Anderson and Wincoop (2004) Trade Costs

7/30/2019 Anderson and Wincoop (2004) Trade Costs

1/62

Trade Costs

James E. Anderson; Eric van Wincoop

Journal of Economic Literature, Vol. 42, No. 3. (Sep., 2004), pp. 691-751.

Stable URL:

http://links.jstor.org/sici?sici=0022-0515%28200409%2942%3A3%3C691%3ATC%3E2.0.CO%3B2-%23

Journal of Economic Literature is currently published by American Economic Association.

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available athttp://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtainedprior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content inthe JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained athttp://www.jstor.org/journals/aea.html.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printedpage of such transmission.

The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academicjournals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers,and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community takeadvantage of advances in technology. For more information regarding JSTOR, please contact [email protected].

http://www.jstor.orgFri Nov 2 16:31:31 2007
http://links.jstor.org/sici?sici=0022-0515%28200409%2942%3A3%3C691%3ATC%3E2.0.CO%3B2-%23http://www.jstor.org/about/terms.htmlhttp://www.jstor.org/journals/aea.htmlhttp://www.jstor.org/journals/aea.htmlhttp://www.jstor.org/about/terms.htmlhttp://links.jstor.org/sici?sici=0022-0515%28200409%2942%3A3%3C691%3ATC%3E2.0.CO%3B2-%23


2/62

Journal of Economic LiteratureVol. XLII (Septem ber2004)pp. 691-751

Trade Costs JAMES E. ANDERSONnd Eric van wincoop1

". . . the report of m y d eath wa s a n exaggeration." - Mark Twain, 18971. Introduction

The death of distance is exaggerated.Trade costs are large, even aside fromtrade-policy barriers and even betweenapparently highly integrated economies.Despite many difficulties in measuring andinferring the height of trade costs and theirdecomposition into economically usefulcomponents, the outlines of a coherent pic-ture emerge from recent developments indata collection and especially in structuralmodeling of costs. Trade costs have econom-ically sensible magnitudes and patternsacross countries and regions and acrossgoods, suggesting useful hypotheses fordeeper understanding. This survey is aprogress report. Much useful work remainsto be done, so we make suggestions below.Trade Costs Mattel: (1)They are large-a170-percent total trade barrier is construct-ed below as a representative rich countryad-valorem tax equivalent estimate. Thisincludes all transport, border-related, andlocal distribution costs from foreign poduc-er to final user in the domestic county. (2 )

' .klderson: Boston College and NB ER. van Wincoop:University of Virginia and NBE R. W e are grateful to JohnMcMillan, two referees, and Jeff Bergstrand for exceptionally generous and detailed comments and sugges-t ions. We have a lso benef i t ted f rom many othercomments by colleagues too numerous to single out,tllrough e-mail and in presentation s at conferences an d atuniversity workshops.

Trade costs are richly linked to economicpolicy. Direct policy instruments (tariffs, thetariff equivalents of quotas and trade barri-ers associated with the exchange-rate sys-tem) are less important than other policies(transport infrastructure investment, lawenforcement and related property-rightsinstitutions, informational institutions, regu-lation, language). (3 ) Trade costs have largewelfare implications. For example,Anderson and van Wincoop (2002) arguethat current policy-related costs are oftenworth more than 10 percent of nationalincome. (4) Maurice Obstfeld and KennethRogoff (2000) argue that all the major puz-zles of international macroeconomics hangon trade costs. ( 5 )Details of trade costs mat-ter to economic geography. For example, thehome market effect hypothesis (big coun-tries produce more of goods with scaleeconomies) hangs on differentiated goodswith scale economies having greater tradecosts than homogeneous goods (DonaldDavis 1998).(6)The cross-commodity struc-ture of policy barriers is important to welfarei e . ~ . .Anderson 1994).i r o a d l u ~ e f i n e d . ' rade costs. broadlvJ Jdefined, include all costs inconed in gettinga good to a other than Ihe margin-al cost of producing the good itself: trans-portation costs (both freight costs and tirne

policy barriers itariffs and llontariffbarriers), information costs, contract


3/62

Jozrrnal of Economic Litelenforcement costs, costs associated with theuse of different currencies, legal and regula-tory costs, and local distribution costs(wholesale and retail). Trade costs arereported in terms of their ad-valorem taxequivalent. The 170-percent headline num-ber breaks down into 55-percent local distri-bution costs and 74-percent internationaltrade costs (1.7=1.55*1.74-1).Both domestic and international tradecosts are included because it is arbitrary tostop counting trade costs once goods cross aborder. It is not even obvious when goodscross a border in the economic sense: Is itwhen they arrive on the dock, leave thedock, arrive 'it the importer?

Costs broadlv defined and reported belowmay include sonle rent. Exclusion of rentrequires a markup theory and its application.Tlzree Sorlrces. We report on trade costsfrom three broad sources. Direct measuresof trade costs are discussed in section 2.Direct measures are remarkably sparse andinaccurate. Two types of indirect measurescomplement the incomplete and inade-quate direct measures: inference fromquantities (trade volumes), discussed in sec-tion 3, and inference from prices, discussedin section 4.Theo ry Looriw Larg e. In our survey, theo-n7looms large. A theoretical approach isinevitable to infer the large portion of tradecosts that cannot be directly measured in thedata. The literature on inference about tradebarriers from final goods prices remainslargely devoid of theory. \lie point to ways inwhich trade theory can be effectively used tofill this gap and learn more about trade bar-riers from evidence on prices. Recent devel-opments have bridged the gap betweenpractice and theory in the inference of tradecosts from trade flows. Readers who pausewith us on the bridge will produce betterwork in the future.The gravity model provides the main linkbetween trade barriers and trade flows.Gravity is often taken to be rather atheoret-ic or justified only under highly restrictive

ratzrre Vol. XL II (S ep tem be r 2004)assumptions. We place gravity in a wide classof trade separable general equilibrium mod-els. Trade separability obtains when the allo-cation of trade across countries is separablefrom the allocation of production andspending within countries. Gravity links thecross-country general equilibrium tradeallocation to the cross-country trade barri-ers, all conditional on the observed con-sumption and production allocations.Inferences about trade costs therefore donot depend on the general equilibriumstructure that lies beneath the observedconsumption and production allocationswithin countries.Appropriate aggregation of trade costs is akey concern. Aggregation of some sort isinevitable due both to the coarseness ofobservations of complex underlying phe-nomena and the desirability of simple meas-ures of very high dimensional information.\lie show how theory can be used to replacecommon atheoretic aggregation methodswith ideal aggregation. \lie also show howtheory can shed light on aggregation biasand what can be done to resolve it.

Trade Costs Are Large and Variable.Mattel's Barbie doll, discussed in RobertFeenstra (1998), illustrates large costs. Theproduction costs for the doll are $1, while itsells for about $10 in the United States. Thecost of transportation, marketing, wholesal-ing, and retailing have an ad-valorem taxequivalent of 900 percent.A rough estimate of the tax equivalent of"representative" trade costs for industrializedcountries is 170 percent. This number breaksdown as follows: 21 percent transportationcosts, 44 percent border-related trade barri-ers, and 55 percent retail and wholesale dis-tribution costs (2.7=1.21* .44*1.55). The21-percent transport cost includes bothdirectly measured freight costs and a 9-per-cent tax equivalent of the time value of goodsin transit. Both are based on estimates forU.S. data. The 44-percent border-relatedbarrier is a combination of direct observationand inferred costs. Total international trade


4/62

693nderson and va n Winco op: Trade Costscosts are then about 74 percent(0.74=1.21*1.44-1). Representative retailand wholesale distribution costs are set at 55percent, close to the average for industrial-ized countries.Direct evidence on border costs showsthat tariff barriers are now low in mostcountries, on average (trade-weighted orarithmetic) less than 5 percent for richcountries, and with a few exceptions are onaverage between 10 percent and 20 percentfor developing countries. Our overall repre-sentative estimate of policy barriers forindustrialized countries (including nontariffbarriers) is about 8 percent. Inferred bor-der costs appear on average to dwarf theeffect of tariff and nontariff policy barriers.An extremely rough breakdown of the44-percent number reported above is asfollows: an 8-percent policy barrier, a 7 -percent language barrier, a 14-percentcurrency barrier (from the use of differentcurrencies), a 6-percent information costbarrier, and a 3-percent security barrier forrich countries.Trade costs are also highly variable acrossboth goods and countries. &ade barriers indeveloping countries are higher than thosereported above for industrialized countries.High value-to-weight goods are less penal-ized by transport costs. The value of timeli-ness varies across goods, explaining modalchoice. Poor institutions and poor infrastruc-ture penalize trade differentially acrosscountries. Sectoral trade barriers appear tovary inversely with elasticities of demand.Policy barriers, especially nontariff barriers(NTBs), also vary significantly across goods.NTBs are highly concentrated: in many sec-tors they are close to zero, but U.S. textilesand apparel have 71 percent of productscovered, with tariff equivalents ranging from5 percent to 33 percent.

2. Direct EvidenceDirect evidence on trade costs comes in

two major categories, costs imposed by

policy (tariffs, quotas and the like) andcosts imposed by the environment (trans-portation, insurance against various haz-ards, time costs). We review evidence oninternational policy barriers, transportcosts and wholesale and retail distributioncosts. We focus on current and recent tradecosts in this survey; see the work ofWilliamson and co-authors (e.g., KevinO'Rourke and Jeffrey Williamson 1999) forhistorical evidence.An important theme is the many difficul-ties faced in obtaining accurate measures oftrade costs. Particularly egregious is thepaucity of good data on policy barriers."Transport-cost data is limited in part by itsprivate nature. Many other trade costs, suchas those associated with information barriersand contract enforcement, cannot be direct-ly measured at all. Better data on trade costsare feasible with institutional resources andwould yield a high payoff.2.1 Policy Barriers2.1.1 Measurement Problerns andLimitations

How high are policy barriers to trade?This seemingly simple question cannot usu-ally be answered with accuracy for mostgoods in most countries at most dates. Theinaccuracy arises from three sources:absence of data, data that are useful only incombination with other missing or fragmen-tary data, and aggregation bias. Each of thedifficulties is discussed further below.The key open3 (to the research communi-ty) source for panel data on policy barriers

"h e grossly inco ~np lete nd inaccurate informatioll onpolicy barriers available to researchers is a scandal and apuzzle. It is natural to assume that trade policy is well-docum ented since th eo q. and politics have emphasized itfor hundreds o f years.U NC TA D sells the TRA INS data each year to com-~nercial usto~ nerswho use it to provide current informa-tion on trade costs to potential traders. It comes with afront end designed for convenience in pulling a maxi~nu~no f 200 lines o f data while preventing a user from gainingaccess to the whole database and using it to compete withUN CTA D's potential sales to other customers.


5/62

694 Journal of Economic Literature Vol. XLII (September2004)to trade is the United Nations Conferenceon Trade and Development's Trade Analvsis& Information System, TRAINS. It containsinformation on trade control measures (tar-iff, para-tarifp and nontariff measures) attariff line level for a maximum of 137 coun-tries beginning in the late 1980s. TRAINSreports all data on bilateral tariffs, nontariffbarriers, and bilateral trade flows at the six-digit level of the Harmonized System (HS)product classification for roughly 5000"products." Countries use a finer productclassification when reporting tariffs toUNCTAD. Thus, multiple tariff "lines"underlie each six-digit aggregate. It is insome cases possible to drill down to thenational tariff line level. Some organizationshave obtained the full TRAINS database(without the front end software) and pro-vide access to a limited set of users insidethe ~ r ~ a n i z a t i o n . ~

Table 1 gives a sense of the substantialincompleteness of TRAINS. For each yearfrom 1989 to 2000, the table reports thefraction of countries that report some lines(though possibly a very limited number) fortariffs, NTBs, and trade flows. Of 121reporting countries in 1999, 43 percentreport tariffs, 30 percent report NTBs, 55percent report trade flows, and 17.4 per-cent have data for all three. The other coun-tries report no data at all for any good.Coverage is not much better in other \.ears.Coverage is better for OECD count&-over 50 percent have tariff and NTB infor-mation recorded in 1999, with considerablevariability in coverage across the years.

The World Bank has very recently puttogether from these elements the most

'' Para-tariff measures include customs surcharges suchas i ~ p o r ticense fees , foreign exchange taxes, stamp s, etc." Fo r example, Boston College has purchased disaggrr-gated tariff information from UNCTAD's TRAINS data-base for the years 1988 through 2001, inclusi\~e.Ve havrdata for 137 countries for at least one year, counting theEuropean Union as a single countq; but far less than themaximum am ount suggested by four tren years t im rs 137countries.

comprehensive svstem for researchers, andin principle prdvides public Itsc ~ e s s . ~World Integrated Trade System (WITS)sofhvare is coupled to TRAINS and to theWorld Trade Organization (WTO)Integrated Data Base and ConsolidatedTariff ~che dule s' along with the UNStatistical Division's COMTRADE tradeflow data. In principle it allows users to drilldown and select data according to their owncriteria, to track the complexities of tradepolicy as finely as the primary inputs allow.8WITS has some other data handling andmodeling functions as well.

National sources in combination with theabove allow better measurement of a singlecountry's trade policy A series from theInstitute for International Economics pro-vides measures of a few national trade policies

' s e e http://wits.worldbank.org . At this writing, thereare still technical glitches facing a user tning to gainaccess. \I7ITS only runs in late mod el \Untiows machines,users may need some IT support to install the software;and a user must to pay fees to UKCT AD for use of CO M -TR AD E and TRA INS. Email queries are not answered, inou r experience, without using friends at the Bank as inter-mediaries.' The \ I T 0 Consolidated Tariff Schedule database l iststhe Most Favored Nation (M FN ) boullti tariffs at the tar-iff line level. The bo und tariffs are the upp er li~ nit s nderthe m em ber countries' \ I T 0 obligatioll for actual tariffscharged to countries not then e ~ n b e r a ss oc ia te d w ithimporter in a free trade agreement or a customs union,anti often exceed the actual duties charged. The \\TO'SIntegrated Data Base colltaills information on the appliedrates at the national tariff line level. Th e tiata is closed toresrarchers. The LIT0 periodically reports on ind ilidualmember-counti?. trad e policy with a published t rade poli-cy review based o n its da ta.'Th e \170rld Bank Tra de allti Prod uction database pro-duces from its resources a set of three- and four-digitaggregates of trade, production, and tariff data. It is pub-lished on the Bank website and presumably will be regu-larly update d. The Trade and Prod uction database covers67 developing and developrd countries over the period

1976-99. Again, the tiescriptioll given misleadingly sug-grs ts a usrful pa nel; the actual data is full of missing ob ser-vations d ue to the untierl!ing limitations of TR AIN S. Th esector disaggregation in the database follows theIlltemational Stand ard Industrial Classification (IS IC )andis prolided at the three-digit level (28 industries) for 67coulltries and at the four-digit level (81 industries) for 24of thesr countr ies.
http://wits.worldbank.org/http://wits.worldbank.org/


6/62

Anderson and van Win coop : Trade CostsTABLE 1PERCENTAGEF COUKTRIESITH DATAN TRAINS

All Countries OE CD CountriesYear Tariff NTB Trade All Tariff NTB Trade All1989 0.8 1.7 1.7 0.8 0.0 0.0 5.6 0.0

2000 36.4 2.5 0.0 0.0 55.6 55.6 0.0Notes: The data are from UNCTAD's TRAINS database (Haveman Repackaging). The table reports the percentageof all countries, based on total of 121 reporting countries, that have at least one type of data for one year availablethrough TRAINS. For OE CD countries the percentages are based on 19countries. "All" indicates that a countryhas reported all three types of data for that year.

for single years.9 See also the WTO's Trade and must be converted to ad valorem equiv-Policy Reviews series. alents using price information that must beData limitations can make even the avail- matched up to the tariffs. Imperfections ofable trade-barrier information difficult to classification or other information introduceinterpret. TRAINS does not routinely report potential measurement error. TRAINSad-valorem equivalents of specific tariffs (on reports the percentage of underlying linesquantities rather than values). Where specif- that have specific tariffs in order to provideic tariffs are prevalent, such as in agriculture some information about how widespreadin many countries, the omission of specific the measurement problem may be. Thetariffs is highly misleading. Many tariffs are World Bank's trade-barrier database sub-in the form of specific taxes (on quantity), stantially supplements TRAINS by supply-ing missing specific tariff information. The

See Gaq- Hufbauer and Kimberly Ann Elliott (1994) match of the tariff line classifications withon the United States; Patrick Messerlin (2001) on the the commodity classifications for trade flowsEuropean Union; Yoko Sazana~ni,Shujiro Urate, and is imperfect as well, introducing measure-Hiroki Kawai (1995) on Japan; Namdoo Kiln (1996) onSouth Korea; and Shuguang Zhang, Zhang Yansheng, and ment error when converting all data to theWan Zhonpn (1998)on China. Harmonized System.

0.0


7/62

696 J o u ~ m a lof Econoinic Literature Vol. X L I I (Septeinber 2004)Nontariff barriers (NTBs) are much moreproblematic than tariff barriers. The WorldBank database unfortunately does not pro-vide NTB data. A user must use TRAINS ormore specialized databases directed at par-ticular NTBs. The TRAINS databaserecords the presence or absence of a nontar-iff barrier (NTB) on each six-digit line. Manvdiffering types of nontariff barriers arkrecorded in TRAINS (a total of eighteentypes). The NTB data requires concordancebetween the differing NTB, tariffs, andtrade classification svstems at the nationallevel, converting to the common HS system.Jon Haveman's extensive work withTRAINS has produced a usable NTB data-

base.'' Haveman follows what has become acustomary grouping of NTBs into hard bar-riers (price and quantity measures), threatmeasures (antidumping and countervailingduty investigations and measures), and qual-ity measures (standards, licensing require-ments, etc .). A fourth category is embargoesand prohibitions. A common use of the NTBdata is to construct a measure of the preva-lence of nontariff barriers, such as the per-centage of HS lines in a given aggregatethat are covered by NTBs. Nontariff barrierinformation in TRAINS is particularlyprone to incompleteness and poor-qualityproblems, so analysts seeking to study parti-cular sectors such as the Multi-FibreArrangement (MFA) will do better to accessspecialized databases such as the WorldBank's MFA data."No information about NTB restrictive-ness is provided in TRAINS, since measur-ing the restrictiveness of each type ofnontariff barrier requires an economicmodel. In some important cases, individualanalysts have developed direct measures of

lo See the Ultimate Trade Barrier Catalog atht //\m~.~v.eiit.org/Protection.

Uationd tmff line infornlation is also very problem-atic when analyzing nontariff barriers. For example,matching up reported trade flows with annual quotasimmedia teb runs into inconsistencies in reporting conven-tions.

the restrictiveness of NTBs based on quotalicense prices where these are available.Indirect methods of measuring the restric-tiveness of NTBs are important because ofthe paucity of direct measures. One methodis to infer the restrictiveness of nontariff bar-riers through the comparison of prices. Someimportant trade lines are well-suited for pricecomparisons (homogeneous products sold onwell-organized exchanges, for example), buteven here there are important issues withdomestic transport and intermediary marginsand the location of wholesale markets rela-tive to import points of entry. Evidence fromprice comparisons is discussed in section 4.The restrictiveness of nontariff barriers canalso be inferred from trade quantities in thecontext of a well-specified model of tradeflows. Inference about nontariff barriersfrom trade flows is discussed in section 3.Alan Deardorff and Robert Stern (1998) andSam Laird and Alexander Yeats (1990) pro-vide other detailed discussions of inferenceabout the restrictiveness of NTBs.Aggregation is an important problem inthe use and analysis of trade barriers. Tariffsand NTBs comprise some 10,000 lines, withlarge variation across the lines. The nationalcustoms authorities are the primary sourcesof trade restrictions, and their classificationsystems do not match up internationally oreven intranationally, as between trade flowson the one hand and tariff and nontariff-barrier classes on the other hand. Matchingup the tariff, nontariff, and trade-flow datarequires aggregation, guided by concor-dances that are imperfect and necessarilygenerate measurement error. Moreover, formany purposes of analysis, the comprehen-sion of the analyst is overwhelmed by detail,and further aggregation is desirable.Atheoretic indices such as arithmetic (equal-ly weighted) and trade-weighted average tar-iffs are commonly used, whileproduction-weighted averages sometimesreplace them. for nontariff barriers, the

indicator is aggregated into a nontar-iff barrier coverage ratio, the arithmetic or


8/62

697nderson and va n W incoop: Trade Coststrade-weighted percentage of componentsectors with nontariff barriers.Ideal aggregation is proposed by JamesAnderson and Peter Neary (1996, 2003)based on the idea of a uniform tariff equiva-lent of differentiated tariffs and NTBs.Theoretically consistent aggregationdepends on the purpose of the analysis, sothe analyst must specify tariff equivalencewith reference to an objective that makessense for the task at hand. Anderson andNeary develop and apply indices for thesmall-country case that are equivalent interms of welfare and in terms of distortedaggregate trade volume,12 and show thatatheoretic aggregation can significantly biasthe measurement of trade restrictiveness.13The theme of appropriate aggregation in thedifferent setting of many countries in gener-al equilibrium plays a prominent role in ourdiscussion of indirect measurement of tradecosts, so we defer a full treatment to thatsection. Ideal aggregation is informationallydemanding, so for that reason and becauseof their familiarity and availability in thework of others, we report the standard trade-weighted and arithmetic averages of tariffsand of NTB coverage ratios below.2.1.2 Evidence on Policy Barriers

Tariffs. Trade-weighted and arithmeticaverage tariffs are reported in table 2 for fifty'"ee also Anderson (1998)for equivalence in terms of

sector-specific factor income and its relationship to thecom~nonlyeported effective rate of protection.

l3 Anderson and Neaq. (2003)report results using avo lu~ ne quivalent uniform tariff- replacement of the dif-ferentiated tariff structure with a uniform tariff such thatthe general equilibrium aggregate value of trade in tiistort-eti products (in terms of external prices) is held constant.The ideal index is usually larger than the trade-weightedaverage-an arithmet ic average across countries in thestudy !ields approximately 11percent for the trade weight-ed average anti 12 percent for the uniform volume equiv-alent, while the U.S. nu~nbersn 1990 are 4 percent and4.8 percent respectively For purposes of co~nparisonbetween the initial tariffs and free trade, the two indexesare quite highly correlated. For a s ~na lle ret of evaluationsof year-on-year changes, in contrast, Anderson anti Neal?show that the ideal and trade-weighted average indexes areuncorrelatect

countries for 1999, based on TRAINSdata.14 The relatively small number of coun-tries reflects reporting difficulties typical ofTRAINS; some earlier years contain data formore countries. The reported numbersaggregate the thousands of individual tarifflines in the underlying data. The table con-firms that tariffs are low among most devel-oped countries (under 5 percent), whiledeveloping countries continue to have high-er tariff barriers (mostly over 10 percent).Dispersion across countries is wide: HongKong and Switzerland have 0 percent tariffs,the United States has a 1.9 percent simpleaverage, and at the high end India has 30.1percent and Bangladesh 22.7 percent.The variation of tariffs across goods isquite large in all countries; typically only afew are large. Intuition suggests that thevariation of tariffs adds to the welfare cost.Marginal deadweight loss is proportional tothe tariff, hence the cumulated dead weightloss triangle varies with the square of the tar-iff. Coefficients of variation (the ratio of thestandard deviation to the mean) of tariffs,either arithmetic or trade-weighted, thussometimes supplement averages. Andersonand Neary (2003) report trade-weightedcoefficients of variation of tariffs for 25 coun-tries around the year 1990, ranging from0.14 to 1.67, many being clustered aroundone. Thev show that a proper analysis quali-fies the simple intuition considerably, withwelfare cost increasing in an appropriatelyweighted coefficient of variation.Bilateral variation of tariffs can also belarge. Preferential trade is mainly responsible

since insiders face a zero tariff while outsidersface the MFN tariff, but aggregation overgoods induces further bilateral variation dueto differing composition of trade across part-ners. James Harrigan (1993) reports bilateralproduction-weighted average tariffs in 28product categories for OECD countries for'"dculations were based on TRAINS annual data-

bases purchased by Boston College without the front-e d oftware and assembled into panel data


9/62

TABLE 2SILIPLE TARIFFKD TRADE-\VEIGHTED A V E R A G E S - ~ ~ ~ ~

County Simple T \VAverage Average

Argentina 14.8 11.3Australia 4.5 4.1Bahamas 0.7 0.8Bahrain 7.8 -Bangladesh 22.7 21.8Barbados 19.2 20.3Belize 19.7 14.9Bhutan 15.3 -Boli~ia 9.7 9.1Brazil 15.5 12.3Canada 4.5 1.3(:bile 10.0 10.0(;oloinbia 12.2 10.7(:osta Rica 6.5 4.0(;zech Republic 5.5 -Ilorninica 18.5 15.8Ecuador 13.8 11.1European Union 3.4 2.7(korgia 10.6 -Grenada 18.9 15.7Guyana 20.7 -Ilonduras 7.5 7.8Hong Kong 0.0 0.0India 30.1 -Illdonesia 11.2 -Jiiunaica 18.8 16.7Japan 2.4 2.9Korea 9.1 5.9Mexico 17.5 6.6hlontserrat 18.0 -New Zealand 2.4 3.0Nicaragua 10.5 11.0Paraguay 13.0 6.1Peru 13.4 12.6Pllilippilles 9.7 -Romania 15.9 8.3Saudi Arabia 12.2 -Singapore 0.0 0.0Slovenia 9.8 11.4Soutll Africa 6.0 4.4St. Kitts 18.7 -St. Lucia 18.7 -St. Uncent 18.3 -Suriname 18.7 -Switzerland 0.0 0.0Taiwan 10.1 6.7Trinidad 19.1 17.0Ur11gua)i 4.9 4.5USA 2.9 1.9Ve~~ezuela 12.4 13.0

Xotcs: Tlle data are from UNCTAD's TRAINS database (Haveman repackagmg).A "-"indicates that trade data for 1999 are ullavailable in TRAINS.


10/62

6 9 9nderson and van Wincoo p: Trade Costs1983. For Canada, the reported range ofbilateral averages runs from 1.2 perce nt to 3.2percent and for Japan from 2.3 percent to 4.5percent. The United States has more modestdifferences, from 1.6 perc ent to 2 .3 percent.15Nontarif Barriers. Table 3 reports theprevalence of nontariff barriers for 34 coun-tries in 1999 based o n data from TRAINS andHaveman's work (rendering it more usable).The smaller number of countries availablethan for tariffs reflects previously noted limi-tations of TRA INS. We repo rt arithmetic andtrade-weighted NTB coverage ratios; the pe r-centage of tariff lines subject to NTBs. Thetrade-weighted NTB coverage ratios general-ly exceed the a rithmetic average NTB cover-age ratios, often considerably so. Forexample, on the narrow definition (definedbelow), the U.S. arithmetic NTB coverage is1.5 percent, while the trade-weighted N TBcoverage is 5. 5 perc ent. This reflects th e factthat NTBs tend to fall on important tradedgoods such as textiles and app arel.Narrow NTB coverage is basically priceand quantity control measures and qualitycontrol measu res, while broad coverage is thenarrow classification plus threat measuresrelated to antidumping. Threats are signifi-cant, since Robert Staiger and Frank Wolak(1994 ) have shown that the th reat ofantidumping impe des trade considerably (onaverage for th e U nited States they find a sev-enteen-percent reduction in trade due to anongoing investigation). Narrow NTBs coverless than 10 perc ent of trade for the richcountries and modest amounts of trade inmost except for Argentina and Brazil. Table 3implies that antidumping is quite common.Fo r example, based on th e TRAINS data, theUnited States in 1999 had 1.5 percent of tar-iff lines subject to NTB when narrowlydefined b ut 2 7.2 perc ent of tariff lines subject

l5 Harrigan's d ata is available to the public. Huiwen Laiand D aniel Trefler (2002 ) are notable for compiling thr ee-digit bilateral tariffs for fourteen importers and 36exporters for 1972, 1977, 1982, 1987, and 1992. Theyemphasize the hard work involved and have not yet madethe data public.

to NTB when broadly defined. These meas-ures do not include withdrawn suits, whichare presum ptively restrictive since th ey facil-itate collusion. Table 3 also shows wide vari-ability over countries in the u se of NTBs. T hebroad definition with arithmetic averages hasSouth Africa at 1 1.3 percent and Argentina at71.8 percent.T he use of NTBs is concentrated in a fewsectors in most economies. Table 4 reportssectoral NTB coverage ratios for the UnitedStates, European Union, Japan, and Canadafor 1999. NTBs are widely used by devel-oped countries in food products (for exam-ple, the trade-weighted NTB coverage ofagriculture, forestry and fishery products is74 perc ent for the U nited States and 24 per-cen t for the Europ ean Union ), textiles/appar-el (7 1 percent for the U nited States and 42percent for the European Union), wood andwood products (39 percent for the UnitedStates and 26 percent for the EuropeanUnion ) and in some oth er areas of manufac-turing. The products involved are quite sig-nificant in the trade of developing countriesbu t also somewhat significant in th e trad e ofdeveloped countries with e ach other.In comparison to tariffs, NTBs are con-cen trated in a smaller numb er of sectors andin those sectors they are mu ch m ore restric-tive. Deardo rff and Stern (1998) survey mostof what limited data is available on quotalicense prices. Table 5 (based on tab le 3 .6 ofDeardo rff and Stern 1998 ) gives estimates oftax equivalents based on annual averages ofweek1 Hon g Kong MFA quo ta l icense(which are themselves averages of

l6 Wh ere m arkets subject to quotas are thick and well-organized an d behavior of all agents is competitive, quo talicense prices provide the best evidence of tariff equiva-lents. Using license price data even under these assump-tions forces the analyst to face the many dimensions onwhich the quo ta is not equivalen t to a tariff-daily pricequote s exhibit within-year variation with ec onomically sig-nificant p atterns (seasonality, year-end jumps and drops)-such that no single index of them can generally beequivalent to a tariff. Neve rtheless, average license pricesin combination with the substantial rent-retaining tariffsfound on most quota-constrained products provide a use-ful measure of the restrictiveness of quotas.


11/62

TABLE 3NOY-TARIFF. \ R R I E R S - ~ ~ ~ ~

NTB ratio T\f7NTB ratio NTB ratio T\f7NTB ratio(narrow) (narrow) (broad) (broad)

AlgeriaArgentinaAustraliaBallrainBhutanBoliviaBrazilCanadaChileColombiaCzech RepublicEcuadorEuropean UnionGuatemalaH u n g a ~IndonesiaLebanonLithuaniaMexicoMoroccoNew ZealandOmanParaguayPeruPolandRomaniaSaudi ArabiaSlo\-eniaSouth AfricaTaiwanTunisiaUruguayUSAVenezuelaXotes : Tlle data are from UNCTAD's TRAINS database (Haveman repackaging). Tlle "narrow" cat ego^ includes, quantity price, quality and ad\-ance payment NTBs, but does not include threat measures such as antidumping in\-estigations and duties. The "br oa d categon includes quantity price, quality, advance payment and threat measures. Tlle ratios are calculated based on six-digit HS categories. 4 "-" indicates that trade data for 1999 are not available.


12/62

Anderson a nd va n Win coop: Trade CostsTABLE 4

NTH COVERAGEATIOSBY SECTOR-^^^^

United States 1999 EU-1 2 1999 Japan 1996 Canada 1999Narrow Broad Narrow Broad Narrow Broad Narrow Broad

NTB-ratio NTB-ratio NTB-ratio KTB-ratio KTB-ratio KTB-ratio NTB-ratio NTB-ratio

ISIC Description1 Agric., Forestry, Fish. ,011 ,052 ,719 ,743 ,001 ,001 ,229 ,2 41 ,153 ,227 ,897 ,962 ,028 ,022 ,878 ,9382 Mining, Quarcng ,000 ,000 ,018 ,099 ,001 ,0 55 ,001 ,055 ,02 8 ,008 ,193 ,706 ,000 ,000 ,02 7 ,014

21 Coal Mining ,000 ,00 0 ,000 ,00 0 ,004 ,00 0 ,004 ,000 ,00 0 ,000 ,667 1.000 ,000 ,000 ,000 ,00022 Crude Petroleum ,000 ,000 ,250 ,105 ,004 ,067 ,004 ,067 ,000 ,000 1.000 1.000 ,000 ,000 ,375 ,01923 Metal Ore Mining ,000 ,000 ,000 ,000 ,000 ,000 ,000 ,000 ,087 ,00 0 ,087 ,000 ,000 ,000 ,000 ,00029 Other Mining ,000 ,000 ,000 ,000 ,00 1 ,038 ,001 ,038 ,014 ,129 ,120 ,184 ,000 ,000 ,000 ,000

3 Manufacturing ,015 ,047 ,245 ,423 ,007 ,042 ,083 ,107 ,044 ,10 2 ,322 ,366 ,171 ,044 ,261 ,19631 Food, Bev., Tobacco ,072 ,120 ,644 ,809 ,004 ,01 1 ,489 ,474 ,185 ,329 ,925 ,893 ,185 ,34 8 ,456 ,45332 Textiles, Apparel ,000 ,002 ,509 ,70 8 ,030 ,255 ,102 ,420 ,022 ,050 ,163 ,120 ,762 ,681 ,816 ,78433 Wood, \Toad Prod. ,000 ,000 ,459 ,389 ,00 0 ,007 ,19 7 ,263 ,000 ,000 ,098 ,025 ,016 ,015 ,262 ,25234 Paper, Paper Prod. ,000 ,000 ,053 ,023 ,00 0 ,000 ,00 0 ,000 ,000 ,000 ,133 ,036 ,000 ,000 ,000 ,0003 5 Chem., Petrol Prod. ,036 ,149 ,114 ,322 ,003 ,011 ,032 ,03 3 ,048 ,169 ,635 ,750 ,013 ,00 7 ,047 ,07336 Kon-Metal Min. Prod. ,000 ,006 ,014 ,029 ,000 ,016 ,000 ,043 ,000 ,000 ,073 ,160 ,000 ,000 ,000 ,00037 Basic Metal Ind. ,003 ,044 ,006 ,04 4 ,002 ,010 ,012 ,016 ,051 ,086 ,375 ,139 ,000 ,000 ,381 ,36238 Fab. Metal Prod. ,002 ,039 ,166 ,450 ,000 ,010 ,0 05 ,012 ,032 ,057 ,095 ,266 ,000 ,00 0 ,048 ,17939 Other Manuf. ,000 ,002 ,122 ,199 ,000 ,017 ,238 ,222 ,000 ,000 ,134 ,112 ,000 ,00 0 ,073 ,012

Total All Products ,015 ,055 ,272 ,389 ,008 ,04 1 ,095 ,106 ,055 ,098 ,369 ,442 ,151 ,039 ,307 ,198Notes: "S" indicates "simple" and "TI1"' indicates "trade-weighted." Data are from UNCTAD's TRAINS database(Haveman repackaging). The "Narrow" category includes, quantity, price, quality and advance payment NTBs, but doesnot include threat measures such as antidumping investigations and duties. The "Broad" category includes quantity,price, quality, advance payment and threat measures. The ratios are calculated for two-digit ISIC categories based onthe six-digit HS classifications used by TRAINS, using HS to ISIC concordances published by the World Bank.

transactions within the week ) for textiles and the license prices to form the full tax equiv-apparel subject to quota behveen controlled alent. Th e table shows fairly high tax equiva-exporters and the United States in 1991 and lents , especially in the largest tra de1993. Th e license prices im puted for other categories (23 percent for products of broad-suppliers de pe nd on arbitrage assumptions woven fabric mills, 3 3 perc ent for appareland especially on relative labor productivity made from purchased material) . There isassumptions which may not be met. The also high variability of license prices an d taxprices are expressed as ad-valorem tax equiv- equivalen ts across comm odities (from 5 per-alents using Ho ng Kong export prices for th e cen t to 23 perc ent for textiles, from 5 per-underlying textile and apparel items, trade- cent to 33 percent for apparel) . Earlier yearsweigh ted across suppliers.17 We a dd th e reveal highe r tax equivalents.18 Th us the re isU.S. tariffs on the corresponding items to

l8 Anderson and Neary (1994) report trade-weightedaverage tax equivalents across the MFA commodity groups

l7 Because the license prices are for transfer between for a set of exporters to the United States in the mid-holders and users and are effectively subject to penalty for 1980's. The Hong Kong average exceeded 19 percent inthe holder, the implied tariff equivalents are lower bounds each year and ranged to over 30 percent in some years,to true measures of restrictiveness; this bias direction while tax equivalents (very likely biased upward) for otherprobably also applies to the intertemporal averaging. countries were much larger, some over 100 percent.


13/62

702 Journal of Economic Literature Vol. XL Il (Septemb er 2004)TABLE 5

TARIFFEQUIULEXTSF U.S. hfFA QI;oT.~s,1991 A N D 1993 (PERCE NT)Sector 1991 1993

Rent Rent S TIV Rent + %USTar Eq . Tar Eq. Tanff Tariff T\T7Tariff Imports

Textiles:Broad\vo\-en fabric millsKarro\v fabric millsYarn mills and textile finishillgThread millsFloor co\-eringsFelt and textile goods, n.e.c.Lace and knit fabric goodsCoated fabrics, not rubberizedTire cord and fabricCordage and twineKon\voven fabric

Apparel and fab, textile products:TVomen's hosiery except socksHosieT, nn.e.c.App'l made from purchased mat']Curtains and draperiesHouse furnishings, n.e.c.Textile bagsCan\-as and related productsPleating, stitching, ... e m br o i de ~Fabricated textile products, n.e.c.LuggageTVomen's handbags and purses

Xotes: "S" indicates "simole" and "71\"'indicates "trade-weirrhted." Rent eaui\-alents for U.S. imoorts from H o w0 DKong were estimated on the basis of average weekly Hong Kong quota prices paid by brokers, using informationfrom International Business and Economic Research Corporation. For countries that do not allocate quota rightsin public auctions, export prices were estimated from Hong Kong export prices, with adjustments for differencesin labor costs and productivity. Sectors and their correspolldillg SIC classifications are detailed in USITC (1995)Table D-1. Quota tariff equivalents are reproduced from Deardorff and Stem (1998), Table 3.6 (Source USITC1993,1995).Tariff a\-erages, rade-weighted tariff a\-erages and U.S. import percentages are calculated using datafrom the UNCTAD TRAINS dataset. SIC to HS concordances from the U.S. Census Bureau are used.

(i ) substantial restrictiveness of M FA quotas and Canadian agriculture. Details are dis-and (ii) very large differentials in q uota p re- cussed in section 4.mia across commodity lines and across Using a variety of methods, Messerlinexporters. (2001 ) makes a notably ambitious attem pt toPrice comparison measures confirm this assemble tariff equivalents of all trade policypicture of the high and highly concentrated barriers for the European Union. He com-nature of NTBs with data from the agricul- bines the NTB tariff equivalents with thetural sector. Eu rop ean a nd Japanese agricul- M FN tariffs. Fo r 1999 th e tariff equivalent ofture is even m ore highly protected than U.S. policy barriers were 5 percent for cereals,


14/62

703nderson and van Wincoop: Trade Costs64.8 percent for meat, 100.3perc ent for dairyand 125 percent for sugar. In mining, the tar-iff equivalent was 71.3 percent. The com-bined arithmetic average protection rate was31.7 percent in agriculture, 22.1 percent intextiles, 30.6 percent in apparel, and muchless in other industrial goods. The combinedarithmetic average protection rate in industri-al goods was 7 .7 percent. W e use 7.7 percentfor our rep resentative trad e policy cost.2.2 Transport C osts

Direct transport costs include freightcharges and insurance which is customarilyadded to the freight charge. Indirect trans-port user costs include holding cost for thegoods in transit, inventory cost due tobuffering the variability of delivery dates,preparation costs associated with shipmentsize (full con taine r load vs. partial loads), an dthe like. Indirect costs must be inferred.2.2.1 Measurement Problems andLimitations

Th ere are three main sources of data fortransport costs. The most direct is industryor shipping firm information. Nuno Limaoand A nthony Venables (2001 ) obtain q uotesfrom shipping firms for a standard co ntainershipped from Baltimore to various destina-t ions. David H umm els (2001a) obtainsindices of ocean shipping and air freightrates from trade journals which presumablyare averages of such quo tes. Direct m ethodsare best but not always feasible due to datalimitations and the very large size of theresulting d atasets.

Alternatively, there are two sources ofinformation on average transport costs.National customs data in some cases allowfine detail. For example, the U.S. Censusprovides data on U.S. imports at th e ten-digitHarm onized System level by exporter coun-try, mode of transport and entry port, byweight (where available) and valued at f.0.b.and c.i.f. bases. Dividing the former valueinto the latter yields an ad-valorem estimateof bilateral transport cost. Humm els (2001 a)

makes use of this source for the UnitedStates and several other coun tries. Th e mostwidely available (many countries and yearsare co vered ) but least satisfactory average advalorem transport costs are the aggregatebilateral c.i.f.1f.o.b ratios p rod uc ed by t heIMF from matching export data (re ortedf.0.b.) to import data (reported c.i .f .)J 9 Th eIM F uses the UN's COM TRA DE database,supplemented in some cases with nationaldata sources. Hummels (2001 a) points outthat a high proportion of observations areimp uted ; this and the compositional shifts inaggregate trade flows which occur over timelead him to co nclude that "quality problemsshould disqualify these data from use as ameasure of transportation costs in evensemi-careful studies."202.2.2 Evidence on Transport Costs

Hum mels (2001a) shows the wide disper-sion in freight rates over commodities andacross countries in 1994. The all-commodi-ties trade-weighted average transport costfrom national customs data ranges from 3.8percent of the f .0.b. price for the UnitedStates to 13.3 perc ent for Paraguay. The all-commodities arithm etic average ranges from7.3 percent for Uruguay to 17.5 percent forBrazil. Th e U.S. average is 10 .7 perc ent.Across commodities for the United Statesthe range of trade-we ighted averages is fromless than 1 percent (for transport equip-me nt) to 27 p ercent for crude fertilizer. Th earithmetic averages range from 5.7 percentfor machinery and transport equipment to15.7 percent for mineral fuels.

Hu mm els (1 999) considers variation overtime. The overall trade-weighted averagetransport cost for the United States declinedover the last thirty years, from 6 perc ent to 4l9 See the Direction of Trade Statzstics a nd theli~t enl ati ona l inancial StatisticsNevertheless, because of then avallabihty and thedifficulty of obtaining better estimates for a w d e range ofcountries and years, apparently careful work such asHarrigan (1993) and Scott Baier and Jeffrey Bergstrand

(2001)uses the I M F data.


15/62

704 Journal of Economic Literature Vol. XL lI (Septemb er 2004)percent. Composition problems are acutebecause world trade in high-value-to-weightmanufactures has grown much faster thantrade in low-value-to-weight primary prod-ucts. Hummels shows that air freight costhas fallen dramatically, while ocean shippingcost has risen (along with the shift to con-tainerization, which improves the quality ofthe shipping senice). He also documentsthe \vide dispersion in the rate of change ofair-freight rates across country pairs over thepast forty years.Notice that alongside tariffs and NTBs,transport costs look to be comparable inaverage magnitude and in variability acrosscountries, commodities, and time. Transportcosts tend to be higher in bulky agriculturalproducts where protection in OECD coun-tries is also high. Policy protection tends tocomplement natural protection, amplifyingthe variability of total trade costs.Limao and \'enables (2001) emphasize thedependence of trade costs on infrastructure.They gather price quotes for shipment of astandardized container from Baltimore tovarious points in the world. Infrastructure ismeasured as an average of the density of theroad network, the paved road network, therail network, and the number of telephonemain lines per person. A deterioration ofinfrastructure from the median to the 75thpercentile of destinations raises transportcosts by 12 percent. The median landlockedcountry has transport costs that are 55 per-cent higher than the median coastal econo-my.21 The infrastructure variables also haveexplanator). power in predicting trade vol-ume. Inescapably, understanding trade costsand their role in determining internationaltrade volumes must incorporate the internalgeography of countries and the associatedinterior trade costs.As for indirect costs, I-Iummels (2001b)imputes a willingness-to-pay for saved time.

21 Limao an d \'enables also repo rt similar results usingth e c.i.f.1f.o.b. ratios of th e IILIF.

Each day in travel is on average worth 0.8percent of the value of manufacturedgoods, equivalent to a 16-percent ad-val-orem tariff for the average-length oceanshipment. Shippers switch from ocean to airin his mode choice model when the full(shipping plus time) cost of ocean exceedsthat for air,22 The use of averages here"masks a lot of variation in the estimatedvalue of time across two-digit manufactur-ing sectors, and is subject to upward aggre-gation bias due to larger growth in tradewhere savings are greatest due to the sub-stitution effect. Infrastructure is likely tohave a considerable effect on the time dostsof trade.In 1998, half of U.S. shipments was by airand half by boat.23 Assigning one day toshipment by air anywhere in the world, asHummels does, and using the twenty-dayaverage for ocean shipping, leads to anaverage 9-percent tax equivalent of timecosts. Hummels argues that faster transport(shifting from shipping to air, and fasterships) has reduced the tax equivalent oftime costs for the United States from 32

percent to 9 percent over the period1950-98.24We combine %percent time costs and10.7 percent U.S. average direct transportcosts for our representative full transportcost of 21 percent (1.107*1.09-1)." Linking p ort of en try for U.S. imports with th e trav-el time to the exporter (a county-averag e of times to theexporter's po rts ), he creates a matrix of ocean shippin gtimes. Air freight is assumed to take one day for pointsanywhere in the world."This ignores trucking and rail modes, which areimportant to trade with Canada and Mexico, the twolar est trade partners of the United States."T he calculation is based on observing that U.S.impo rts, excluding Canada a nd Mexico, had 0-percent airshiprnent in 1950 and 50-percent air shipment in 1998.The average ocean shiprnent time was halved from fortydays to twent). days over the same tim e. Th e net effect is as a ~ i n gof 29.5 days, equal to forty days for 195 0 minus 10 .5days for 1998. The latter is equal to .5 times hvent). daysfor ocean shipping plus .5 times on e day for air freight. Formanufacturing, at 0.8 per cent ad valorem per day for thevalue of time in shipping, the s a ~ in g f 29.5 days is worth

a fall from 3 2 percent to 9 perce nt ad valorern.


16/62

Anderson and van Winco op: Trade CostsTABLE 6DISTRIBUTIONARGINSOR HOUSEHOLDONSUMPTIONND CAPITALGOODS

Select Aus. Bel. Can. Ger. Ita. Jap. Net. UK USProduct Categories 95 90 90 93 92 95 90 90 92Rice 1.239 1.237 1.867 1.423 1.549 1.335 1.434 1.511 1.435Fresh, frozen beef 1.485 1.626 1.544 1.423 1.605 1.681 1.640 1.390 1.534Beer 1.185 1.435 1.213 1.423 1.240 1.710 1.373 2.210 1.863Cigarettes 1.191 1.133 1.505 1.423 1.240 1.398 1.230 1.129 1.582Ladies' clothing 1.858 1.845 1.826 2.039 1.562 2.295 1.855 2.005 2.159Refrigerators, freezers 1.236 1.586 1.744 1.826 1.783 1.638 1.661 2.080 1.682Passenger vehicles 1.585 1.198 1.227 1.374 1.457 1.760 1.247 1.216 1.203Books 1.882 1.452 1.294 2.039 1.778 1.665 1.680 1.625 1.751Office, data proc. mach. 1.715 1.072 1.035 1.153 1.603 138 9 1.217* 1.040 1.228Electronic equip. , etc. 1.715 1.080 1.198 1.160 1.576 1.432 1.224* 1.080 1.139Simple Average(125 categories) 1.574 1.420 1.571 1.535 1.577 1.703 1.502 1.562 1.681

Notes: The table is reproduced from Bradford and Lawrence, "Paying the Price: The Cost of FragmentedInternational Markets", Institute of International Economics, forthcoming (2003).Margins represent the ratioof purchaser price to producer price. Margms data on capital goods are not available for the Netherlands, so anaverage of the four European countries' margins is used.

2.3 Wholesale and Retail Distribution Costs show that their input-output estimates ofU.S. distribution costs are roughly consis-

Wholesale and retail distribution costs tent with survey data from the U.S.enter retail prices in each country. Since Department of Agriculture for agriculturalwholesale and retail costs vary widely by goods and from the 1992 Census ofcountry, this would appear to affect Wholesale and Retail Trade. For other G-7exporters' decisions. Local trade costs apply countries they report distribution costs into both imported and domestic goods, how- the range of 35-50 percent.ever, so relative prices to buyers don't Scott Bradford and Robert Lawrencechange and neither does the pattern of (2003) use the same input-output sources totrade. Section 3 gives a formal argument. measure distribution costs for the UnitedSection 4 discusses the effect of distribution States and eight other industrialized coun-margins on inference about international tries, but instead divide by the producertrade costs from retail prices. price, consistent with the approach in this

Ariel Burstein, Joao Neves, and Sergio survey of reporting trade barriers in terms ofRebelo (2003) construct domestic distribu- ad valorem tax equivalents. Table 6 reportstion costs from national input-output data distribution costs for selected tradablefor tradable consumption goods (which household consumption goods and an arith-correspond most closely to the goods for metic average for 125 goods. The averageswhich narrowly defined trade costs are rel- range over countries from 42 percent inevant). They report a weighted average of Belgium to 70 percent in Japan. Average41.9 percent for the United States in 1992 U.S. distribution costs are 68 percent of pro-as a fraction of the retail price. They also ducer prices. The range of distribution costs


17/62

706 Journal of Economic Literature V ol. X L I I (September 2004)is much larger across goods than acrosscountries, for example running from 14 per-cent on electronic equipment to 216 percenton ladies clothing in the United Wetake 55 percent as our representativedomestic distribution cost, close to the trad-able consumption goods average of OECDcountries.

3. Inference of Trade Costsfrom Trade F1ou;sTrade costs can be inferred from an eco-

nomic model linking trade flows to observ-able variables and unobservable trade costs.Inference has mainly used the gravity

The economic theor). of gravity isdeveloped here extensively, revealing newproperties that clarify procedures for goodempirical work, reporting results, and doingsensible comparative statics. Our develop-ment embeds gravity within the classic con-cerns of trade economists with theequilibrium allocation of production andexpenditure within nations.

A variety of ad hoc trade cost functionshave been used to relate the unobservablecost to observable variables. Plausibly, costfalls with common language and customs,better information, better enforcement andso forth. Many implementations imposerestrictions that seem implausible. Someproxies may not be exogenous, such asmembership in a currency union or regionaltrade agreement; their effects will not beuniform; functional form is often too simpleand so forth. Further economic theory isneeded to identify the underlying structureof trade costs.

"j Iye sshould warn that some of these distribution costestimates include rents in the form of lnonopolisticmarkups rather than actual costs incurred by the local dis-tribution sector. Although it n~ould e desirable to take outthese rents, no data are available to do so." '4 subset of the gravity literature uses it to discrimi-nate among theories of the determinants of trade. It is not

very well-suited for this purpose, since many trade models\\ill lead to gra~ity Deardorff 1998).

3.1 Traditional GravityMost estimated gravity equations take the

form

where xg is the log of exports from i to j, y,and yJ are the log of GDP of the exporterand importer, and zy (m=l , .. . M) is a set ofobservables to which bilateral trade barri-ers are related. The disturbance term is E, .A large recent literature developed estimat-ing this type of gravity equation after a sur-prising finding by McCallum (1995) thatthe U.S.-Canada border has a big impact ontrade." McCallum estimated a version of(1)for U.S. states and provinces with two zvariables: bilateral distance and an indica-tor variable that is equal to one if the tworegions are located in the same country andequal to zero otherwise. He found thattrade between provinces is more thantwenty times trade between states andprovinces after controlling for distance andsize. The subsequent literature has oftenadded so-called remoteness variables,which are intended to capture the averagedistance of countries or regions from theirtrading partners.

Gravity equations can be derived from avariety of different theories. None lead to tra-ditional gravity, despite this literature's use ofreferences such as Anderson (1979) andDeardorff (1998) o justify estimation of3.2 Theoy-Based Gravity

Gravity-like structure obtains in a wideclass of models, those where the allocationof trade across countries can be analyzed

" See Michael Anderson and Stephen Smith (1999a,b),Natalie Chen (2002), Carolyn Evans (2003a,b,c), JohnHelli\vell (1996, 1997, 1998), Hellin~ell and JohnMcCallum (1995), Helli\vell and Genevieve F7erdier(2001), Russell Hillberry (1998, 2002), 1701ker Nitsh(2000), Shang-Jin \17ei (1996), and Holger LVolf (2000a,b).

For recent surveys of gravity theory, see Harrigan(2002), and Feenstra (2002,2003).


18/62

707nderson and c an W incoop: Trade Costsseparately from the allocation of productionand consumption within countries.29 Let{Y:, Elk] be the value of production andexpenditure in country i for product class k.A product class can be either a final or anintermediate good. A model is trade separa-ble if the allocation of {ytkk, for each coun-tk}try i is separable from the bilateral allocationof trade across countries.

Trade separability obtains under theassumption of separable preferences andtechnology. Each product class has a distinctnatural aggregator of varieties of goods dis-tinguished by country of origin. Thisassumption allows the two-stage budgetingneeded to separate the allocation of expen-diture across product classes from the allo-cation of expenditure within a product classacross countries of origin.3o What exactly isimplied by product differentiation within aproduct class will differ depending on theproduct class and its myriad characteristics.For inputs of physically standardized prod-ucts (grains, petroleum, steel plate), differ-entiation may be on terms of delivery orsubtle qualities that affect productivity whilebranded final goods suggest hedonic differ-entiation over varieties of clothing or per-fume. Imposing some form of differentiationis unavoidable if empirical trade models areto fit the bilateral trade data, as Armingtonnoted long ago.

The class of trade separable modelsyields bilateral trade without having tomake any assumptions about what specificmodel accounts for the observed outputstructure {Y:} and expenditure allocations

"The general equilibrium of trade in a many-countryworld is enorlnously simplified by this natural assumption.

30Explicitly, separabil iv restricts the dual cost functionc ( p , w , y )where p is a vector of traded input prices and w isa vector of non-traded primary input prices n~hi le is thelevel of output. The separability restriction is imposed inc ( p , u - , y ) = f l g ( p ) , u - , y ] where g ( p ) is a homogenous ofdegree one and concave function of p By Shephard'sLemma,f, is the aggregate demand for the traded inputclass while f,g, is the demand vector for the individualproducts. A similar structure characterizes the assumptionof separability in final demand.

{ ~ , k } . ~ lilateral trade is determined in con-ditional general equilibrium whereby prod-uct markets for each good (each brand)produced in each country clear conditionalon the allocations {ytkk,tk]. nference abouttrade costs is based on this conditional gen-eral equilibrium. Comparative static analy-sis, in contrast, requires consideration ofthe full general equilibrium. A change intrade barriers, for example, will generallyaffect the allocations {Y:, E:].

Two additional restrictions to the class oftrade separable models yield gravity. Theseare: the aggregator of varieties is identicalacross countries and CES; and ad-valoremtax equivalents of trade costs do not dependon the quantity of trade. The CES formimposes homothetic preferences and thehomogeneity equivalent for intermediateinput demand. These assumptions simplifythe demand equations and market clearingequations. These two restrictions can berelaxed in various useful ways discussedbelow. Our derivation provides much morecontext to Deardorff's (1998) remark, "I sus-pect that just about any plausible model oftrade would yield something very like thegravity equation .. .".If X; is defined as exports from i to j inproduct class k, the CES demand structureimplies (under the expositional simplifica-tion of equal weights for each country oforigin)

where a, is the elasticity of substitutionamong brands, pi is the price charged by ifor exports toj and P; is the CES price index:

" Specifically it does not matter what one assumesabout production functions, technology, the degree ofcompetition, or specialization patterns. The n atur e of pref-erences and technology that gives rise to the observedexpenditure allocations E: also does not matter.


19/62

70 8 Journal of Econoinic Literature Vo l. XLII (Sep temb er 2004)

The assumption that trade costs are pro-portional to trade implies that the price p:can be written as pft,;, where p f is the "sup-ply price" received by producers in country kand ti-1 is the ad-valorem tax equivalent oftrade costs, independent of volume.If supply is monopolistic, the "supplyprice" is the product of marginal productioncost and the markup. So long as the markupis invariant over destination^,^^ t i containsonly trade costs. Otherwise the tax equiva-lent must be interpreted to contain markups.With competitive supply, this issue does notarise. Markups arising from monopolypower in the distribution sector itself are amore important issue with interpreting ti-1as a cost.Imposing the market-clearing conditions

for all i and k yields gravity. Solve for thesupply prices p: from the market-clearingconditions and substitute the result in (2)and (3).This yields the system

where yk is world output in sector k . Theindices P; and flk can be solved as a function'3 2 LVith Cournot competition, the markup is invariant

over destinations in symmetric rnonopolistic competition.Generally it is equal to l/ [l -l /~ (l -s ,~ )] is the mar-here sket share of i inj.

of trade barriers (t:] and the entire set( Y : , E ~ ] .rade flows therefore also dependon trade barriers and the set ( Y ; , E , ~ ) . ~ ~A wide variety of production and expen-diture models may lie behind (E; ,Y,~} .nparticular, some of the E's and 2"s may beequal to zero. Previous derivations of gravi-ty have usual1 made much more restrictiveassumptions.JThe main insight from the theory is thatbilateral trade depends on relathe trade bar-riers. The key variables fl: and P; are out-ward and inward multilateral resistance,respectively. They summarize the averagetrade resistance between a country and itstrading partners in an ideal aggregation

sense, which we develop below. Basicdemand theory suggests that the flow of goodi in to j is increased (givenu>l)by high tradecosts from other suppliers toj as captured byinward multilateral resistance P;. But, lessobviously, high resistance to shipments fromi to its other markets, caP ured in outwardmultilateral resistance Hi,ips more tradeback into i's market in j. Gravity gives theseinsights an elegant1 sim le form in ( 5 ) .Y P

The trade cost tV may include local distri-bution costs in the destination market, butthose domestic costs do not affect trade flows.This rather surprising and important resultfollows from basic gravity theory. Suppose forany goods class k that each destination j hasits own domestic margin m" If we multiply all

""Theoretical gravity equations in Anderson (1979),Jeffrey Bergstrand (1985,1989,1990)and Scott Baier andBergstrand (2001) look far more complicated than (3 ,with a large number of prices and price indices. Andersonand van Lf'incoop (2003) show that all these prices can besurnrnarized by just two price indices, one for theimporter and one for the exporter, which are solved as afunction of trade barriers and total supply and dernand ineach location.

34 The first paper to formally derive a grality equationfrom a general equilibrium model with trade costs isAnderson (1979), who assumes that every country pro-duces a particular variety. Bergstrand (1989, 1990) andBaier and Bergstrand (2001) derive gravity equations inmodels with rnonopolistic competition, endogenizing vari-ety. Anderson and van \\'incoop (2003) assume each coun-try has an endo\vment of its good.


20/62

709nderson and uan W incoo p: Trade Costsnarrowly defined trade barriers tlJ by domes-tic trade costsmk in the destination market, itis easily verified from (6)-(7) that the P aremultiplied by m;, the are unchanged, andtherefore trade flows are also unchanged.The invariance of trade patterns to domes-tic distribution costs that apply to all goodshas another important practical implication.We can only identify relative costs with thegravity model. One way to interpret aninferred system of trade costs {ti] s to I icksome region i and normalize t,=l.Essentially this procedure treats the tradecost of i with itself as a pure local distribu-tion cost and divides all other trade costs bythe local distribution cost in region i.

As in most gravity papers, Anderson andvan Wincoop (2003) consider a one-sectoreconomy. They show that when consumershave CES preferences with common elas-ticity a among all goods, the gravity equa-tion can be written, omitting thesuperscripts k , as

where Y, and Y are levels of GDP, Y isworld GDP, and 0, is the income share ofcountry i. This is a special case of (5) withexpenditure equal to output because it is aone-sector economy (E,=Y,).With symmetryof trade costs (t,=t,,), n,=P, .A couple of comments are in order abouthomogeneous goods trade. When we let theelasticity of substitution ak in (5)-(7) go toinfinity, trade converges to that in a homoge-neous goods model. However, no informa-tion about trade barriers can be inferred. Asan example, consider a two-country modelwith trade in a homogeneous good k. Ifcountry 1exports k to country 2, its export is

equal to ~,k-Y,k. Gravity equation ( 5 ) ,accounting for multilateral resistance,indeed converges to this for the two-count17case as a, approaches infinity, assuming anynon-zero international trade barrier (andnormalizing domestic barriers to zero).Since the bilateral trade flow in the condi-tional general equilibrium does not dependon the magnitude of trade barriers, nothingcan be learned about trade barriers. Moregenerally, it is difficult to learn much abouttrade barriers from a gravity equation forsectors where the elasticity of substitution isextremely high. The expressions t,:-"hnwhich trade flows (directly and through mul-tilateral resistance) depend are virtually zerowhen a, is very high as long as the trade bar-rier is positive (t:>l). Conditional on a tradebarrier being positive, the size of the tradebarrier does not matter much to the patternof trade flows in the conditional generalequilibrium, hence we cannot learn muchabout their size.Several authors have derived gravity equa-tions for homogeneous goods trade whentrade is an aggregate of a variety of homoge-neous goods. Deardorff (1998) derived agravity equation in the Heckscher Ohlinmodel with complete specialization. This isessentially a differentiated goods modelthough, with each country producing a dif-ferent brand. It does not mean much to saythat a good is homogeneous when there isonly one producer.A real homogeneous goods model, withmultiple producers of the same homoge-neous good, is the Ricardian model ofJonathan Eaton and Samual Kortum (2002).Their model leads to a gravity equation foran aggregate of homogeneous goods. It isalso a model with trade separability,although the rationale is somewhat different.Production is Ricardian, with the cost of pro-duction in country i in good k given byc,lz(k)where z(k) is the realization of tech-nology in good k, an element in a continuumof goods. Productivity is drawn from aFrdchet distribution. The distribution has


21/62

710 Journal of Economic Literature Vol. X L I I (September 2004)two parameters. The first is T,,with higher T ,meaning a higher average realization forcountry i . The second is 8, with a largervalue implying lower productivity differ-ences across countries. For a particulargood, users always buy from the cheapestsource. The price is the production costtimes the trade cost t,. Each good is pro-duced with both labor and a bundle of inter-mediate goods that consists of the same CESindex of all final goods as the utility functionover final goods.Since there is a continuum of goods andthe setup is the same for all goods (same pro-duction function, same productivity distribu-tion, same trade cost), the fraction thatcountnj spends on goods from i is equal tothe probability for any particular good thatjsources from i. With the assumed Frdchetdistribution this is equal to

The probability of shipment from countryi is lowered by the trade cost of getting thegood to country j , relative to the averagetrade cost of shipping from all other destina-tions, and lowered by a higher cost of labor.The same mathematical representation ofthe allocation of trade arises as with the CESstructure of demand for goods differentiatedby place of origin. This equation is the sameas (2), with a=8+1 and pI1-" replaced byT,cll-".The p, is essentially replaced by c,,which can be solved in the same way fromthe conditional general equilibrium. Thisgives rise to the same gravity equation asbefore.l5It is worth noting that gross output is nowlarger than net output due to the input ofintermediates. The output in the gravityequation (8) is gross output. Since Eatonand Kortum assume that intermediates are afraction 1-P of the production cost c,, with

ji Eaton and Kortum onl) derwe a gra \l tv spe c~f icat~ onfo r X /X

labor a fraction P, gross output is 11P timesvalue added. If we interpret Y, in (8) asvalue added, the gravity equation must bemultiplied by 1IP.3.3 The Trade Cost Function

The theoretical gravity model allows infer-ence about unobservable trade costs by (i)linking trade costs to observable cost proxiesand (ii) making an assumption about errorterms which link observable trade flows totheoretically predicted values. Here wefocus on (i ), the next section deals with (ii).For now we will focus on inference abouttrade barriers from the aggregate gravityequation (8). In a section about aggregationbelow we will return to the disaggregatedgravity equation (5).Bilateral trade barriers are assumed tobe a function of observables z;, commonlyloglinear:

W

Normalizing such that z$=1 measures zerotrade barriers associated with this variable,(z;)ril is equal to one plus the tax equivalentof trade barriers associated with variable m .The list of observable arguments 2:;' whichhave been used in the trade cost function inthe literature includes directly measuredtrade costs, distance, adjacency, preferentialtrade membership, common language, and ahost of others. Gravity theory has used arbi-trary assumptions regarding functional formof the trade cost function, the list of vari-ables, and regularity conditions.

As an illustration of the functional formproblem, consider distance. By far the mostcommon assumption is that t,=d;. GeneGrossman (1998) argues that a more reason-able assumption is that ~,=t,- l=d; since onecan think of z, as transport costs per dollar ofshipments. Hummels (2001a) estimates thep in the second specification by using data onad-valorem freight rates and finds a value ofabout 0.3. Limao and Venables (2001) esti-mate the first specification using c.i.f.1f.o.b.


22/62

Anderson and uan W incoop: Trade Costsdata and also find an estimate of p of about0.3. Although these numbers are the same,they are inconsistent with each other. If theGrossman specification is correct with p=0.3,one would expect a distance elasticity of t, of0.3ziJ(1+q1), evaluated at some average z,which is much less than 0.3. Highly mislead-ing results for trade barrier estimates arisewhen the wrong functional form is adopted.Eaton and Kortum (2002) generalize thetreatment of distance with a flexible formwhich can approximate both of the precedingspecifications. They assume that there are dif-ferent trade barriers for six different distanceintervals. While implicitly they still assume aparticular functional form, in the form of astep function, this spline approach is likely tobe more robust to specification error.Another functional form issue is that themost common setup (11)is multiplicative inthe various cost factors. Hummels (2001a)argues that an additive specification is moresensible. A multi-dimensional generalizationof the approach by Eaton and Kortum(2002) may be applied, although there is atradeoff between degrees of freedom andgenerality of the specification. To the extentthat theory has something to say about thefunctional form, it is preferable to use thisinformation over econometric solutions thatwaste degrees of freedom.The second problem is which observablesto include. Empirical practice can improvewith a more theoretical approach to the z's.Especially for abstract trade barriers such asinformation costs, it is often unclear whatspecific variables are meant to capture. Evenin the absence of a specific theory, it is usefulto ponder the relationship between tradebarriers and observed variables. For example,common empirical practice uses a languagevariable that is one if two countries speak thesame language and zero otherwise. JacquesMelitz (2003) considers ways in which lan-guage differences affect trade and developsseveral variables that each capture differentaspects of communication. One such variableis "direct communication," which depends on

the percentages of people in two countriesthat can speak the same language. Another isthe binary variable "open-circuit communica-tion," which is one if two countries have thesame official language or the same languageis spoken by at least 20 percent of the popu-lations of both countries. The first variablereflects that trade requires direct communi-cation, while the second variable is meant tocapture an established network of translation.Another example is distance. It is common tomodel distance as the Great Circle distancebetween capitals. Where these differ fromcommercial centers it is sometimes taken tobe superior to use distance between com-mercial centers. But then what of countrieswith more than one commercial center, ofinterior inf ra st r~ ct ur e? ~~Implausibly strong regularity (commoncoefficients) conditions are often implicitlyimposed on the trade cost function. Forexample, the effect of membership in a cus-toms union or a monetary union on tradecosts is often assumed to be uniform for allmembers. As for customs unions, a uniformexternal tariff is indeed approximately thetrade policy (though NTBs remain inherent-ly discriminatory), while free trade agree-ments continue to have different nationalexternal tariffs and thus different effects. Asfor monetary unions, the effect of switchingfrom national to common currencies is like-ly to be quite different depending on thenational currency. Similar objections can beraised to a number of the other commonlyused proxies z y such as common language oradjacency dummies. NTBs present an acuteform of this problem. The effect on tradebarriers of NTBs in a country i will general-ly vary across trading partners j, goods k , andtime t . Regression residuals reflect theNTBs but also random error. To identify thetariff equivalent of NTBs Harrigan (1993)

36 Some investigators (e.g., Bergstrand) measure bilat-eral distance between ports, supplemented by twice theland distance between ports and commercial centers,reflecting the rough difference in cost between water andland shipment.


23/62

712 Journal of Economic Literature Vol. XLIl (Septeinber 2004)assumes, not very plausibly, that the import-ing country's NTB has the same trade dis-placement effect for each exporter i that itbuys good k from. Trefler (1993) assumeseven less plausibly that U.S. NTBs have thesame trade-reducing effect for all goods kthat it imports from the rest of the world. 37We sympathize with efforts to identifytrade costs with simple forms of (11). Ourcriticism of the ad hoc functional form andthe regularity assumptions aims to stimulateimprovement in estimation and useful com-parative statics. Unpacking the reducedform to its plausible structural elements willaid both.3.4 Estimation of Trade Barriers

Given the trade cost function, the loga-rithmic form of the empirical gravity equa-tion becomes (dropping the constant term)

where x,=ln(X, ), y,=ln(Y,),and Am=(l-o)~,,r, and y , are oiservables, and 5 is the errorterm.The theoretical gravity equation can beestimated in three different ways.38Anderson and van Wincoop (2003) estimatethe structural equation with nonlinear leastsquares after solving for the multilateralresistance indices as a function of theobservables 2,; (bilateral distances and adummy variable for international borders)and the parameters A,,. Another approach,

37 Both authors implicitly impose a further regularit).condition. NTB coverage ratios for each good are theexplanatory variable, so all changes in this ratio areassumed to be equally important.

38 A potential fourth method is to infer all bilateraltrade barriers for a group of countries or regions from theresiduals of the bilateral trade flows from the prediction ofthe frictionless gravit). model. The information in thismeasure would be drowned in random error, however:there is an unboundedly large confidence intend aboutthe point estimates because all degrees of freedom areused up.

which also gives an unbiased estimate of theparameters A,,,, is to replace the inward andoutward multilateral resistance indices andproduction variables, y,-(1-o)lnP, andy,-( 1-o)ln(Il, ), with inward and outwardregion-specific dummies. With symmetry, asingle set of region-specific dummies suffice.This approach is adopted by Anderson andvan Wincoop (2003), Eaton and Kortum(2002),Asier Minondo (2002),Andrew Roseand van Wincoop (2002), and Hummels(2001a). Keith Head and Thierry Mayer(2001) and Head and John Ries (2001) fol-low an estimation approach that is identicalto replacing multilateral resistance variableswith country dummies in the case whereinternal trade data X,, exist for all regions orcountries. Assuming that z,=zJ,, it followsfrom (12) that

The parameters Ak can then be estimatedthrough a linear regression.A third method is to use data for the priceindices and estimate with OLS. Thisrequires data on price levels for a cross-sec-tion regression or changes in price indiceswhen there are at least two years of data.The latter is the approach taken byBergstrand (1985, 1989, 1990), Baier andBergstrand (2001), and Head and Mayer(2000).As discussed in Baier and Berstrand(2001), it is often hard or impossible tomeasure the theoretical price indices in thedata. Price indices, such as the consumerprice index, also include nontradables andare affected by local taxes and subsidies.Nominal rigidities also affect observedprices, and have a big impact on relativeprices when combined with nominalexchange rate fluctuations. Anderson andvan Wincoop (2003) also argue that certaintrade barriers, such as a home bias in prefer-ences, do not show up in prices. Similarly,Deardorff and Stern (1998) explain why cer-tain NTBs affect trade but not prices.


24/62

713nderson and van Winc oop: Trade CostsFeenstra (2003) sums it up by writing that"the myriad of costs . . . involved in makingtransactions across the border are probablynot reflected in aggregate price indices . . .".This does not mean that prices of individualtradable goods are entirely uninformativeabout trade costs. We turn to that topic insection 4.The tax equivalent of trade barriersbetween i andj associated with variable m isestimated as

This shows that we need an estimate of CJin order to obtain an estimate of trade barri-ers. Assumptions about o can make quite adifference. For example, the estimated taxequivalent is approximately nine times larg-er when using o =2 instead of o=10.Gravity can only measure trade barriersrelative to some benchmark, as noted above.The literature tends to compare trade barri-ers between countries to barriers withincountries, or barriers between regions tobarriers within regions. This is problematicsince different countries or regions have dif-ferent barriers for internal trade. Theresults will also depend on the measure ofbarriers within a region or country, such asthe treatment of distance within a countryor region. This is essentially an aggregationproblem since a country or region is itself anaggregate. Head and Mayer (2001),Helliwell and Verdier (2001), and RussellHillberry and David Hummels (2002a)show that the different measures of internaldistance can make a big difference for theresults.The error term in the empirical gravityliterature is usually taken to reflect meas-urement error. Trade flow data are notori-ously rife with measurement error,39 whichis taken to justify a normally distributedadditive error term which is orthogonal to

39 For example, mirror-image trade flow data do notmatch

the regressors. More careful considerationof measurement error is likely to be produc-tive. Some recent implementations recog-nize the panel nature of the data inconstructing standard errors (all bilateralobservations from i or into ma have com-mon disturbance elements).lJ Improvedeconometric techniques based on carefulconsideration of the error structure are like-ly to pay off. Recent literature on spatialeconometrics (e.g. H. H. Kelejian andIngmar Prucha 1999) may be helpful.The error term may also reflect unobserv-able variables in the trade cost function(11). If the trade cost t , is multiplied by eEy,there is again an additive error term in theempirical gravity equation (12). In this casethe structural estimation technique dis-cussed above would have to be modifiedsince the multilateral resistance variablesalso depend on these error terms. But thesecond estimation technique, replacing themultilateral resistance variables with regionor country-specific dummies, is still appro-priate. Eaton and Kortum (2002) introducethe error term in this way and adopt the sec-ond estimation method.Non-orthogonality of the error term hastwo sources, omitted variables and endo-geneity. If the error term reflects omittedtrade frictions which are correlated with theincluded Z: variables, it causes estimationbias. Endogeneity problems arise in severalways. Concerned that trade can affect out-put, many papers use instrumental variablesfor output Y, and Y,. The most commoninstruments are population or factor endow-ments. l1With coefficients on the y's con-strained to one, the natural implementationis to make the dependent variable size-adjusted trade x,-y,-y,, as Anderson and vanWincoop (2003) do. Endogeneity issues alsoarise when the proxies for trade costs are

40 See for example James Anderson and DouglasMarcou~l l e r(2002)41 See for example Exans (2003a),Harngan (19931,an d\\'el (1996)


25/62

714 J o ~ ~ r n a lf Economic Li te ra ture Vol. XLII (September 2004)endogenous. Examples are membership incurrency unions or free trade agreements.Section 3.7.2 below discusses how th e liter-ature has addressed endogeneity in thesespecific cases.3.5 Estimation Bias with TraditionalGrauity Equations

T he traditional gravity equation (1)omitsth e multilateral resistance terms of the theo-retical ly co nsistent model (12 ) . In theabsence of multilateral resistance the twoequations are th e same, with & ,=Am . n thatcase ( Z , ~ ) P " " ( ~ ~ ~ 's an estimate of on e plus th etax equivalent of trade barriers associatedwith this variable. This interpretation of theresults from estimating (1)is generally incor-rect because the z:;' are correlated with themultilateral resistance indices, which arethemselves a function of trade barriers.In o rder to b etter illustrate this point, con-sider th e following simplified environm ent.There are two countries, the United Statesand Canada, with respectively N states andM provinces. T he only trade barrier is a bor-der barrier between states and provinces(ignore distance and o ther factors gen eratingtrade barriers). In that case t,=bsg, whe re b isone plus the tariff equivalent of trade barri-ers associated with the border, and 6, isequal to zero when hvo regions are locatedin the same country and equal to one other-wise. Th e gravity equation is then (droppin gthe constant term)

If we ignore multilateral resistance, theestimate of (l-o)ln(b) is equal to the aver-age within-country size adjusted trademinus the average cross-country size-adjusted trade. When there are N observa-tions of trade within the United States(between states) and M observations oftrade within Canada (between provinces), i tis easy to check that the est imate of(l-o)ln(b) when ignoring the multilateral

resistance terms is equal to (l -o )ln (b) plusthe biasM - N(1- a)- (InP,, - InPC,)N + hl (16)

wh ere P,, an d PC, are the m ultilateral resist-ance indices for, respectively, U.S. states a ndCanadian provinces. P,,>P,, becauseCanada is smaller and provinces face borderbarriers with trad e to all of th e U nited States.The result is that for a> l the estimate of(l-a)ln(b) is biased upwards as long asM >N . This result is intuitive. If, for example,the only within-country trade in th e sample isbetw een provinces, th e average size-adjustedwithin-coun try trade is very large. Th e reasonis that relative trade barriers within Canadaare lower than within the United States du eto larger m ultilateral resistance, so that size-adjusted trade is larger between provincesthan between states. This point way empha-sized by Anderson and van Wincoop (20 03).The example above emphasizes size, butestimation bias holds moreConsider the role of distance. When t,=d{bgeneral11the gravity equation becomes (4 ) with theterm (l-o)pln(d,) added . Th ere are twotypes of bias when attempting to estimatethe border barrier b in a traditional gravityequation that omits multilateral resistance.First, the d istance elasticity (l -o )p is gener-ally incorrectly estim ated since bilateral dis-tance is correlated with the multilateraldistance terms that are left in the e rror term .Second, even when (l- o) p is estimated cor-rectly, we still obtain t he same bias as in (1 6).The bias results as long as the multilateralresistance terms P C , an d P,, are differe nt.Differences arise du e to size, but also du e togeography. For example, Canadianprovinces are located on the NorthAmerican periphery. As a result, their dis-tances from main trading partners ten d to berelatively large, so tha t P,,>P,,. AMcCallum type gravity equation with N=Owould then imply a positive U.S.-Canadaborder barrier even when none existed. Ifthe geographic size of Canada were much


26/62

715nderson and van Wincoop: Trade Costssmaller, so that trading distances betweenprovinces are much smaller, PC, would besmaller and the bias from estimating theborder barrier b with McCallum's equation(N=O) would be smaller. Serge Coulombe(2002) emphasizes the role of these topolog-ical issues related to the special structure ofthe regions.423.6 The Elasticity of Substitution

Estimates of trade costs from trade flowsare very sensitive to assumptions about theelasticity of substitution a, so a look at theevidence is worthwhile. Although manypapers have estimated this elasticity frombilateral trade flow models, only a few haveused theory-based gravity.One way to obtain an estimate of a is touse information from directly observed tradebarriers. Harrigan (1993), Hummels (2001a),Head and Ries (2001), and Bai

Documents

Anderson and Wincoop (2004) Trade Costs