buy from the lowestcost suppler, so the prce of good l n country n s p ( l) mn p ( l), 1,..., N. For the purposes of ths paper, t s necessary to separate the total trade costs d nto polcyrelated trade costs and non-polcy-related trade costs t. The polcy-related trade barrers (tarffs and tarff equvalents of NTBs) can be mposed on the f.o.b. or c..f. values of goods. If they are mposed on the f.o.b. values, then d 1. If they are mposed on the c..f. values, then d 1 t t 1. The NAFTA smulatons n ths paper assume that the polcyrelated barrers are mposed on the f.o.b. values, whch corresponds to the practce n the Unted States, Canada, and Mexco (for NAFTA countres). T cd p s G ( p) 1 F c d / p1 e. The N p N p s G ( p) G ( p) 1 e, where T c d The dstrbuton (cdf) of prces dstrbuton of 1 summarzes technology, nput costs, and transport costs around the world. The exact prce ndex 1 for the wthn-ndustry CES obectve functon s then p p ( l) dl 0 1 1/ 1 1 1/ 1 1/ 1/ p dg ( p E P, where 1 1 /. 6 Ths prce ndex can also be wrtten as ) 0 1/ d c, 1 p 1/ 1 N p T (3) 1 where s a constant. Pluggng ths prce ndex nto (1), the cost equaton becomes c r w m 1 J N Tnmdnmcnm. (4) m1 n1 To derve the ndustry-level blateral trade flows, we note that the probablty that a producer from country has the lowest prce n country n for good l s Pr p ( l) mn pns ( l); s s 1 Gns ( p) dg ( p) T cd / p 0. Snce there s a contnuum of goods on the nterval [ 0,1], ths probablty s also the fracton of ndustry goods that country n buys from. It s also the fracton of n 's expendture spent on ndustry goods from or X / X (ths s true because condtonal on the fact that country actually

5 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, supples a partcular good, the dstrbuton of the prce of ths good s the same regardless of the source ). So, the ndustry-level blateral trade s gven by X X T cd p, (5) where X s the spendng of country n on ndustry goods produced n country and X s the total spendng n country n on ndustry goods. Parameter T represents ndustry-level productvty and, therefore, determnes comparatve advantage across ndustres. For example, country n has a comparatve advantage n ndustry f T / T T / T. 7 Parameter determnes the comparatve advantage across goods wthn an nm m ndustry. Lower value of means more dsperson of productvtes among producers, leadng to stronger forces of wthn-ndustry comparatve advantage. Industry output Q s determned as follows. The goods market clearng equaton s Q N n1 X N n1 N X Z C, (6) n1 where Z and C are amounts spent by country n on ndustry 's ntermedate and consumpton goods, respectvely. The spendng on ntermedate goods s Z m wnlnm 1 m m / m gm, where L nm s the stock of labor employed n ndustry m of country n. 8 Consumer preferences are two-ter: Cobb-Douglas across ndustres and, as prevously mentoned, CES across goods wthn each ndustry. Therefore, C Yn, where Y n s the total ncome (GDP) n country n and 0 s a parameter of the model. 9 Pluggng the expressons for ntermedate and consumpton spendng nto (6), the output equaton becomes Q N n1 J m 1m m wn Lnm Yn. (7) m1 m Snce producton s Cobb-Douglas, ndustry factor employments are gven by K Q / r and L Q / w. Factors of producton can be freely and nstantaneously moved across ndustres wthn a country, subect to the constrants the country factor stocks, whch are fxed. K K J J 1 and 1 L L, where K and L are Due to data lmtatons, only the manufacturng ndustres are modeled. The nonmanufacturng sector's prce ndex s normalzed to 1 and ts purchases of the manufacturng ntermedates are

6 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, treated as fnal consumpton. Country ncome Y s the sum of the manufacturng ncome and nonmanufacturng ncome O Y : M Y Y M O O Y Y w L r K Y. (8) The manufacturng s assumed to be a constant proporton of the GDP, so that Y O Y, where 0 s a parameter of the model. Factor stocks K and L are specfc to manufacturng. Captal and labor are not moble between the manufacturng and nonmanufacturng sectors. The model s gven by equatons (3)-(5), (7)-(8), and the four factor employment and factor clearng equatons. Model parameters are,, m,,, t,, T, K, L, and. The model solves for all other varables ncludng all prces, ndustry factor employments, output, and trade Assgnng parameter values The model s parametrzed usng 1989 data for 8 two-dgt manufacturng ndustres n 19 OECD countres. 11 The ncluded countres and ndustres can be seen n Table 1. The values for parameters,,, and are taken from the data or lterature. The value of the technology m dstrbuton parameter s taken from Eaton and Kortum (2002), where t s estmated to be 8.28 usng trade and prce data. 12 Sources for all data are descrbed n the appendx. Estmaton of the trade costs s dscussed n Sectons and The values for parameters, T, K, L, and are obtaned by fttng a subset of the model to data, whch s descrbed n Secton Total trade barrers Ths secton estmates total trade costs d. The next secton wll descrbe the magntudes of polcy-related barrers obtaned from data. The non-polcy-related barrers t are calculated as the dfference between d and. Total blateral trade barrers d are estmated by applyng the approach of Eaton and Kortum (2002) at the ndustry level. The rato of n 's spendng on 's goods to ts spendng on ts domestcally-made goods s obtaned from equaton (5): n X X n T T d c c. (9) To relate the unobservable trade cost to the observable country-par characterstcs, the followng trade cost functon s used:

7 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, phys log d d b l f m, (10) where k d 1,...,6 phys k k s the effect of physcal dstance lyng n the k th nterval 13, b s the effect of common border, l s the effect of common language, f s the effect of belongng to the same free trade area, m n s the overall destnaton effect, and n s the sum of transport costs that are due to all other factors. Then, the gravty-lke estmatng equaton s obtaned by takng logs of both sdes of (9): log X X n phys exp mp d b l f D D, (11) k where D exp T c s the exporter dummy, mp D m log T c s the mporter dummy. 14,15 The average (across country pars and ndustres) estmated transport cost s Ths transport cost ncludes all costs necessary to move goods between countres, such as freght, nsurance, tarffs, non-tarff barrers, and theft n transt. Trade costs vary across country pars and ndustres. For example, the Machnery and Textle products are typcally cheaper to move between countres than the Wood and Food products Polcy-related trade barrers In order to smulate the effects of NAFTA, t s necessary to know the extent of the polcyrelated trade barrers before ts mplementaton. Ths ncludes both tarff and non-tarff barrers, the latter expressed n terms of ad-valorem tarff equvalents. Obtanng such data s not trval. The man source for the magntudes of the pre-nafta polcy-related trade barrers used n ths paper s Ncta and Olarreaga (2007) that has nformaton on both tarff and non-tarff barrers. Ths paper uses 1989 appled tarff data for Canada and the Unted States and 1991 data for Mexco, whch s the closest avalable year. 16,17 Compared to tarffs, the tarff equvalents of non-tarff barrers are much harder to collect and estmate. Consequently, there are fewer sources of ths data. The earlest years for whch the advalorem equvalents of NTBs for Canada, Mexco, and the U.S. are avalable n Ncta and Olarreaga (2007) are In the absence of other data, ths paper uses the nformaton from to proxy for 1989 magntudes. Due to NAFTA's reducton of NTBs, the average levels of NTBs for the NAFTA countres have probably fallen between 1989 and Therefore, usng the levels results n smaller forecasted growth n trade due to NAFTA. 18 However, the pattern of NTBs across ndustres s less lkely to have changed. The tarffs and tarff equvalents of non-tarff barrers used n the smulatons are presented n Tables 2(a)-(c). The average tarffs mposed by Canada, Mexco, and the Unted States on manufacturng goods, shown n the last column, were about 8.5%, 13.7%, and 4.7% respectvely,

8 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, whle the tarff equvalents of NTBs were 3.3%, 17.7%, and 4.1%. The total polcy-related barrers, therefore, were 11.8% n Canada, 31.4% n Mexco, and 8.8% n the Unted States. Tables 2(a)-(c) also report tarffs and tarff equvalents of NTBs by ndustry. There s notceable heterogenety of protecton levels across ndustres. The Textle ndustry s one of the most protected ndustres n all three countres wth total polcy-related barrers rangng between 16% n the Unted States to 40% n Mexco. The Paper ndustry s the least protected ndustry n all three countres wth barrers rangng between 1.29% n the Unted States to 17% n Mexco. The rankng of ndustres accordng to total protecton levels vared across countres. For example, the Wood ndustry s farly heavly protected n Canada, but relatvely less protected n Mexco and the Unted States. The same s true of the Metals ndustry. On the other hand, the Nonmetals ndustry has less protecton relatve to other ndustres n Canada than t does n Mexco or the Unted States. The prevalent type of polcy-related protecton also vared across ndustres and countres. For example, n the Unted States, the Textle and Nonmetals ndustres are protected mostly by tarffs, whle the Food ndustry s protected mostly by NTBs. In Canada, the Chemcals and Nonmetals ndustres are protected mostly by tarffs, whle the Metals ndustry s protected mostly by NTBs Technology and other ftted parameters The parameters, T, K, L, and are obtaned by fttng a subset of the smulaton model, together wth a long-run equlbrum condton, to domestc data. 19 The subset of the model ncludes the cost equaton (4), reproduced here: c r w m 1 J N Tnmdnmcnm, (12) m1 n1 and a smplfed verson of the output equaton (ndustry output): Q N X, (13) n1 where mport shares are gven by the followng equaton, derved from equatons (5) and (3): T cd T c d N 1. The values of Q, X, and w are taken from data. Data on the rates of return r s not avalable, so ther values n the base year are approxmated at 20%. 20 The values of trade costs (14)

9 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, d were estmated n the prevous secton. Equatons (12)-(14) are solved to fnd the 1989 values of T nm and c. The system (12)-(14) s exactly dentfed snce there are as many equatons ( 2 NJ ) are unknowns. The ndustry employments of captal and labor are then calculated as L / Q w. 21 The country factor stocks K and K Q / r and L are calculated as the sum of ndustry 1 r K w L / Y, where the factor employments. Nonmanufacturng share s calculated as total ncome Y s taken from data. The taste parameters are calculated as C / Y, where the J consumpton s calculated as C X Z X m m1 km lm Qm The estmated ndustry technology parameters relatve to the Unted States are presented n Table 1. They show that countres have dfferent relatve technologes n dfferent ndustres. These dfferences provde ndustry-level (nter-ndustry) Rcardan comparatve advantages. 3. Evaluatng the predctons of the model The smulaton of NAFTA performed n ths paper entals the removal of all polcy-related trade barrers reported n Table 2(c) between the three NAFTA countres, whle mantanng all other trade barrers t. Ths secton evaluates the accuracy of the model's predctons regardng the changes n total manufacturng trade and trade n ndvdual ndustres. The smulatons results are compared wth the actual post-nafta data and results of several prevous NAFTA smulatons Benchmarks The data aganst whch the results of the smulatons are compared s from The ntal pont (1989) s gven by the year n whch the model was parametrzed. 24 The end year (2008) s the latest year for whch trade data s avalable. NAFTA was mplemented n 1994 and provded for graduate blateral trade barrer reductons between the partcpatng countres. Both the tarff barrers and non-tarff barrers were to be reduced over a perod of tme. The average length of phase-outs for U.S. tarffs was 1.4 years and Mexcan tarffs 5.6 years (Kowalczyk and Davs, 1996). Therefore, by the year 2008, whch s the 15th year of NAFTA mplementaton, the vast maorty of the trade barrers that were to be elmnated under NAFTA had been elmnated. 25 The effects of NAFTA have been prevously forecasted by several teams of researchers. 26 The forecasts employed computable general equlbrum models that utlzed the Armngton assumpton. These models assumed ether constant or ncreasng returns to scale. 27 Assumng IRS resulted n greater predcted effects of NAFTA. Some models had constant captal stock, whle others allowed captal accumulaton. Allowng nternatonal movement of captal typcally caused large nflows of captal nto Mexco.

10 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, The effects of NAFTA were smulated by removng the pre-nafta polcy-related trade barrers. Some studes, such as Brown et al. (1992a) and Roland-Holst et al. (1994), smulated the removal of the non-tarff barrers (NTBs) n addton to the removal of tarffs, whch resulted n greater predcted effects of NAFTA. 28 Very few studes exst that systematcally evaluate the predctons made by economc models regardng NAFTA. Ths s unfortunate, snce NAFTA was an mportant test for general equlbrum models of trade, so t s nterestng to see how they performed. Because the purpose of the CGE models typcally s to make economc forecasts, the qualty of those forecasts s an mportant crteron by whch the models should be udged. Instead, most post-nafta studes focus on analyzng the effects of NAFTA, typcally usng the gravty model rather than a CGE model. 29 These studes generally fnd that NAFTA had a relatvely small effect on employment, prces, and welfare, as pre-nafta studes predcted. They also fnd that NAFTA had a large effect on trade, whch s where the pre-nafta economc forecasts badly stumbled. One paper that evaluates the performance of the pre-nafta forecasts s Kehoe (2005). 30 By systematcally comparng model predctons to data, he fnds that many of the predctons made before NAFTA turned out to be sgnfcantly off. 31 Specfcally, the pre-nafta forecasts sgnfcantly underestmated the effects of NAFTA on trade, sometmes by several orders of magntude. In addton, the models dd poorly n explanng the varaton of changes n trade flows across countres and ndustres. Secton 3.2 wll compare the forecasts of the HPPC model to the forecasts of the Brown- Deardorff-Stern (BDS) and Roland-Holst-Renert-Shells (RRS) models. 32 Secton 4 wll dscuss possble reasons for the dfferences n forecasts. 3.2 Results Table 3 shows the overall effect of NAFTA: change n the share of NAFTA trade n the total trade of the NAFTA countres. 33 The HPPC model predcts that ths share would grow 25.9% whle t actually grew 23.8%. Relatve to the total NAFTA ncome, the predcted growth n NAFTA trade s 62.2% whle the actual growth s 66.5%. Therefore, the HPPC model slghtly overestmates the change n NAFTA trade relatve to the total trade of the NAFTA countres, and underestmates somewhat the change n NAFTA trade relatve to NAFTA ncome. Ths means that the total trade of the NAFTA countres relatve to ther ncome grew more than what the model predcts. Ths could be due to a decrease n non-polcy related trade costs across the world, for example. 34 Table 4 gves a more detaled look at the changes n trade of the NAFTA countres. It shows the actual and predcted percent changes n the total exports and mports of Canada, Mexco, and the Unted States, relatve to ther respectve GDPs. The numbers are also plotted n Fgure 1. The frst column shows the actual changes. Mexcan exports and mports have grown the most, followed by Canada's and then the Unted States'. The changes predcted by the RRS

11 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, and BDS models are many tmes smaller than the actual changes. The RRS model, whether wth constant or ncreasng returns to scale, performs the worst n terms of correlaton wth data. The BDS model performs better, but ts predcted changes n Canadan and Mexcan exports and mports are smaller than the actual changes by an order of magntude. The HPPC model performs the best: ts predcted changes are the closest to the actual, as can easly be seen on Fgure 1. Its predcted changes also exhbt the best correlaton wth the actual changes: Next, we nvestgate the accuracy of the models' forecasts at the ndustry level. Only the HPPC and BDS models are consdered. The BDS model s chosen because t seems to be the bestperformng out of prevous NAFTA smulatons and because of the avalablty of the detaled smulaton results. 35 Tables 5a-c show the actual vs. predcted percent changes n the mport shares for each par of the NAFTA countres by ndustry. The share of country n ndustry mports of country n s X / IM, where IM are the total mports of ndustry goods n country n. Fgure 2 plots the data shown n Table 5. Fgures 2a and 2b plot the changes for the US-Canada and US- Mexco trade, whch together consttute about 99% of NAFTA trade. 36 It can be seen from these fgures that the predctons of the HPPC model are generally close to the actual values, whle the BDS model tends to sgnfcantly underpredct trade changes. The HPPC model s also better able to explan the varaton of changes n trade across ndustres: the correlaton of ts predctons wth data s 0.95, whle for the BDS model t s As an example, consder U.S. mports from Canada and Mexco. The HPPC model correctly predcts that the largest ncreases wll occur n the Textle ndustry, whle the BDS model does not. Fgures 2c and 2d plot the actual vs. predcted changes n mport shares for the Canada-Mexco trade. Trade between these two countres exhbts the largest dscrepances between the actual and predcted changes. Trade between Canada and Mexco trade s small: t consttuted ust under 1% of NAFTA trade n 1989 and about 1.5% n Canadan exports to Mexco were only $400M n 1989 and $3B n I beleve that small trade flows (most lkely done by ust a handful of frms n a few transactons) are more susceptble to data rregulartes than larger trade flows. 37 Very extreme observatons are reported n some ndustres: the Wood ndustry data shows a 1580% ncrease n Mexcan exports to Canada and the Chemcals ndustry data shows a 413% ncrease n Mexcan exports to Canada. 38 The predctons of both the HPPC and BDS models correlate poorly wth the actual changes n trade: the correlaton s 0.08 for the HPPC and for the BDS model. Table 6 shows the correlatons between the actual and predcted changes n mport shares for each par of countres. It also shows the estmated ntercepts and slopes for the regressons of actual on predcted changes. Ideally, we would lke the ntercept to be zero and slope one. The correlaton s a measure of how much of the varaton n the data s explaned by the model. 39 The table shows, for example, that on average the HPPC's estmates of changes n Mexcan mport shares n the U.S. have to be multpled by 0.93 and the product reduced by 15.70

12 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, percentage ponts to match the actual changes n those mport shares. By comparson, BDS model's predcted changes have to be multpled by 2.23 and the product ncreased by percentage ponts to match the actual changes. The correlaton between the actual and predcted changes s 0.98 for the HPPC model and 0.44 for the BDS model. 4. Analyss of the results The prevous secton has shown that there are sgnfcant dspartes between the NAFTA forecasts of the HPPC and the other models. Specfcally, t noted two problems wth the forecasts of the other models: the overall magntude of the predcted changes n trade s too low and the correlaton (across ndustres and country pars) between the predcted and actual changes s poor. Ths secton wll nvestgate the causes for the dfferences n the forecasts. 40 Equaton (14), whch s obtaned by combnng equatons (5) and (3), shows how trade flows are predcted n the HPPC model. Its analogue n the Armngton model s N s1 1 1 p, 1 1 p ns ns (15) where s the (Armngton) elastcty of substtuton between goods sourced from dfferent countres and s the weght placed n the utlty functon of country n on ndustry goods sourced from country (see, for example, Eaton and Kortum (2002) or Ruhl (2008)). In the presence of tarffs, prce p s equal to c 1, where c s the cost of producng country 's varety of good and s the tarff. We can see that parameter n the HPPC model and n the Armngton model are key to determnng how changes n trade costs affect trade flows. The HPPC model sets whle the BDS model sets the Armngton elastcty at around As dscussed n Ruhl (2008), lower values of are estmated by studes of hgh-frequency changes n prces, whle hgher values are produced by cross-country studes and studes of changes n trade polcy. Holdng everythng else equal, usng nstead of 3 results n about tmes greater predcted changes n trade flows. So the use of the low Armngton elastctes by the BDS model can explan ts small forecasted changes n trade flows due to NAFTA. However, snce the BDS model uses very smlar elastctes n dfferent ndustres (they vary between 2. 7 and 3. 0 across ndustres), the values of cannot explan the poor correlaton of the forecasted and actual changes n trade across country pars and ndustres. HPPC model assumes constant n all ndustres. To check the effects of the dfference n magntude between and on the forecasts, I set 3 and re-smulate the effects of NAFTA. 42 The results of ths and other smulatons dscussed n ths secton are shown n Table 7. The columns present varous measures of the relatonshp between the actual and predcted changes n ndustry-level mport shares (excludng Canada-Mexco trade). The measures nclude correlaton, ntercept and slope from the regresson

13 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, of the actual on predcted trade changes, and the average absolute change n mport shares (whch helps to measure the overall magntude of the predcted trade changes). The frst lne of the table shows the results for the orgnal model confguratons and parameter values. It shows that the forecasts of the HPPC model correlate hghly wth the data and are of smlar magntude to the data, as evdenced by the average absolute change and regresson results. 43 The second lne shows that settng 3 results n much smaller predcted changes n trade. The overall magntudes of the forecasted changes n trade n ths case are smlar to those of the BDS model, but the correlaton between the predcted and actual changes s much hgher at 0.87 (vs for the BDS model). So why do the changes n trade flows predcted by the BDS model correlate much more poorly wth the data than the changes n trade flows predcted by the HPPC model? One possble reason s that the BDS model (same as all the Armngton-based models) treats the polcy barrers as beng mposed on the c..f. goods values nstead of ther f.o.b. values, whch results n greater percent changes n trade costs and, therefore, greater forecasted changes n trade. Wth 8. 28, the ncrease s about 50% on average for the HPPC model's NAFTA forecasts. 44 However, the choce of the assumpton does not have a bg mpact on the correlaton between the actual and predcted changes. Though, d / 1 s less correlated wth the actual changes n NAFTA trade than d / 1 t, 0.06 vs. 0.2, the correlaton s low n ether case, meanng that the varaton n trade costs can explan only a fracton of the varaton n the actual trade changes. Also, assumng that polcy barrers are mposed on the c..f. nstead of f.o.b. goods' values n the HPPC model actually slghtly ncreases the correlaton between the forecasted and actual changes n trade (for 3, the ncrease s from 0.87 to 0.93 as shown on lne 3 of Table 7). We should also consder the fact that the BDS and HPPC models use dfferent data on pre- NAFTA polcy barrers, whch can be a reason for ther dfferent forecasts. We note that the tarff rates used by the BDS model (shown n Brown et al. (1992a)) are hghly correlated (0.8) wth the tarff rates used by the HPPC model. Usng the HPPC model wth the BDS tarff levels (keepng 3 and c..f. polcy barrers) reduces the correlaton wth the data from 0.93 to 0.88, as shown on lne 4 of Table 7. The tarff equvalents of non-tarff barrers are solved for endogenously by the BDS model and are not shown. 45 However, snce the BDS model only consders NTBs mposed by the U.S. on Mexcan Food and Textle products, I wll smply omt these two trade flows from the analyss to gauge the mpact of the BDS's NTB treatment. The correlaton between the remanng 30 predcted and actual trade flows for the BDS model s 0.44, whch s an ncrease over 0.32 for all 32 trade flows. So, BDS's treatment of the NTBs may be contrbutng to the poor qualty of ts forecasts. Smulatng the effects of NAFTA usng the HPPC model wth 3, polcy-related trade costs mposed on c..f. values, BDS's tarff rates, and no NTB removal, the correlaton between the predcted and actual changes for the 30 trade flows falls to 0.74, as shown on the last lne of

14 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, Table Ths s not as good as usng HPPC's own parameter values (0.95), but stll substantally better than the BDS's result of However, the gap between the correlatons n ths case s about a half of the gap that exsts when HPPC's own parameter values are used (the gap s 0.3 vs. 0.63). Table 7 shows that of all parameter values, BDS model's treatment of NTBs contrbute the most to the poor qualty of ts forecasts (t explans more than 3/4 of the change n the correlaton gap). The values of the Armngton elastctes and tarff levels used by the BDS model also deterorate HPPC model's forecasts, but to a much smaller degree (n terms of correlaton, not magntudes). The rest of the dfference n the performance of the HPPC and BDS models must be explaned by the values of other parameters, such as the nput-output shares. Unfortunately, the values of these parameters are not publshed by the authors of the BDS model. Therefore, a comparson of ther values n the BDS and HPPC models s not possble. 47 In addton to the model propertes and parameter values dscussed n ths secton, the gap between the forecasted and actual changes n trade could have been caused by varous post- NAFTA events not accounted for by the smulaton models. Technologcal change s one such possble event (Kehoe, 2005). However, technologcal change affects both trade and GDP, so ts effect on the trade-to-gdp ratos s lkely to be small. In addton, whle by most estmates there was some postve technologcal change n Mexco n the late 1990s, t was farly small and not larger than the contemporary technologcal change n the Unted States. So the technology of Mexco relatve to the U.S. has not changed notceably durng that tme. There may have been dfferences of technologcal growth across ndustres. It s unknown how much these dfferences have contrbuted to the varaton between predcted and actual changes n trade across ndustres. The devaluaton of Mexcan peso n 1995 s another post-nafta event that s not part of the smulatons. However, the effects of the peso devaluaton have most lkely dsspated by the year Concluson Beng a maor event n recent North Amercan economc hstory, NAFTA s a natural experment that s useful for evaluatng models of trade. Unfortunately, the currently avalable computable general equlbrum models of trade have not done a good ob forecastng the effects of NAFTA. The changes n trade flows predcted by those models are much smaller than the actual changes that occurred after NAFTA. In addton, the models have done a poor ob explanng the varaton n trade changes across countres and ndustres. Whle the exstng computable models of trade use the Armngton (1969) methodology to explan ntra-ndustry trade, ths paper presents a new model, called the HPPC model, that uses the Eaton and Kortum (2002) methodology for that purpose. Usng ths framework on the ndustry level results n a hghly tractable model that has all the usual neoclasscal features wth

15 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, room for both Rcardan and Heckscher-Ohln reasons for trade and asymmetrcal ndustry-level trade costs. After descrbng the model, the paper evaluates t by usng t to forecast the effects of NAFTA from the pont of vew of The results are then compared wth the post-nafta data and forecasts of other models of trade. The results show that the HPPC model performs better than the exstng models of trade. Its forecast s closer n magntude to the actual data and ts ablty to predct varaton of trade changes across countres and ndustres s better. The paper nvestgates why the HPPC model s able to produce better forecasts than the exstng CGE models, focusng specfcally on the Brown et al. (1992a) model as an example of an exstng CGE model. It fnds that the BDS model suffers from usng Armngton elastctes that are too low. It also fnds that the BDS's treatment of the non-tarff barrers has contrbuted to the poor qualty of ts forecasts. Endnotes * Emal: serge.shkher@suffolk.edu; The author would lke to thank Tolga Ergün, Jonathan Haughton, Tmothy Kehoe, and the anonymous referees for ther helpful comments. 1. Predctons made by poltcans vared from sgnfcant beneft to the U.S. to sgnfcant loss of U.S. obs to Mexco. For example, Presdent Clnton, who sgned NAFTA n 1993, promoted ts benefts to the U.S. and the world. Ross Perot, an ndependent presdental canddate n 1992, predcted sgnfcant ob losses n the U.S. 2. Factor endowments are fxed n ths model. An unpublshed appendx (avalable from the author upon request) presents an extenson that allows domestc accumulaton (followng the Solow model) and nternatonal moblty of captal (that equalzes the rates of return across countres). The assumpton of fxed captal stock n ths paper s motvated by several reasons. Frst, the exstng models of NAFTA that allow captal moblty use varous and often ad-hoc closure rules, makng a formal comparson wth ther forecasts dffcult. Second, allowng captal accumulaton and nternatonal moblty has only a small effect on the forecasted changes n trade (see footnote 34). 3. Calendo and Parro (2010) also have a model that uses the Eaton-Kortum methodology on the ndustry level. Compared to the HPPC model, ther model has only one factor of producton and s parametrzed dfferently. Note that ths paper and the HPPC model predate ther paper and model. 4. Kortum (1997) and Eaton and Kortum (1999) provde mcrofoundatons for ths approach. Parameter T governs the mean of the dstrbuton, whle parameter, whch s common to all countres and ndustres, governs the varance. The support of the Fréchet dstrbuton s ( 0, ).

16 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, To receve $1 of product n country n requres sendng d 1 dollars of product from country. By defnton, domestc transport costs are set to one: d 1. Trade barrers result n d The last equalty follows from a known statstcal result. These dervatons are explaned n greater detal n Eaton and Kortum (2002). 7. The technology parameter T s dfferent from the total factor productvty (TFP). Parameter T determnes the mean of the Fréchet dstrbuton and s exogenous n ths model. TFP, on the other hand, s endogenous. Fncell, Pagano, and Sbraca (2007) derve the analytc relatonshp between the T of an ndustry and the mean productvty of the frms that actually operate n that ndustry. 8. It obtans as follows: Z Znm m p m M nm m m nmm nm, where Z nm s the amount spent by ndustry m on ntermedate goods from ndustry, M s the quantty of ntermedate goods, and the last equalty follows from (2). Then from (1) M w L 1 /. nm nm n nm m m gm 9. Consumpton C ncludes prvate consumpton and government consumpton The model has N J 5NJ 3N unknowns and the same number of equatons. The unknowns n the model are X, c, p, K, L, Q, Y n, w n, and r n. 11. The countres and ndustres ncluded n the model are chosen because of the avalablty of data, especally wages and ndustry-level output and spendng. The year 1989 s chosen because the CGE models wth whch the HPPC model s compared n Secton Sect: Evaluaton are parametrzed wth data from 1989 or The other estmate of n Eaton and Kortum (2002), 3.6, results n abnormally large estmates of trade barrers d (6.6 average across all countres and ndustres). The results presented n ths paper are affected by the value of. See Secton 4 for analyss. 13. Followng Eaton and Kortum (2002), the physcal dstance s dvded nto 6 ntervals: [ 0,375), [ 375, 750), [ 750,1500), [ 1500,3000), [ 3000, 6000), and [ 6000, maxmum ) to create phys d k. 14. Note that the estmatng equaton ncludes the export and mport dummy varables, smlarly to the theoretcally-derved gravty equaton of Anderson and van Wncoop (2003). n

17 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, For some pars of countres, trade values are mssng for Therefore,, whch are part of the dstance measure, could not be estmated for some n,, and. There are 19*18*8 =2,736 observatons of possble n the data, of whch 105 or 3.8% are mssng. Most mssng observatons are proxed by estmates from neghborng years. Sx observatons that could not be proxed n ths manner were proxed by the estmates of n for total manufacturng. 16. Snce Mexco has been lberalzng ts trade even before NAFTA, usng 1991 nstead of 1989 trade barrers n the smulatons may have reduced the forecasted growth n Mexcan mports. 17. Note that the trade barrers between the U.S. and Canada were not zero n Even though the U.S.-Canada FTA went nto effect n 1988, t called for a gradual removal of all blateral trade barrers. Therefore, many tarffs and NTBs were stll n place n Secton 4 shows the trade forecasts f the NTB barrers are gnored altogether. 19. Ths procedure s dfferent from the approach used by Eaton and Kortum (2002) to fnd the technology parameters. They calculate technology parameters from the estmated mporter and exporter dummes and data on wages. 20. The rates of return r are gross rates. The rate of 20% s obtaned by assumng 10% net return and 10% deprecaton. The results presented n the paper are not senstve to these values. 21. The correlaton between the calculated ndustry-level captal stocks and the captal stocks n the data s The same number for labor s The correlaton between the predcted and actual trade flows n the base year s near 1, but t needs to be remembered that the model has very few degrees of freedom. The average Grubel-Lloyd ndex (t measures the sze of ntra-ndustry trade) s n the model and n the data. 23. The focus wll prmarly be on the predctons regardng trade, and not GDP, prces, or welfare. The reason s that t s dffcult to fnd data on prces, and GDP and prces were both sgnfcantly affected by events other than NAFTA. 24. The HPPC and BDS models were parametrzed wth data from The RRS model was parametrzed wth data from 1988, whch should not make notceable dfference for the analyss of ths paper because trade data s very slow-movng. 25. Data presented n López-Córdova (2002) shows that the percentage of U.S. manufacturng mports to Mexco that were ether not subect to tarff or pad at most 5% tarff was 10 n

18 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, and 93 n The percentage that pad more than 10% tarff was 85 n 1993 and less than 1 n They nclude Sobarzo (1992), Roland-Holst, Renert, and Shells (1994, Cox and Harrs (1992), Brown et al. (1992a), Bachrach and Mzrah (1992), Shells and Shelburne (1992), Hunter, Markusen, and Rutherford (1992) (auto ndustry), and McCleery (1992). The pre- NAFTA studes were ntally collected together by the U.S. Internatonal Trade Commsson (1992). The updated versons of some of these studes, together wth the several new ones were later collected n Francos and Shells (1994) and Kehoe and Kehoe (1995b). Summares of these studes are presented n Brown, Deardorff, and Stern (1992b), Baldwn and Venables (1995), and Kehoe and Kehoe (1995a). 27. See Baldwn and Venables (1995) for a revew and classfcaton of these models. 28. Unfortunately, some studes dd not report the sze of the trade barrers that were removed durng ther smulatons of NAFTA. 29. The examples nclude Gould (1998) and Krueger (1999). Unfortunately, many of these studes do not use the theoretcally-derved specfcaton of the gravty equaton of Anderson and van Wncoop (2003). Romals (2007) uses a CGE model wth the Armngton assumpton, parametrzed wth the post-nafta data, to study the effects of NAFTA. Revews of the post-nafta lterature can be found n Burfsher et al. (2001) and Romals (2007). 30. Fox (2000) evaluates the performance of CGE models n predctng the effects of U.S.- Canada free trade agreement. 31. Kehoe revews the forecasts of the Brown-Deardorff-Stern, Cox-Harrs and Sobarzo models. 32. The sources for the BDS results are Brown et al. (1992a) and Kehoe (2005); the source for the RRS results s Roland-Holst et al. (1994). 33. Total NAFTA trade s the sum of all blateral trade flows n NAFTA: n H X X,, n where H s the set of NAFTA countres. Total trade of the NAFTA countres s n H EX IM, where EM and IM are total exports and mports of ndustry goods n country n. Measurng blateral trade relatve to total trade or ncome helps to control for events other than NAFTA that affected the economes of the NAFTA countres durng the post-nafta perod. See Secton 4 for the dscusson of such events. 34. As mentoned n footnote 2, an unpublshed extenson of the HPPC model ncorporates domestc accumulaton and nternatonal moblty of captal. In the smulaton of NAFTA, ths extenson predcts an accumulaton of captal stock n the NAFTA countres, mostly due to a transfer from the non-nafta countres. It predcts that the captal stocks of Canada, Mexco, and the U.S. would grow 9.1, 10, and 0.65%, respectvely. The total NAFTA trade

19 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, relatve to the total trade of the NAFTA countres n ths case s predcted to grow 28.4% and the total NAFTA trade relatve to the total ncome of the NAFTA countres s predcted to grow 66.2%. Therefore, allowng domestc accumulaton and nternatonal moblty of captal has farly small effects on these measures of NAFTA trade. 35. The BDS model uses a 21-ndustry classfcaton. Ther results were aggregated to the 8- ndustry classfcaton used by the HPPC model. The authors of the RRS model do not provde a concordance between ther ndustres classfcaton and SIC. Judgng by ndustry names, some ndustres n the RRS model have no equvalents n SIC. 36. Trade between U.S. and Canada was 76% of the NAFTA trade n 1989 and 59% n Trade between U.S. and Mexco was 23% of the NAFTA trade n 1989 and 39% n Alternatvely, the data s correct and the models have trouble makng predctons n the vcnty of autarky. 38. The 1989 Mexcan Wood exports to Canada are reported at only $8.2M. 39. Correlaton s 2 R. 40. As n the prevous secton, the emphass here wll be on the BDS model. As explaned earler, I chose the BDS model because t seems to have made good NAFTA forecasts and because ts results (.e. varables, ndustral structure) are readly comparable to the results of the HPPC model. 41. The RRS model sets at around 1. Note that settng equal to 3 or 1 n the HPPC model would result n very unreasonably hgh estmates of d. 42. Note that the estmates of total trade costs d change when the value of changes. 43. More precsely, there s evdence of small overpredcton by the model. 44. The percent change n trade costs durng trade lberalzaton s d 1 t f.o.b. tarffs and / 1 / n case of the d n case of the c..f. tarffs. For a back-of-the-envelope analyss of ths dfference, let's take 8, t , and , whch are average for the HPPC model's NAFTA smulaton. Wth these numbers, the percent change n trade costs s about 50% greater n case of the c..f. tarff than n case of the f.o.b. tarff. 45. The BDS model ncorporates NTBs by endogenously solvng for the ad valorem tarff rate that wll hold mports wthn each product category covered by NTBs at a predetermned level. (Brown et al., 1992a).

20 Shkher, Journal of Internatonal and Global Economc Studes, 5(2), December 2012, The overall magntude of trade changes decreases to be agan roughly smlar to that of the BDS model. 47. The HPPC and BDS models also have dfferent assumptons regardng the returns to scale. The HPPC model assumes constant returns to scale, whle the BDS model assumes ncreasng returns. However, the forecasts of the constant and ncreasng returns versons of the RRS model have smlar correlatons wth the data (across countres, see Table 4 - RRS do not present comparable ndustry-level results for both versons of the model, so I cannot check the smlarty of cross-ndustry correlatons.), whch suggests that the degree of the returns to scale does not play a bg role n determnng the varaton n trade changes. References Anderson, J. E. and van Wncoop, E Borders, trade, and welfare, n D. Rodrk and S. Collns (eds), Brookngs Trade Forum 2001, Brookngs Insttuton. Anderson, J. E. and van Wncoop, E Gravty wth gravtas: A soluton to the border puzzle, Amercan Economc Revew 93(1). Armngton, P. S A theory of demand for products dstngushed by place of producton, IMF Staff Papers 16(1): Bachrach, C. and Mzrah, L The economc mpact of a free trade agreement between the Unted States and Mexco: A CGE analyss, Economy-Wde Modelng of the Economc Implcatons of a FTA wth Mexco and a NAFTA wth Canada and Mexco, USITC, Washngton, DC. Baldwn, R. E. and Venables, A. J Regonal economc ntegraton, n G. M. Grossman and K. Rogoff (eds), Handbook of Internatonal Economcs, Volume III, Elsever, chapter 31. Brown, D. K., Deardorff, A. V. and Stern, R. M. 1992a. A North Amercan free trade agreement: Analytcal ssues and a computatonal assessment, The World Economy 15(1): Brown, D. K., Deardorff, A. V. and Stern, R. M. 1992b. North Amercan ntegraton, The Economc Journal 102: Burfsher, M. E., Robnson, S. and Therfelder, K The mpact of NAFTA on the Unted States, The Journal of Economc Perspectves 15(1): Calendo, L. and Parro, F Estmates of the trade and welfare effects of NAFTA, Unversty of Chcago mdeo. Cox, D. and Harrs, R. G North Amercan free trade and ts mplcatons for Canada: Results from a CGE model of North Amercan free trade, Economy-Wde Modelng of the