Border Prices and Retail Prices

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 972 May 2009 Border Prices and Retail Prices David Berger, Jon Faust, John H. Rogers, and Kai Steverson NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. References in publications to International Finance Discussion Papers (other than an acknowledgment that the writer has had access to unpublished material) should be cleared with the author or authors. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/.

Border Prices and Retail Prices (cid:3) David Berger (Yale), Jon Faust (John Hopkins), John H. Rogers (Federal Reserve Board), Kai Steverson (Federal Reserve Board) April 9, 2009 Abstract We analyze retail prices and at-the-dock (import) prices of speci(cid:133)c items in the Bureau of Labor Statistics(cid:146)(BLS)CPIandIPPdatabases,usingbothdatabasessimultaneouslytoidentifyitemsthatare identical in description at the dock and when sold at retail. This identi(cid:133)cation allows us to measure the distributionwedgeassociatedwithbringingtradedgoodsfromthepointofentryintotheUnitedStatesto theirretailoutlet. We(cid:133)ndthatoverallU.S.distributionwedgesare50-70%,around10to20percentage points higher than that reported in the literature. We discuss the implications of this for measuring the size of the "pure" tradeables sector, exchange rate pass-through, and real exchange rate determination. We(cid:133)ndthatdistributionwedgesareverystableovertimebutthereisconsiderablevariationacrossitems. Thereissomevariationacrossthecountryoforiginfortheimporteditem,forourmajortradingpartners, butnotasmuchasthecross-itemvariation. Wealsoinvestigatethedeterminantsofdistributionwedges, (cid:133)nding that wedges do not vary systematically with exchange rates, but are related to other features of the micro data. Keywords: prices, distribution, exchange rates. JEL Classi(cid:133)cation: F30 We thank Ariel Burstein and Linda Goldberg for comments and Craig Brown, Rob McClelland, Daryl Slusher, and Rozi (cid:3) Ulics, for their generous assistance. The views in this paper are solely the responsibility of the authors and should not be interpreted as re(cid:135)ecting the views of the Board of Governors of the Federal Reserve System or of any other person associated with the FederalReserve System.

1 Introduction Weanalyzeretailpricesandat-the-dock(import)pricesofspeci(cid:133)citemsintheBureauofLaborStatistics(cid:146)(BLS)CPIandIPPdatabases. Previousworkhasexploitedthesedataseparately, usingeithertheCPI (Bils-Klenow (2004) and Nakamura-Steinsson (2007)) or the IPP (Gopinath-Rigobon, 2007). We use both databases simultaneously in order to compare prices of items that are identical in description at the dock andwhensoldatretail. OurprimarystatisticofinterestistheCPIpricerelativetoimportprice, astatistic that has confusingly been referred to as both the distribution cost or distribution (or pro(cid:133)t) margin in the previousliterature. Instead,weprefertocallthisgapthedistributionwedgebecauseitcaptureseverything thatencompassesthegapbetweentheretailpriceandthepriceatthedockincludingbothpro(cid:133)tmarginsand local distribution costs.1 We think this term is conceptually appealing because while it is clear that at least one these components is necessary to explain real exchange rate dynamics, it is an open question whether the failure of the law of one price for traded goods is primarily driven by variation in pro(cid:133)t margins (Engel 1999) or driven by locals costs of distribution (Burstein, Neves and Rebelo (2005)).2 Unfortunately, we are unable to decompose the distribution wedge into its local cost component and its markup component in a nice, nonparametric way. Nonetheless, the total wedge that we measure is 10-20% larger than previous estimates of the distribution wedge and this implies signi(cid:133)cantly less exchange rate pass-through to retail prices than previous estimates in the literature. After documenting the size of distribution wedges along several cuts of the data, we investigate the determinants of these wedges. It is well known that distribution costs are large. In their authoritative survey, Anderson and van Wincoop(2004)emphasizetheimportanceofdistributioncostsasacrucialcomponentofoveralltradecosts. They note, (cid:147)Trade costs, broadly de(cid:133)ned, include all costs incurred in getting a good to a (cid:133)nal user other than the marginal cost of producing the good itself: transportation costs, policy barriers, information costs, contract enforcement costs, costs associated with the use of di⁄erent currencies, legal and regulatory costs, and local distribution costs (wholesale and retail).(cid:148)They further estimate the contribution of distribution to overall trade costs: (cid:147)The 170 percent headline number for overall trade costs [on an ad valorem tax equivalent basis] breaks down into 55 percent local distribution costs and 74 percent international trade costs [1:7=(1:55 1:74) 1]:(cid:148)Thus, according to the evidence in Anderson and van Wincoop, distribution (cid:3) (cid:0) costs for the United States are large and economically important. 1Spec(cid:133)cally, we follow the previous literature (Burstein, Neves, and Rebelo 2003) and refer to the distrution wedge as PCPI PDOCK P(cid:0)CPI 2The issue isfurtherconfused by the factthatEngel(cid:146)spaperfocuseson the realexchange rate ofthe U.S. and otherlarge, developedeconomieswhereasBurstein,NevesandRebelo(BER)focusontherealexchangeratebetweentheU.S.andemerging economies. There is reason to believe that this di⁄erence underlies some their diverging results. For instance, BER (cid:133)nd that pass-throughintodockpricesforArgentinaisalmost100%whereasGopinath,ItshokiandRigobon(cid:133)ndthatthecorresponding (cid:133)gure forthe U.S.is closerto 10%. 1

Onthemacroside,severalrecentpapershaveexplicitlyanalyzedthee⁄ectsofmodelingadistribution sector. Burstein,Neves,andRebelo(2003),GoldbergandCampa(2006),andChoudri,Faruqee,andHakura (2005)emphasizetheimportanceofadistributionsectorinaccountingforexchangeratepass-through. Each of these papers shows that incorporating a distribution sector into an otherwise standard model improves the ability of the model to explain observed rates of exchange rate pass-through. In models without a distribution sector, predicted rates of pass-through are counterfactually high. More generally, several authors argue that the distribution sector can be crucial for understanding andgeneratingmodelsthatdisplayrealisticrealexchangeratedynamics. Burstein,EichenbaumandRebelo (2005) present evidence on the importance of properly accounting for the role of distribution services as a component of the prices of goods traditionally classi(cid:133)ed as "traded". They undertake an Engel (1999)-style decomposition of real exchange rate movements and document that results from this type of accounting exercise are very di⁄erent when distribution services are included, at least for several large devaluation episodes. Devereux, Engel, and Tille (2003) incorporate a distribution sector in their work on the welfare e⁄ects of moving to a single currency in the euro area. In a series of papers, Corsetti, Dedola, and Leduc have worked extensively on modeling the distribution sector [Corsetti and Dedola (2002), Corsetti, Dedola, andLeduc(2008a,2008b)]. Theyrevisitseveralclassicquestionsininternationalmacroeconomics,including exchangeratepass-through,thelackofcorrelationbetweentherealexchangerateandrelative(home-foreign) consumption and the international transmission of real and monetary shocks.3 There is also a large literature in international trade and (cid:133)nance that argues that variable markups are essential for explaining real exchange rate dynamics. Recent papers that contribute to the literature are Atkeson and Burstein (2008), Goldberg and Hellerstein (2008) and Nakamura (2008). The latter two papers examine speci(cid:133)c industries to understand the sources of incomplete pass-through to retail prices. Consistent with the theoretical work of Atkeson and Burstein, Goldberg and Hellerstein (cid:133)nd that 32%4 of the imperfect pass-through is a result of variable markups at the wholesale level. They also (cid:133)nd that retail markup variation is much less important and does not seem to vary systematically with exchange rates. This is consistent with a recent paper by Gopinath, Gourinchas and Hsieh (2008) which (cid:133)nds that border price di⁄erences are driven by di⁄erences in marginal costs not by variable markups. Despite the importance of variable markups in explaining incomplete pass-through, both industry studies cited above 3In Devereux, Engel and Tille (2003) the modeling was quite simple and designed solely to have two di⁄erent prices, one fordomesticconsumersandoneforexports. In theirset-up,retailersdid notuseanyresourcessuch aslabor. IntheCorsetti, DedolaandLeducframeworks,retailersuselabor,sodomesticfactorcostsmatterfortheconsumerpriceofimportedgoods. In additiontheabsenceofsubstitutabilitybetweenlaborinretailandintheimportedgoodmakestheretailpriceoftheimported good a linear bundle of the imported good and local labor. As a result the foreign exporter faces a non-constant elasticity of demand,leading to severalinteresting (cid:133)ndings,such as limited exchange rate pass-through even with (cid:135)exible prices. 4Intheirexamplecompletepass-throughwouldbe100%passthrough. Localcostsatthewholesalelevellowerpass-through to 50%. Variable markupsatthe wholesale levellowerthe pass-trhough percentto 18%. (ora 32% markup varation)close to what is observed in the data. 2

(cid:133)nd that local distribution costs explain the majority of incomplete pass-through. Our distribution wedge measurecapturesthelevelofthesemarkups,notthechangesandwhilewecannotaccuratelymeasureeither the level or the change in markups, we show in the next section that under plausible assumptions (about the relative magnitudes of the average markup and local costs and about the amount of markup adjustment at the wholesale level in response to an exchange shock), our estimates of the distribution wedge imply signi(cid:133)cantlylesspass-throughintoretailpricesinresponsetoalargedevaluationofthedollarthanprevious studies have found. In short, the overall size of the aggregate distribution wedge is still important. There are a handful of estimates of the size of the distribution wedge in the literature. Almost all of theseestimatesaretakenfromnationalinput-outputtables, andhenceareperformedatafairlyhighdegree of aggregation. Burstein, Neves, Rebelo (2003) estimate that distribution wedges for tradable consumption goods are quite large, on average around 40 percent of the retail price of these goods for the United States and 60 percent for Argentina. Their primary source of data is national input-output tables. Unlike us, they attribute 100% of the wedge to distribution costs which is why they refer to it as distribution costs rather than the distribution wedge. This result follows naturally from the fact that they assume in their theoretical model that the distribution sector is perfectly competitive so these (cid:133)rms earn zero pro(cid:133)ts in equilibrium. Their data work implicitly makes the same assumption because their primary source of data is national input-output tables and these tables are derived under the assumption that all production units have constant returns to scale technologies.5 GoldbergandCampa(2006)documentthesizeofthedistributionsectorfortheUnitesStatesand20 other OECD countries. Their primary data source is also input-output tables so they are assuming that the entire wedge is due to distribution costs. Across countries, distribution wedges on household consumption goodsarebetween30and50percentofpurchasersprices; theestimatefortheUnitedStatesis43%. Forthe eight countries for which Goldberg and Campa have time series data, it is found that distribution wedges are sensitive to exchange rate movements. Bradford and Lawrence (2003) also use input-output sources to measuredistributioncostsinover100consumercategoriesfortheUnitedStatesandeightotherindustrialized countries. For the United States, Bradford and Lawrence report wedges as a fraction of producer prices of 68% on average, or 40% as a fraction of purchaser prices. There is considerable variation across categories of items and across countries, with Japan and the United States on the high end. In this paper, we measure the distribution wedge associated with bringing traded goods from the point of entry into the United States to their retail outlet. An important distinguishing feature of our work isthatwecomputewedgesusingmicrodata. Asnotedabove,wecomparepricesofspeci(cid:133)citemsatthedock andwhensoldatretail. A"matchingprocedure",describedindetailintheappendix,veri(cid:133)esthattheitems 5Hence,in the absence ofotherdistortions,these production units earn zero pro(cid:133)ts in equilibrium. 3

beingcomparedareidenticalindescription. Toourknowledge,nootherstudyofdistributionwedgesusesas detailed a data set as ours. This allows for a cleaner calculation of the distribution wedge than was possible beforeanditallowsthefurtheradvantagetoinvestigatethedeterminantsofthewedgebycomparingitwith other observed covariates. After computing estimates of distribution wedges, we make rough attempts to uncover the determinants of these wedges. Of particular interest, given the focus on this question in the existing literature, is whether wedges vary systematically with exchange rates. Potentially, this can help explain the relative importance of the cost versus margin components. We also relate margins to various features of the micro data, such as the frequency of price changes for an item. This is something papers using input-output data are of course unable to do. We (cid:133)nd that overall distribution wedges are around 50-70% for U.S. data between January 1994- July 2007. This number is about 10 to 20 percentage points higher than that reported by other researchers. Distribution wedges are quite stable over time but vary considerably across items. Margins are typically lower for sale price CPI items, as expected, but do not di⁄er signi(cid:133)cantly across c.i.f. versus f.o.b. import price basis considerations. Surprisingly, intra-company transfer pricing considerations did not have a sizable e⁄ectonthesizeofdistributionwedges. Thereissomevariationacrossthecountryoforiginfortheimported item,forourmajortradingpartners,butnotasmuchasthecross-itemvariation. Wedonot(cid:133)ndthatwedges vary systematically with exchange rates although wedges are well explained by other characteristics of the micro data. We take this lack of correlation with the exchange rate as evidence that the majority of our distribution wedge is capturing distribution costs, not pro(cid:133)t margins. If distribution wedges were largely composedofvariablewholesalemarkupsandifpricingtomarketisimportant,thenchangesinwedgeswould strong negatively covary with nominal exchange rate changes. 2 Distribution Margin or Distribution Cost Asmentionedintheintroduction,wemeasuretheaggregatewedgebetweentheretailpriceandtheprice atthedock-awedgethatincludesbothretailanddistributormarkupsandlocaldistributionandmarketing costs. Unfortunately, despite our intensive work with these rich data sets, we are unable to disentangle thesetwocomponentsinanice, nonparametricway. Onewaytoproceedistofollowthepreviousliterature (Burstein, Neves and Rebelo 2003) and assume that the distribution wedge is equal to the distribution cost. Given that we measure the distribution wedge to be between 10-20% larger than previous estimates, if we performed a similar exercise to the one done by Burstein, Eichabaum and Rebelo (2005), then one would (cid:133)nd that our measured wedge implies signi(cid:133)cantly less pass-through into retail prices than what Burstein, 4

Eichabaum and Rebelo found. This is shown explicitly in the next section of the paper. Wethink, however, thattheassumptionthatthedistributionwedgeisequaltothedistributioncost is unappealing because it contradicts a long empirical IO literature arguing that many (cid:133)rms are imperfectly competitive. Furthermore, if there is no markup, there is no role for markup adjustment, contradicting recent empirical work by Goldberg and Hellerstein (2008) and Nakamura (2008), both of whom (cid:133)nd that markup adjustment at the wholesale level is important for explaining incomplete pass-through. Another approach is to make a rough approximation of the relative magnitudes of the markup components and the distribution costs components so that we can consider both margins in the exercise we perform in the next section. To (cid:133)x ideas consider a simple decomposition of the retail price for a single good: PR =PNT +(cid:22) +PD t t t t where PR, PNT, PD and (cid:22) are the retail price, the distribution cost, the price at the dock and the markup. Concretely, one can imagine the case where the foreign manufacturer owns the wholesaler in the U.S. as is the case in the Beer industry. (Goldberg and Hellerstein 2008) PD is the foreign manufacturer(cid:146)s unit marginal cost, (cid:22) is the markup the wholesaler charges the retail (cid:133)rm to purchase the item, and PNT is the total distribution costs required to bring the product to market: 6 The distribution wedge is de(cid:133)ned as PR PD d =100 t (cid:0) t 60% t (cid:3) PR (cid:25) (cid:18) t (cid:19) Consistent with the upper estimates from the empirical IO literature, assume that unit margins are equal to 25% (cid:22) =:20 :7 Thisimpliesthatthefractionoftheretailpricespentonlocaldistributionandmarketing PR costs(cid:0)is 35%. C(cid:1)onsistent with the empirical literature highlighting the importance of pricing to market we also assume that the markup varies negatively with the exchange rate. Speci(cid:133)cally, in response to a 1% unexpected depreciation of the dollar, we assume that wholesale markup falls by .317%.8 Interestingly no matter which set of assumptions one makes the implied pass-through into retail prices is much less than previous studies found. 6Alternatively, one could consider the case where a wholesaler purchases the item from an overseas manufacturer at price PD, it requires PNT total distirbution costs to bring the item to retail and (cid:22) is the total markup of the wholesaler and the retailer. 7Note that we are measuring the markup relative to the retail price, where traditionally the markup is measured relative to marginal cost. Thus our assumption that the markup relative to the retail price is a (signi(cid:133)cant) lower bound the the conventionally measured markup. 8See Goldberg and Hellerstein (2008) Table 13 column 2. 5

3 Distribution Wedges, Measuring the Tradeables Sector, and Exchange Rate Pass-Through One concern about the large US external de(cid:133)cits is that they will eventually boost in(cid:135)ation. Typically the purported e⁄ects, some of which are quite dire (Obstfeld and Rogo⁄, 2005), work through the exchange rate. Such concerns are of course tempered when purely traded goods are a small fraction of our economy. Our estimates of the size of distribution wedges have a direct bearing on this issue. To illustrate, let(cid:146)s map distribution wedges into a calculation of the size of the pure traded goods sector, following the extension of Engel(cid:146)s (1999) simple example by Burstein, Eichenbaum and Rebelo (2005). Think of the CPI as a geometric average between the retail price of traded goods PT and the prices t of nontradable goods and services PN: t PCPI =(PT)1 !(PN)! (1) t t (cid:0) t where ! is the weight of nontradables in the CPI not taking into account distribution wedges. The retail price of traded goods includes both goods that are actually traded, whose retail price is PI , and traded t goodsthatareconsumedlocally,whoseretailpriceisdenotedPL. Assumethattheretailpriceoftradables t can be written in Cobb-Douglas form as: PT =(PI)1 (cid:18)(PL)(cid:18) (2) t t (cid:0) t where (cid:18) is the share of local goods in tradable goods. Assume that selling at retail one unit of traded or local goods requires nontradable distribution services and that the price of distribution services is the same as the price of nontradable goods. For simplicity, assume the technology to transform traded and local goods into retail tradable goods is Cobb- Douglas. Denote the weight of distribution services by (cid:30) and for simplicity assume that it is the same for both traded and local goods. We also assume that the markup the retailer charges is zero. What matters forincompletepass-through isvariationin the markup and theempiricalevidence suggests thatvariation in the retail markup explains less than 1% of incomplete pass-through. (Goldberg and Hellerstein 2008). Consistent with the discussion in the previous section, we assume that the distributor of the traded good charges a markup (cid:22) over the dock price. For simplicity, we assume that there is perfect competition in the distribution sector of the local good. These assumptions imply that the retail price of traded and local 6

goods is given by PI =((cid:22) P I )1 (cid:30)(PN)(cid:30) (3) t t t (cid:0) t and PL =(P L )1 (cid:30)(PN)(cid:30) (4) t t (cid:0) t I L where P is the price at the dock of the traded good and P is the price of the local good exclusive of t t distribution costs. Assuming that the price of local goods exclusive of distribution costs is the same as the price of nontradable goods: P L =PN (5) t t Equations 1-5 imply that PCPI =(cid:22)1 (cid:11)(P I )1 (cid:11)(PN)(cid:11) (6) t (cid:0) t (cid:0) t where (cid:11)=1 (1 (cid:18))(1 (cid:30))(1 !)isthe totalweightof nontradablesinthe CPIbasket. AssumethatPPP (cid:0) (cid:0) (cid:0) (cid:0) holds for the purely traded component of the CPI: I I P t =kS t P t (cid:3) (7) I where S t is the nominal exchange rate, P t (cid:3) is the foreign price of truly traded goods and k is a constant of proportionality. Plugging equation 7 into 6 and taking log di⁄erences gives: (cid:25)CPI =(1 (cid:11))ln((cid:22) =(cid:22) )+(1 a)ln(S =S )+(1 (cid:11))(cid:25)I +(cid:11)(cid:25)N (8) t (cid:0) t t (cid:0) 1 (cid:0) t t (cid:0) 1 (cid:0) t(cid:3) t Consistent with Goldberg and Hellerstein (2008), Nakamura (2008) and Atkeson and Burstein (2008) we assume that the distributor(cid:146)s markup varies with the exchange rate. This allows us to write changes in the markup in terms of changes in the exchange rate. ln((cid:22) =(cid:22) )= (cid:23)ln(S =S ) (9) t t (cid:0) 1 (cid:0) t t (cid:0) 1 where (cid:23) (0;1) . We choose v equal to 0.317 (see table 13 of Goldberg and Hellerstein 2008) which 2 implies that for every 1% depreciation of the dollar, distributor(cid:146)s markups fall by 0.317%. This is because an exchange rate depreciation leads foreign (cid:133)rms to change the markup or pro(cid:133)t margin they earn per unit 7

sold in the U.S. Using equations 7 and 8 we can now express CPI in(cid:135)ation in terms of the exchange rate (cid:25)CPI =(1 a)(1 (cid:23))ln(S =S )+(1 (cid:11))(cid:25)I +(cid:11)(cid:25)N (10) t (cid:0) (cid:0) t t (cid:0) 1 (cid:0) t(cid:3) t It is easy to see that the smaller the weight that pure tradables have in the CPI, the lower the expected change in the CPI will be in response to a devaluation. Also, the more the distributor(cid:146)s markup responds to changes in the nominal exchange rate the lower pass-through will be. There are two mechanisms that are often mentioned in discussions of how U.S. CPI in(cid:135)ation might be "imported" from abroad. The most common begins with the concern that large U.S. trade de(cid:133)cits will ultimately cause the dollar to depreciate sharply against other (cid:135)oating currencies. More recent discussions explainhowCPIin(cid:135)ationmightbeimportedevenwhentheforeigncountryhasa(cid:133)xedormanagedexchange rate. Note that both mechanisms are captured by equation 8: the former refers to changes in the (cid:133)rst term, whilethelatterreferstochangesinthesecondterm. Intheabsenceofsystematicmarkupvariation((cid:23) =0), given that these mechanisms enter the in(cid:135)ation equation in a symmetric way, they imply the same results for depreciations holding foreign costs constant and for (equivalent-size) foreign cost increases holding the nominalexchangerateconstant. Whenmarkupsvarysystematicallywiththeexchangeratethenonewould expect more pass-through in the latter case. To understand the importance of this paper(cid:146)s measuring the size of U.S. distribution wedges, let(cid:146)s use equation 9 to do some back of the envelope calculations. We consider two calibrations. For the (cid:133)rst calibration we follow BER and assume that the distribution wedge is equal to the distribution cost and that thereisnomarkupvariation. Sincewe(cid:133)ndthattheaveragedistributionwedgeisequalto60%,thisimplies that (cid:30) = 0:6 and (cid:23) = 0: In the second calibration we follow the empirical IO literature and assume that (cid:30)=0:35 and (cid:23) =0:317: BER (2003) estimate the average distribution cost in the US to be 45%. In 2006, the share of services (nontradables) in the US CPI-U was 60% (! = 0:6). They estimate that the share of traded goods consumed locally could be as large as 22% of traded goods ((cid:18) = 0:22). This suggests that a lower bound on the amount of pure tradable goods with BER(cid:146)s estimates is :4 :55 :22 :4=13:2% (cid:3) (cid:0) (cid:3) and an upper bound is :4 :55=22% (cid:3) Under our (cid:133)rst calibration, since the distribution cost is 60% the lower bound on the purely tradable com- 8

ponent of the CPI drops to :4 :4 :22 :4=7:2% (cid:3) (cid:0) (cid:3) with an upper bound is :4 :4=16% (cid:3) With such a small tradable component to the CPI that would be guided by PPP, there is less reason to be concerned that U.S. in(cid:135)ation would be highly a⁄ected by even a large and sharp depreciation. Using this lower bound, we can get a back-of-the-envelope guess at what would happen to U.S. in(cid:135)ation if there was a big depreciation but foreign prices remained constant. (Remember that this is equivalent in our framework to considering a large increase in foreign prices holding the exchange constant when there is no markup adjustment). Assume that nontradable good prices in the U.S. grow at 2% a year. Assume also that we have a 25% unexpected depreciation in the nominal trade weighted exchange rate, that foreign prices do not change over this period and that pure traded goods obey PPP. Under the (cid:133)rst calibration this implies that the share of nontradables in the CPI is 87:5% 9Then the expected in(cid:135)ation rate using equation 9 gives Expected (cid:25)CPI =(:125 :25)+(:875 :02)=4:88% (cid:3) (cid:3) Underthesecondcalibration,theshareofnontradablesintheCPIisequalto79.7%andv =0:317:Expected in(cid:135)ation under this scenario is given by Expected (cid:25)CPI =(:203 (1 :317) :25)+(:797 :02)=5:06% (cid:3) (cid:0) (cid:3) (cid:3) Using the BER calibration with an estimated 45% distribution margin produces an expected in(cid:135)ation rate of 6%. Therefore our (higher) estimate of distribution wedges lowers the expected pass-through to domestic in(cid:135)ation by 19% relative to BER in the (cid:133)rst calibration and by 16% relative to BER under the second calibration. This a non-trivial di⁄erence for an economy like the United States which has a low and stable in(cid:135)ation rate. However, the main lesson of both BER and our paper is that when distribution costs and markup adjusted are included, the expected change in the CPI in response to even a very large unexpected depreciation is not that alarming. Excluding distribution costs of course implies much higher expected in(cid:135)ation - our simple example produces in(cid:135)ation rates of around 9% under the (cid:133)rst calibrationeven greater when local goods are excluded. Of course, with larger unexpected depreciation rates the in(cid:135)ationary implications are larger, e.g., with a 45% depreciation, the expected CPI in(cid:135)ation rate is 7.4% 9(cid:11)=[1 (1 :6) (1 :6) (1 :22)]=:875where the distribution wedge is 60%,the share of services in the CPI is 60% (cid:0) (cid:0) (cid:3) (cid:0) (cid:3) (cid:0) and localgoods make up 22% oftraded goods. This represents an upperbound forthe share ofnontradables. 9

when distribution costs are taken into account. Qualitatively,with even higher distribution wedges than had previously been reported in the literature, the in(cid:135)ationary consequences for the United States of a large depreciation are even less disastrous. Of course these are simple back of the envelope calculations and do not necessarily emerge from a more desirable general equilibrium analysis. 4 Measuring Distribution Wedges We measure distribution wedges in two ways. First, using the detailed information on product characteristics in the CPI and IPP databases, we match items at the dock to those sold at retail that are identical in description. Our matching procedure is done on a category by category basis depending on available information, as described in detail in the appendix. Under this procedure we are highly con(cid:133)dent that we are comparing at-the-dock prices and retail prices of items that are identical in description. Unfortunately, this procedure also necessitates that we discard a lot of data, either because exact matches did not exist or because there was insu¢ cient evidence to determine the quality of a match. Inlightofthislastconsideration,wecheckrobustnessusingasecondmeasureofdistributionwedges. Under this procedure we construct weighted-average price levels for fairly disaggregated item categories in the CPI and import price data bases. The level of aggregation is by entry level item (ELI) in the CPI, or approximately 10-digit SIC code for imports. We use prices of only those CPI items that we could reliablydeterminetohavebeenimportedratherthanmadeintheUnitedStates. Thisalternativeprocedure allowsustomeasurethedistributionwedgeforitemcategoriessuchas(imported)(cid:147)beer(cid:148),(cid:147)televisions(cid:148),and (cid:147)bananas(cid:148). Under this procedure we utilize the prices of many more of the items in the sample but use less of the item-speci(cid:133)c information that is contained in the database. Under both strategies, the distribution wedge for item category i, d , is calculated as d = (CPI i i i IPP ) / CPI i i where CPI is the retail price of the item (or its weighted average price level) and IPP is the import i i price.10 We use monthly data from January 1994 to July 2007. Thecalculationofd couldbea⁄ectedbyseveralimportant(cid:147)pricebasis(cid:148)considerations.11 The(cid:133)rst i is whether the CPI item(cid:146)s price is a sale price or a regular price. Second, is whether the imported item is priced on a c.i.f. or f.o.b. basis. Finally, we must distinguish between imports that are intra-company 10Variousauthorsreportwedgesindi⁄erentways. Withtheformulaabovethewedgeisboundedbyzero(whenCPIpriceis greaterthan IPP)and unity. Ithastheintuitiveinterpretation asthefraction oftheretailprice,which consumersdo observe, that is accounted forby transportation costs,overhead,retailerpro(cid:133)t,etc. 11This is relevant for the matching procedure but not the alternative procedure where we calculate weighted-average price levelsforELIcategories. Underthelatterwedonotutilizesuchinformationaswearetryingtousepricesofasmanyitemsas possible,irrespective ofwhetherthe database contains speci(cid:133)c information on the item. 10

transfers and those that are arm(cid:146)s length transactions that more accurately re(cid:135)ect market prices. Each of these could have non-trivial e⁄ects on the distribution wedge. In light of this, we report results in a few di⁄erent ways re(cid:135)ecting combinations of these price bases considerations. In Table 1 we report the median distribution wedge for all items under the (cid:133)rst of our measurement procedures. Results using the (cid:147)matching procedure(cid:148)described in the appendix are contained in part A of the table, while those of the alternative procedure using weighted average price levels are in part B. For the former we report wedges in four ways: when the CPI price is regular and the import price basis is cif, CPI price is regular and import price is fob, and the analogies for cases in which the CPI price is a sale price. In the upper panels intra-company transfer prices are excluded. In the lower panels we report the same calculations using only the intra-company transfer prices. 4.1 Distribution Wedges: all items AccordingtotheupperpanelofTable1A,whentransferpricesareexcludedfromthesamplethemedian distribution wedge across all regular-priced CPI items is 0.57 (0.68) for imports priced on a cif (fob) basis. For sale-price CPI items the respective distribution wedges are 0.50 (0.60). The analogous numbers for transfer prices are, contrary to our prior expectations, generally quite similar: 0.58 (0.62) and 0.57 (0.49), as seen in the lower panel of Table 1A. The distribution wedges reported in Table 1 are distinctly higher than the estimates reported for U.S.consumptiongoodsbyotherresearchers. Burstein,Neves,andRebelo(2003)estimateU.S.distribution wedges12 to be 42% in 1992 and 43% in 1997, using the national input-output tables. The wedge is about the same when the authors use data from the 1992 U.S. Census of Wholesale and Retail Trade. Goldberg andCampa(cid:146)s(2006)cross-countryevidencecon(cid:133)rmsthe43%estimateofthedistributionwedgesforallU.S. (cid:133)nal household consumption in 1997 (also using national input-output data), estimating that most of this is duetodistributionwedgesinthewholesale-retailsectorratherthantransportation. BradfordandLawrence (2003)reportanoveralldistributionwedgefortheUnitedStatesin1992of40%asapercentageofpurchaser price. 4.1.1 Alternative procedure Table 1B reports distribution wedges computed under the alternative procedure where we construct weighted-average price levels for fairly disaggregated item categories. These wedges are slightly higher than those obtained from the matching procedure: 0.70 or 0.64 depending on how we weight item categories. 12Remember,they make assumptions so that the distribution margin is equalto the distribution cost. 11

This indicates a general robustness to using prices of considerably more items than was possible under the matching procedure. 4.1.2 Stability over time The wedges are quite stable over time. Lumping all items together without distinguishing between cif and fob, sale price or not, etc., our matching procedure gives us wedges of 0.62, 0.67, 0.63, 0.57, 0.59, 0.60, 0.58, 0.61, 0.59, 0.57, 0.58, 0.60, 0.60 and 0.61 in the years 1994 through 2007 respectively. As noted above, a relatively stable overall distribution wedge is also found by Burstein, Neves and Rebelo (2003). This stability of wedges against the backdrop of considerable (cid:135)uctuations in the dollar foreshadows our (cid:133)nding below concerning the lack of a relationship between distribution wedges and exchange rates. 4.2 Results by item In our data sample, there is considerable variation across items, with wedges ranging from around 20 percent to 80 percent. These results are presented in Table 2A for the 21 item categories from which we were able to uncover a su¢ cient number of high-grade matches (see the appendix tables for the number of observations in each category). The lowest wedges are for televisions, video cameras, VCRs, cameras, telephones and microwave ovens. The highest wedges are found for drugs, our two apparel categories (men(cid:146)s andwomen(cid:146)spants),watches,(cid:133)lm,andourtwofreshfoodscategories(bananasandtomatoes). Asexpected, wedgesaretypicallylowerforsalepriceitems,andinsomecasesthedi⁄erenceisnearly20percentagepoints. Wedges do not di⁄er systematically between the cif and fob price basis, though on average fob wedges are higher as expected. Table2Breportsresultsbyitemcategorywhenwecomputedistributionwedgesusingthealternative procedure.13 Consistent with the results under the matching procedure, the largest wedges are observed for watches, olive oil, and bananas, with wedges for television sets (and alcoholic beverages here) being at the low end. Below we relate the cross-section of distribution wedges to features of the micro data underlying our sample. 4.3 Composition e⁄ects and results by brand The results so far could be masking important composition e⁄ects, in principle across brand, time, and country of origin. The item categories above, while certainly disaggregated to some extent, still contain 13Wewereabletoconstructreliableestimatesofweighted-averagepricelevelsforonlyabouthalfoftheitem categoriesused in the matching procedure. This was due to data limitations. Particularly constraining was getting information on whether a particularCPIitem was produced in the United States orabroad. 12

product heterogeneity. There are, for example, the wedges associated with small-screen television sets (13inch diameter) and those associated with large, high-end televisions. These wedges are averaged together in theresultsabove. Ifthereareimportantcompositione⁄ects, itmaybemisleadingtocompareourestimates to those of the existing literature, or to compare results across various slices of our own data set. Concerning the comparison with existing estimates of distribution wedges, note that in his conference discussion of our paper (FRB-NY, December 2007), Ariel Burstein presented results indicating that composition e⁄ects do not explain the higher wedges we report relative to BER (2003). That is, Burstein showed that the distribution wedges in the NIPA data used by BER are consistently lower than what we report from the BLS data, category by category. Furthermore, we compute distribution wedges by brand for cases in which we have at least ten observations. Table 3a reports results for Alcoholic Beverages for the case in which transfer prices are excluded. Table 3b repeats the exercise for beer and television sets. (Con(cid:133)dentiality considerations between the BLS and companies preclude us from naming the actual brands.) Composition e⁄ects across brands do not appear to be a predominant factor: distribution wedges for individual brands are in most cases close to the aggregate distribution wedge calculated for all brands together. Television sets are a notable exception (table 3b). Consider the seemingly anomalous result that the wedge associated with fob prices is higher for televisionsetswherethereisasalepriceintheCPIthanforregular-pricedCPItelevisions, 0.35versus0.21. Making the same comparison by brand, we see that in each case the wedge for sale-price televisions is, as expected, lower than for regular-priced items. 4.4 Distribution wedges by country-of-origin We also calculate distribution wedges based on a di⁄erent cut of the data. Here we lump together all items that were imported from a particular trading partner and calculate distribution wedges based on the matching procedure. Table 4 presents the results for our major trading partners. wedges range from a low of 0.36 for Japan (for imports priced on an fob basis) to 0.75 for Mexico (cif basis). However, most of the wedges fall within the range of 50% to 60% reported in the all-items tables above. 5 Determinants of Distribution wedges Weattempttoexplainthedeterminantsofdistributionwedgesbyrelatingtheestimatesdescribedabove to various quanti(cid:133)able features of the micro data underlying our sample. In terms of broad brush, we go about this with two distinct approaches that examine various (cid:147)within(cid:148)and (cid:147)between(cid:148)determinants. First is to explain the determinants of distribution wedges across item categories themselves or across export- 13

ing countries themselves. Under this approach we compute summary numbers (medians) for a particular item category and explain the variation in (median) distribution wedges across item categories. The other approach is to look within an item category and examine the determinants of distribution wedges across items in that category. We repeat this exercise for particular trading partner countries, to examine the determinants of distribution wedges across all items originating from that particular country. 5.1 Accounting for Variation in Distribution wedges As a (cid:133)rst pass at understanding the determinants of distribution wedges, we ask which of the two components of the wedge, the IPP price or the CPI price, accounts for more of the movements in the wedge itself? Or alternatively, we ask are there o⁄setting movements in the IPP and CPI prices such that distribution wedges are unrelated to either? In Table 5 we present the results of a simple decomposition of the variation in distribution wedges. In particular, we report the R2 values from three di⁄erent bivariate regressionsinvolving,respectively,theIPPpriceandtheCPIprice,theCPIpriceandthedistributionwedge, andtheIPPpriceanddistributionwedge. Thevariablesareinlog(cid:133)rst-di⁄erences,andallregressionscontain a constant but no lags. As with the calculations above, we report results for three di⁄erent slices of the data: by item category, by year, and by country of origin of the import. The bulk of the evidence presented in Table 5 indicates that variation in the wedge is primarily accounted for by movements in the retail price, with very little due to movements in the import price. In addition, There is very little evidence of a systematic relationship between movements in the IPP price and the CPI price. The latter is consistent with the literature on exchange rate pass-through into the United States. More speci(cid:133)cally, these conclusions hold by item category according to Table 5a: only for computer accessories and cameras is signi(cid:133)cantly more of the variation in the distribution wedge accounted for by the IPP price than the CPI price. This pattern is relatively stable over time, according to Table 5b, though there is a noticeable decline in the contribution of the CPI price in the latter part of the sample. During this period there is a rise in the contribution of the IPP price to the wedge but it is not large.14 Finally, when we perform this simple decomposition by country of origin we see that movements in the CPI price account for more of the variation in wedges than do movements in the IPP price for each trading partner. In addition, only in the case of imports coming from Canada is there a non-zero relationship between movements in the IPP and CPI prices. 14There are noticeable jumps in the R2 values in the (cid:133)nal year, 2007, which runs only through the month of July. We thus await an updating ofthe data and decomposition results before putting too much stock in these jumps. 14

5.2 Distribution wedges, Endogenous Exits, Law of One Price Deviations and Sticky Prices Next we relate distribution wedges to various features of the BLS micro data. Two of these features are measuresoflawofonepricedeviationsandpricestickiness. Theyarewell-known, andwesimplyfollowthe existing literature in calculating them from the BLS micro data. The third feature is our own construct, whose explanation we turn to next. 5.2.1 Endogenous Exits OnestrikingfeatureofthemicrodataintheIPPdatabaseisthatparticularitemsimportedfromparticular countries are relatively short-lived. We construct a variable that summarizes the short-lived nature of such items, and see if this is systematically related to distribution wedges. We label this new variable the (cid:147)endogenous exits(cid:148)ratio. An endogenous exit is said to occur any time (1) the importing company has gone out of business; (2) the BLS industry analyst, in consultation with the company, concludes that a product is (cid:147)out of scope(cid:148), indicating that there is no longer a meaningful market for the product; or (3) highly signi(cid:133)cant changes in quality are made to an existing item. We then count, within each of the item categoriesusedinthematchingprocedure,thenumberofitemsinthatcategoryexperiencinganendogenous exit during the sample period. The variable of interest, the endogenous exits ratio, is the ratio of this count to the total number of items in that category.15 Theunconditionalrelationshipbetweenendogenousexitsanddistributionwedgesisdisplayedinthe scatterplotintheupperleftpanelofFigure1. Eachdotrepresentsasingleitemcategory,forexamplebeer. The relationship is negative, with a simple correlation coe¢ cient of -0.37 and no clear outlier observations. Thus,inourdata,itemcategorieswithsmalldistributionwedgesarethosewithrelativelymanyendogenous exits. We o⁄er interpretations of this (cid:133)nding below. 5.2.2 Law of One Price Deviations We next examine the relationship between distribution wedges and deviations from the law of one price for the CPI items used in the matching procedure. Conceptually, if the law of one price is closer to holding in the market for a particular item we may expect that market to be more competitive and hence exhibit smallerdistributionwedges. Tobespeci(cid:133)c, wecalculateabsolutedeviationsfromthelawofonepriceacross citiesforaparticulartypeof, e.g., televisionset. Recallthatwehavealreadydeterminedparticularitemsto 15In any given year, about one-(cid:133)fth to one-fourth of the items exit the sample (cid:147)endogenously(cid:148)in this way. This (cid:133)gure has remained fairly constant over time. Somewhat to our surprise, there is not a lot of cross-country variation when we count endogenous exits by country oforigin ofthe imported item. 15

be (cid:147)identical in description(cid:148)through our matching procedure. These are the only items whose (cross-city) price we are comparing in this exercise. For each category like televisions we calculate one number: the median law of one price deviation across all of the individual city-pair observations. We then relate this to the distribution wedge already calculated for that item category. The relationship is depicted in the upper right panel of the (cid:133)gure. There is clearly a positive relationship in our data: categories such as televisions, VCRs, microwave ovens have very small deviations from the law of one price at the retail level; these are also the categories with the smallest distribution wedges. The tri-variate relationship among distribution wedges, endogenous exits, and law of one price deviations at retail provides some economic insights. In item categories where the distribution wedge is small there is a relatively large amount of product churning or turnover, directly observed at the import stage, either because of signi(cid:133)cant quality changes or other market forces that render products obsolete relatively quickly. Thesesmall-wedge(andhighexits)itemcategoriesarealsothoseinwhichmarketforceskeepprices relatively in line with the law of one price at the retail level. 5.2.3 Sticky Prices Finally we ask whether distribution wedges are related to measures of price stickiness. On a priori grounds, we may expect that the sectors with low wedges and in which the law of one price comes close to holding, are also characterized by relatively (cid:135)exible prices. This appears to be the case. In the bottom rowofthe(cid:133)gurewedepictscatterplotsofdistributionwedgesagainsttheprobability thataniteminthat category experienced a price change. We calculate these probabilities for both the CPI price (lower left panel)andtheIPPprice(lowerright)ofthatitem. WefollowNakamuraandSteinsson(2007)incalculating the probabilities (we include both sales prices and regular prices in our CPI calculations). As seen in the (cid:133)gure, the relationship is a⁄ected by a small number of outlier observations. Excluding the outliers, the relationship is strongly negative, -0.42 for CPI and -0.31 for IPP, so that lower wedges are associated with more frequent price changes.16 On the CPI side, the one outlier category is tomatoes, for which prices are quite (cid:135)exible while distribution wedges are high. This presumably re(cid:135)ects a relatively unique combination of (1) supply-side competition, product homogeneity and low demand elasticities inducing frequent price changes and (2) costly transport and storage needs that keep wedges high. On the IPP side, tomatoes are again an outlier, as are bananas and olive oil. This simple, non-structural examination of the determinants of distribution wedges suggests an 16Withtheoutliers,thecorrelationisessentiallyzero. Notethatthisnegativerelationshipisconsistentwiththetheorectical and empiricalwork presented in Gopinath and Itshoki(2008b). 16

interesting relationship among distribution wedges and three "micro features" of the BLS data: endogenous exits, law of one price deviations, and sticky prices. The relationship points to the likely strong role of factors that we would expect to see in(cid:135)uencing distribution wedges (cid:150)competition, product substitutability, transportation and storage costs. 5.3 Do Distribution wedges move Systematically with Exchange Rates? One of the important goals of Goldberg and Campa (2006) was to examine the relationship between distribution wedges and exchange rate changes. As shown in our example of section 2, there is a simple mappingbetweendistributionwedgesandthesizeofthepuretradeablessector. Thus,a(cid:133)ndingthatwedges move with exchange rates suggests that exchange rate changes would, all else constant, be associated with a change in the size of this sector. CampaandGoldberg(2006)reportthathomecurrencydepreciationsareassociatedwithstatistically signi(cid:133)cantly lower distribution wedges in a panel regression containing the United States and 9 European countries. They use national data, at a fairly high level of aggregation, over the period 1995-2001. Their estimates indicate that a 1% real depreciation results in a 0.47 percent decline in the distribution wedge. Such an elasticity of the wedge with respect to exchange rate changes is large enough that it could, in principle, be able to account for much of the discrepancy between our estimates of wedges and those in the existing literature. As noted above, however, there is prima facie evidence against (cid:133)nding a relationship between distribution wedges and exchange rates in our data. Recall that our average annual distribution wedge across all items during the period 1994-2007 (cid:135)uctuates between 0.57 and 0.67, with all of the estimates after 1996 lyingbetween0.57and0.61. Duringthisperiodthedollarmovedbyaconsiderableamount: againstthecurrenciesofourmajortradingpartners, thedollar(cid:133)rstappreciatedbymorethan20percentandsubsequently depreciated by more than 30 percent. Were we to apply the Goldberg-Campa elasticity to the recent U.S. experience, the large drop in the dollar would have been associated with about a 6 percentage point drop in the distribution wedge, all else constant. More formally, in Table 6 we present the results of a regression of the change in distribution wedges on the contemporaneous change in the exchange rate, two lagged changes in the exchange rate, and two lagged changes in the foreign CPI.17 The data are monthly from January 1994-July 2007. We run the regression on two di⁄erent cuts of the data, (cid:133)rst by item (Table 6A) and second by country of origin of the import item(Table 6B). In the formerregression the exchange rate is the trade-weighted value of the dollar, 17Thisistheprototyperegressionfoundintheliteratureonexchangeratepass-through. GoldbergandCampa,(2006)provide details. 17

while in the latter regression the exchange rate is the bilateral rate of the dollar against the currency of the exporting country. We report coe¢ cient estimates on the contemporaneous exchange rate change as well as the Fstatistic from a test of the null hypothesis that the exchange rate changes are jointly zero. We also report the regression R2 values. According to Table 6A there is only modest evidence that exchanges rates signi(cid:133)cantly a⁄ect distribution wedges. In two categories, Drugs and Microwave ovens, the coe¢ cient on the contemporaneous exchange rate change is negative and statistically signi(cid:133)cant. In each case the F-test indicates that the exchange rate changes are jointly non-zero and, in the case of drugs, the regression R2 is sizable. However, for most items the regressions indicate that there is essentially no relationship between exchange rates and distribution wedges. Table 6B repeats the analysis, this time on a country-by-country basis lumping all items together. Here we (cid:133)nd even less evidence that exchange rate changes are signi(cid:133)cant. For Canada, the estimated coe¢ cient on the contemporaneous exchange rate is large at -0.61, but has a standard error of 0.4. The F-statisticfortheCanadianregressiondoesindicatethatexchangeratechangesarejointlynon-zerobutthe regression R2 is quite small. The absence of a signi(cid:133)cant relationship between distribution wedges and exchange rate changes would seem to contradict Goldberg and Campa (2006). However, the data sets used in the two papers are quite di⁄erent, with the Goldberg-Campa data set being richer in the cross-country dimension. Our data for the United States, a relatively closed economy, is richer across item categories. No matter how we sliced our data set, however, there was no consistently signi(cid:133)cant relationship between distribution wedges and exchange rates. 5.4 A Multivariate Empirical Model of Distribution wedges Weconcludethesectionwithanexaminationofthejointsigni(cid:133)canceofeachofthevariablesdiscussedso far in explaining distribution wedges. For each item category, we regress the distribution wedge for product i and time t on several variables: d =(cid:11)+(cid:12) BRAND +(cid:12) SALE +(cid:12) BUSINESS +(cid:12) PBASIS +(cid:12) LOPDEV +(cid:12) EXITS it 1(cid:3) i 2(cid:3) it 3(cid:3) i 4(cid:3) it 5(cid:3) i 6(cid:3) i where BRAND is the brand name of product i, SALE is a zero-one dummy indicating whether the price of the CPI item is a regular price or sale, BUSINESS is also zero-one denoting the type of retail outlet where the item was sold (large retailer or discount store versus small convenience-type of store), PBASIS is the price basis for the imported item (cif versus fob), LOPDEV is the law of one price deviation for the CPI 18

item, and EXITS is the endogenous exit count for that item category. We then repeat this analysis for each country of origin, running the regression above over all items imported from that country. The results are depicted in Table 7A (by item) and 7B (by country of origin). We are primarily interested in the contribution to R2 of each explanatory variable. These are reported in square brackets in the cells. We report the R2 of the full model in column 1. These range from just below 0.2 for some categories up to nearly 0.9 for others. Product BRAND typically has a sizable contribution to overall R2 values, according to column 2. BUSINESS type and SALE (cid:135)ag dummies have the expected signs (positive and negative, respectively) and in some cases contribute signi(cid:133)cantly to the overall (cid:133)t, especially the former variable. Within item category, law of one price deviations are often negatively related to distribution wedges. This does not necessarily contradict the cross-item evidence shown in the scatter plot above. It may be the case, for example, that this within-category relationship re(cid:135)ects the e⁄ect of a factor such as distance between cities in causing both a larger deviation from the law of one price and, through transport costs, smaller distribution wedges. Endogenous EXITs are typically negative (though there are exceptions here too) and often signi(cid:133)cant: within item categories, higher product turnover is associated with smaller wedges just as in the cross-item results in the scatter plot. Finally, Table 7B presents the within-country evidence. Once again, R2 values for the overall model arefairlyhigh,upto0.75foritemsimportedfromMexicoandChina. AsinTable7A,BRANDconsiderations tendtobeveryimportantwhileBUSINESStypeandSALE(cid:135)agsagainhavetheexpectedsigns. Endogenous EXITs are typically unimportant according to the contributions to R2 statistics in brackets. 6 Conclusions: Five Facts about Prices... Retail and at-the-Dock18 Using the detailed information on product characteristics in the CPI and IPP databases of the U.S. Bureau of Labor Statistics, we match items imported into the United States to those sold at retail that are identical in description. We compute the size of the resulting distribution wedge of CPI price relative to import price and then investigate the determinants of these wedges. We also map distribution wedges into a calculation of the size of the pure traded goods sector, and discuss the implications of our (cid:133)ndings for exchange rate pass-through. We (cid:133)nd the following, 1. Distribution wedges for the United States are large. Ourcalculation is in the range of 50-70% forU.S. data between January 1994- July2007. Wedges are slightly higher under the "alternative procedure" than baseline calculations obtained from the 18With apologies to Emiand Jon. 19

detailed "matching procedure". Back of the envelope calculations using a simple modeling framework of BER (2005) imply that the size of the "pure" tradeables sector in the U.S. is thus in the range of 7-16%. 2. Wedges are larger than previously reported. Our headline number is about 10 to 20 percentage points higher than a consensus estimate of 40-45% which was essentially obtained using NIPA data (Burstein-Neves-Rebelo, Goldberg-Campa, Bradford-Lawrence). This maps into a calculation of the "pure" tradeables sector that is 5 to 10 percentage points lower than the 22% number reported by BER. Di⁄erences between our results and thoseoftheexisitingliteratureappeartobedrivenbydi⁄erencesinthedatasetsused, ratherthanby compositional e⁄ects. Since our calculations using the BLS data are built up from the microeconomic level, we hope they provide a cleaner calculation of distribution wedges than was possible before. 3. Wedges are stable over time but vary considerably across items. Under the matching procedure, the average annual distribution wedge across all items is 0.62, 0.67, 0.63, 0.57, 0.59, 0.60, 0.58, 0.61, 0.59, 0.57, 0.58, 0.60, 0.60 and 0.61 in the years 1994 through 2007 respectively. The relative stability of wedges coincides with large (cid:135)uctuations in the dollar over time. Across item categories, several exhibit low wedges: televisions, video cameras, VCRs, cameras, telephones,microwaveovens,whileothercategorieshavehighwedges: drugs,apparel(men(cid:146)s,women(cid:146)s pants),watches,(cid:133)lm,bananas,tomatoes. 4. Wedges do not vary dramatically with exchange rates or across major exporters. Our average annual distribution wedge across all items lies between 0.57 and 0.61 during a period when dollar (cid:133)rst appreciated by more than 20 percent and subsequently depreciated by more than 30 percent. Moreformalregressionresultsusingtheindividualitemdatacon(cid:133)rmsthelackofarelationship betweenchangesinwedgesandexchangerates. Whenweslicethedatabythecountryofexport,most of the wedges fall within the range of 50% to 60%. 5. Variation in wedges is explained by proxies for sectoral characteristics Between categories, distribution wedges vary negatively with endogenous exits and frequency of price changes, and positively with law of one price deviations in the retail market. Thus, in categories where the wedge is small there is a relatively large amount of product churning or turnover, directly observed at the import stage. This turnover is because signi(cid:133)cant quality changes are made to the product or because other market forces render that product obsolete relatively quickly. These smallwedge item categories are also those in which market forces lead to relatively frequent price changes 20

andkeeppricesrelativelyinlinewiththelawofonepriceattheretaillevel. Withincategories,wedges are explained by brand e⁄ects, the type of retail outlet, and on whether the item is sold at sale price. 21

7 References Anderson, J., van Wincoop, E., 2004. "Trade costs". Journal of Economic Literature 42(3), 691-751. Atkeson, A., Burstein, A., 2008. "Pricing to market, trade costs, and international relative prices". American Economic Review, forthcoming. Bils, M., Klenow, P., 2004. "Some evidence on the importance of sticky prices". Journal of Political Economy 112(5), 947-985. Bradford, S., Lawrence, R., 2003. Paying the price: the cost of fragmented international markets. The Peterson Institute for International Economics. Burstein, A., Eichenbaum, M., Rebelo, S., 2005. "Large devaluations and the real exchange rate". Journal of Political Economy 113(4), 742-784. Burstein, A., Neves, J., Rebelo, S., 2003. "Distribution costs and real exchange rate dynamics". Journal of Monetary Economics 52(6), 1189-1214. Choudri, E., Faruqee, H., Hakura, 2005. "Explaining exchange rate pass-through in di⁄erent prices". Journal of International Economics 65(2), 349-374. Corsetti,G.,Dedola,L.,2005. "Amacroeconomicmodelofpricediscrimination,"JournalofInternational Economics 67(1), 129-156. Corsetti,G.,Dedola,L.,Leduc,S.,2008a. "Internationalrisksharingandthetransmissionofproductivity shocks," Review of Economic Studies 75, 443-473. Corsetti,G.,Dedola,L.,Leduc,S.,2008b. "Modelsofhighexchangeratevolatilityandlowpass-through," Journal of Monetary Economics, forthcoming. Devereux,M.,Engel,C.,Tille,C,2003. "Exchangeratepass-throughandthewelfaree⁄ectsoftheeuro". International Economic Review 44(1), 223-242. Engel,C.,1999. "Accountingforrealexchangeratechanges". JournalofPoliticalEconomy107,507-538. Goldberg, L., Campa, J., 2006. "Distribution margins, imported inputs, and the sensitivity of the CPI to exchange rates". Review of Economics and Statistics, forthcoming. Goldberg, P., Hellerstein, R., 2008. "A framework for identifying the sources of local-currency price stability with an empirical application". working paper. Gopinath, G., Rigobon, R., 2008. "Sticky borders". Quarterly Journal of Economics 123(2), 531-575. Gopinath,G.,Ishtoki,O.,Rigobon,R.,2007. "Currencychoiceandexchangeratepass-through". American Economic Review, forthcoming. Gopinath, G., Ishtoki, O., 2008. "Frequencyofpriceadjustmentandpass-through". HarvardUniversity. Gopinath, G., Gourinchas, P.O., Hsieh, C., 2008. "Cross-border prices, costs, and mark-ups". Harvard 22

University. Nakamura, E., Steinsson, J., 2008. "Five facts about prices". Quarterly Journal of Economics, 123(4), 1415-1464. Nakamura, E., 2008. "Pass-through in retail and wholesale". American Economic Review, forthcoming. Obstfeld, M., Rogo⁄, K., 2005. "The unsustainable U.S. current account position revisited". NBER working paper #10869. 23

8 Appendix A; The Matching Procedure As noted in the text, in calculating distribution wedges d we compare the price of an item in the IPP database to that of a matched item in the CPI database. We match items that are identical in description. This appendix provides details on the criteria we used to construct these matches. Naturally these criteria di⁄ered across item categories. Each potential match was given a grade that depended on how many of the criteria were met successfully. For example, as described below, there were 5 criteria that had to be met in order for there to be an (cid:147)A Grade(cid:148)match for that item: product (e.g., vodka), proof (e.g., 80), size of the container (e.g., 1 liter), brand, and country of origin. When a particular criteria was not met, it was usually because that piece of information was missing. In those cases when there was an obvious mismatch on a criteria, e.g., brand of beer, an F grade was given. In our empirical work we used only A grade and B grade matches. Category: Alcohol Grades: size, proof, BRAND, product, country of origin (mostly A(cid:146)s) Category: Audioplayer Grade Scale A = Brand, model number + other character B = Model number + other charact C = partial model number +other characteristics Category: Bananas Match criteria: Brand, Country, Quality = A B = a country (or brand) discrepancy from a mid sample switch in the CPI C = country & brand tend to be o⁄ Category: Beer Grades: Brand, type, bottles vs. cans, country of origin A = got them all B = usually type and/or country of origin unknown in IPP C = unknown type and container (usually in IPP) and unknown country of origin (usually in CPI) Cans vs. bottles given an F Discrepancy in container size given an F Category: Calculator Grades: Model Number and Brand (C tends to be o⁄on a TI 83 "Plus" vs. no Plus) Category: Cameras 24

If it matched on brand and model number gave it A; If model number was missing but brand matched gave it a C; If it matched model number but brand was missing gave it a B/C depending on how unique the model number seemed; If brands were di⁄erent gave it an F. Category: Computer Accessories A match on brand, model number, screen size/resolution B match on screen size and model number +other characteristics BCmatchonscreensize,othercharacteristicsandpartialserialnumber(notenoughinfoinIPPtoknow for sure) Category: Drugs Grade Scale A = matches on brand, type and size of package and form of drug B = believed to be an exact match but a major product characteristic is not listed in one of the descriptions C=onemajorcharateristiciso⁄matchesonbrandbutippsizeishalfofwhatwewant(C2); or(brand x) vs. (brand x max strength) F = not a match Category: Film Criteria: Brand, Shutter Speed, Exposures Number = Ratio of #rolls in CPI to IPP Category: Mens Pants Qualities: style number, brand, item description A = style number, brand, and basic similarity of item description B = often a downgrading from an A when the style number changed in the CPI sample C = info just too spotty F=clear country of origin contract (cpi=made in USA) Category: Microwaves Letter Grade: Brand, cubic feet, watts A = all 3 B = brand info missing in IPP, cu ft and watt match C = cu ft or watt info missing or o⁄in IPP, brand info missing in IPP F = watt and cu ft o⁄or missing, brand info missing 25

Category: Miscellaneous Kitchen Appliances Brand + model number = A. Really long model number and other miscellaneous characteristics = A. Model number match = B. Anything else C or below. Category: Oliveoil A Matches on Brand, Type, Size and Bottle Type (plastic vs glass vs can) B Matches on Brand, Type, Size C Matches on Brand and Size F Not a Match Category: Phones A: if it matched model/brand and serial OR serial was at least 7 digits and it matched other characteristics B: matched serial only and serial was at least 4 digits C: Matched serial and serial was 3 digits or les F:matchednothingORtherewasde(cid:133)nitiveevidencethetwoweredi⁄erentproducts. Category: Stoves Grades: mostly matched on serial numbers (Brand like George Foreman grill, specs like Bun Warmer) (not much else to go on other than proximity of the serial numbers) Category: Tomatoes We matched on brand, country of origin, type (cherry vs roma) and how it was grown (vine vs green house). If it hit brand and type and at least one of country and how it was grown (and no discrepancy in other) then it was an A. Otherwise if it type and at least one of country and how it was grown, then it was a B. Category: Televisions A matches on brand, model # and size at least B matches on model# and size at least (and may have contradictory country of origin info) C partial model # match and size F not a match (wrong size etc.) or made in USA Category: VCR Qualities: Verbal description of item, model number, brand, country of origin A = got everything essentially B = model number and item description mostly, sometimes a brand match as well (still gave it B) 26

Category: Videocameras Grades: mostly Bs = model number (good matches) and basic item description. Country of origin typically not in CPI, brand typically not in IPP Category: Watch A full serial number, country of origin/brand and other indeti(cid:133)ers B serial number, country of origin + other indenti(cid:133)ers C partial serial number, country of origin F not a match Category: Womens Pants A identi(cid:133)able brand, style number type of pants and country of origin B usually no brand or country of origin C incomplete style number, no brand F made in USA or not a match. 27

Table 1; Distribution Wedges, all items. A. Matching Procedure Intra-company transfer prices excluded regular sale cif 0.57 0.50 fob 0.68 0.60 Intra-company transfer prices only regular sale cif 0.58 0.57 fob 0.62 0.49 *Cells report the median of the distribution wedge, (cid:58), across all items in the sample. Regular (sale) denotes that the CPI price of the item was a regular (sale) price. Cif (fob) denote the price basis of the import price in the IPP database. The calculations in the upper panel exclude all imports whose prices are reported as being intra-company transfer prices. The calculations in the lower panel include only imports whose prices are reported as being intracompany transfer prices.. B. Alternative Procedure Price Levels mean weighted-average (cid:58) 0.70 0.64 *The distribution wedge (cid:58) is calculated using weighted-average price levels for several disaggregated item categories in the CPI and import price data bases. The level of aggregation is by entry level item (ELI) in the CPI, or approximately 10-digit SIC code for imports. The item categories are listed in Table 2. The cells above report the simple mean and the expenditure share-weighted average distribution wedge across those item categories.

Table 2; Distribution Wedges by Item Categories A. Matching procedure Intra-company transfer prices excluded Regular Price (CPI) Sale price (CPI) cif fob cif fob Alcoholic beverages 0.55 0.58 0.51 0.44 Audio players 0.58 0.55 0.52 0.47 Bananas --- 0.72 --- 0.59 Beer 0.53 0.66 0.42 0.62 Calculators --- 0.72 --- 0.70 Cameras --- 0.47 --- 0.40 Computer accessories 0.29 0.36 --- 0.31 Drugs 0.67 0.84 --- --- Film 0.86 0.74 0.82 --- Men’s pants --- 0.75 --- 0.70 Microwave ovens --- 0.46 --- 0.36 Kitchen equip. (misc.) --- 0.66 --- 0.62 Olive oil --- 0.72 --- 0.62 Telephones 0.35 0.42 --- 0.27 Stoves 0.56 0.78 0.55 0.61 Tomatoes 0.83 0.78 0.76 0.70 Televisions 0.28 0.21 0.24 0.35 VCRs 0.44 0.40 0.41 0.34 Video cameras 0.32 0.29 --- 0.23 Watches --- 0.78 --- 0.79 Women’s pants --- 0.64 --- 0.66

Intra-company transfer prices only Regular Price (CPI) Sale price (CPI) cif fob cif fob Alcoholic beverages 0.65 0.58 0.57 --- Audio players 0.52 0.48 0.51 0.42 Bananas --- 0.73 --- 0.61 Beer --- 0.65 --- 0.57 Calculators --- 0.54 --- --- Cameras --- 0.51 --- 0.39 Computer accessories --- 0.43 --- 0.47 Drugs --- 0.85 --- --- Film --- 0.71 --- --- Men’s pants 0.61 0.58 --- 0.49 Microwave ovens --- 0.55 --- 0.34 Kitchen equip. (misc.) --- 0.60 --- --- Olive oil --- 0.81 --- 0.82 Telephones --- 0.39 --- 0.35 Stoves 0.57 0.53 0.56 --- Tomatoes 0.31 0.84 --- --- Televisions 0.47 0.35 0.53 0.36 VCRs --- 0.36 --- 0.32 Video cameras --- 0.33 --- 0.29 Watches --- 0.86 --- --- Women’s pants 0.70 0.82 0.66 ---

B. Alternative Procedure Category Wedge Alcoholic beverages 0.41 Bananas 0.72 Beer 0.69 Computer accessories 0.69 Refrigerator 0.58 Men’s pants 0.62 Olive Oil 0.74 Televisions 0.50 Watches 0.90 Women’s pants 0.59

Table 3a; Distribution wedge by brand, Alcoholic Beverages, transfer prices excluded. Regular Price (CPI) Sale price (CPI) cif fob cif fob All alcoholic beverages 0.55 0.58 0.51 0.44 Brand 1 0.57 0.65 0.53 0.40 Brand 2 --- 0.40 --- --- Brand 3 0.46 0.40 0.51 --- Brand 4 --- 0.52 --- 0.52 Brand 5 0.78 --- --- --- Brand 6 --- --- --- --- Brand 7 0.43 0.48 0.34 --- Brand 8 0.50 0.62 0.46 0.53 Brand 9 0.49 0.56 0.40 0.44 Brand 10 0.58 0.65 0.53 0.54 Brand 11 0.60 0.55 --- --- Brand 12 0.62 --- --- --- Brand 13 --- 0.54 --- --- Brand 14 0.55 0.63 --- --- Brand 15 --- 0.64 --- --- Brand 16 0.48 --- --- --- Brand 17 --- 0.54 --- --- Brand 18 0.58 0.63 --- 0.54 Brand 19 0.56 0.64 0.56 0.58 Brand 20 0.49 --- --- --- Brand 21 0.75 --- --- --- Brand 22 0.59 --- --- --- Brand 23 --- 0.61 --- --- Brand 24 0.61 --- 0.50 --- Brand 25 0.70 0.55 --- 0.39

Table 3b; Distribution wedge by brand, Beer and Television sets, transfer prices excluded. Regular Price (CPI) Sale price (CPI) cif fob cif fob All beer 0.53 0.66 0.42 0.62 Brand 1 0.31 0.66 --- 0.62 Brand 2 --- 0.62 --- 0.58 Brand 3 --- 0.44 --- 0.28 Brand 4 --- 0.66 --- --- Brand 5 0.72 0.73 0.65 0.69 Brand 6 --- 0.82 --- --- Brand 7 0.65 0.66 --- 0.60 Brand 8 --- 0.75 --- --- Brand 9 0.48 0.63 0.42 0.58 Brand 10 0.82 0.73 --- --- Brand 11 --- 0.76 --- 0.68 Brand 12 --- 0.78 --- 0.76 Brand 13 0.60 0.59 --- --- Brand 14 --- 0.49 --- --- Brand 15 --- 0.73 --- --- Brand 16 --- 0.68 --- --- Brand 17 0.41 0.60 0.39 0.56 Brand 18 0.42 0.79 --- --- All televisions 0.28 0.21 0.24 0.35 Brand 1 --- 0.40 --- 0.31 Brand 2 --- 0.15 --- 0.11 Brand 3 0.28 --- 0.24 --- Brand 4 --- 0.08 --- 0.04 Brand 5 --- 0.44 --- 0.40 Brand 6 --- 0.27 --- 0.22

Table 4. Distribution Wedges by Country of Origin cif fob Euro Area 0.48 0.60 Canada 0.55 0.59 China 0.54 0.50 Japan n.a. 0.36 Mexico 0.75 0.55 United Kingdom 0.56 0.59 *Cells contain the median distribution wedge across all items imported from the listed country, under the matching procedure. Sale price and regular price CPI items are included together, and intra-company transfer prices are included along with “arm’s length transaction” prices. (n.a.) Insufficient number of item categories.

Table 5; Decomposing the Variance of Distribution Wedges A. By Category Category IPP on CPI CPI on Wedge IPP on Wedge Alcohol 0.00 0.53 0.15 Audio Players -0.01 0.82 0.08 Bananas 0.00 0.81 0.05 Beer 0.00 0.73 0.07 Calculators 0.00 0.95 0.04 Cameras 0.00 0.35 0.56 Computer Accessories 0.24 0.03 0.71 Men’s Pants 0.00 0.60 0.07 Microwave Ovens -0.02 0.91 0.09 Kitchen Equip. (misc.) -0.01 0.93 0.04 Olive Oil 0.00 0.84 0.03 Telephones -0.01 0.80 0.33 Stoves 0.00 0.89 0.02 Tomatoes 0.03 0.25 0.29 Televisions 0.00 0.73 0.44 VCRS 0.00 0.69 0.08 Video Cameras -0.01 0.90 0.05 Women’s Pants -0.01 0.56 0.07

Table 5, cont’d B. By Year Year IPP on CPI CPI on Wedge IPP on Wedge 1994 0.00 0.64 0.02 1995 0.00 0.85 0.01 1996 0.04 0.69 0.03 1997 0.00 0.71 0.14 1998 0.00 0.84 0.06 1999 0.00 0.68 0.05 2000 0.00 0.78 0.08 2001 0.00 0.55 0.16 2002 0.01 0.41 0.15 2003 0.01 0.56 0.08 2004 0.02 0.45 0.10 2005 0.01 0.37 0.22 2006 0.00 0.38 0.12 2007 0.00 0.17 0.47 C. By Country IPP on CPI CPI on Wedge IPP on Wedge Euro Area Canada 0.24 0.68 0.25 China 0.0 0.53 0.28 Japan 0.0 0.69 0.04 Mexico 0.02 0.28 0.18 United Kingdom 0.04 0.80 0.39 Notes: The cells above contain the R-squared values from a regression of the IPP price of the item on its corresponding CPI price (first column), CPI price on the distribution wedge (second column), or IPP price on the wedge (final column).The variables are in log first-differences. All regressions contain a constant but no lags.

Table 6; Changes in distribution Wedges and changes in exchange rates A. Results by Item Category Contemporaneous Exchange Rate Joint F Stat Adjusted R Squared n 1.24 3.37* 0.034 326 Video cameras (.63) 0.71 1.50 -0.01 382 Telephones (.87) -0.16 0.31 0.02 152 Watches (.13) -6.35 0.67 -0.02 135 Computer accessories (8.91) 0.02 3.18* 0 3639 Alcoholic beverages (.11) -0.009 0.07 0.003 1345 Televisions (.36) -0.86 0.62 -0.01 197 Women’s pants (.60) 0.02 1.57 0.01 212 Olive oil (.34) -0.04 0.04 0 3979 Beer (.04) 0.11 0.16 0 7122 Bananas (.12) -0.02 0.09 0.004 484 Audio players (.48) 0.36 0.04 -0.01 335 Cameras (.35) -0.17 3.49* 0.54 115 Drugs (.07)* -0.53 1.49 0.002 527 Film (.26) 0.21 0.32 -0.01 472 Men’s pants (.36) -3.06 0.58 -0.004 245 Kitchen equip. (misc.) (1.79) -0.91 6.48** 0.04 133 Microwave ovens (.65) -0.21 1.51 -0.007 364 Stoves (.40) 0.33 1.71 0.005 1767 Tomatoes (.17) 0.01 0.007 0.002 707 VCRs (.24)

B. Results by Country of Origin Country Contemporaneous Exchange Rate Joint F Stat Adjusted R Squared n .03 Euro Area 1.04 0 2333 (.07) -.62 Canada 2.35* 0.002 1147 (.41) 1.95 China 1.07 -0.001 1369 (2.66) .08 Japan 0.06 -0.006 906 (.24) .03 Mexico 0.2 0.001 5132 (.10) -.05 -0.004 United Kingdom 0.024 1010 (.14) Notes: Standard errors in parenthesis. A * (**) denotes Significant at the 5% (1%) level.

Table 7; An Empirical Model of Distribution Wedges A. By Item Category Category Full Model Brand Business Type Sales Flag Price Basis LOP Deviations Endo Exits Alcoholic .10** -.08** .03* .55** .06** beverages [0.24] [0.22] (0.01) [0.03] (0.01) [0.04] (0.01) [0.01] (0.06) [0.01] (0.01) [0.05] .04 -.08** .06* .16** .01 Audio players [0.69] [0.49] (0.02) [0.23] (0.01) [0.03] (0.01) [0.04] (0.02) [0.08] (0.01) [0.01] .13** -.14** .95** -.06** Bananas [0.30] [0.01] (0.01) [0.04] (0.002) [0.23] na (0.09) [0.008] (0.004) [0] .09** -.05** .09** -.35** .02 Beer [0.61] [0.48] (0.01) [0.03] (0.003) [0.01] (0.002) [0.13] (0.03) [0.01] (0.03) [0.05] .10** -.08** -.02 .78** -.10* Calculators [0.78] [0.35] (0.004) [0.01] (0.01) [-0.002] (0.02) [0.2] (0.04) [0.32] (0.02) [0.19] .04* -.07** .06 .004 Cameras [0.57] [0.63] (0.01) [0.01] (0.01) [0.07] na (0.11) [0.01] (0.03) [0.004] Drugs na na na na na na na .03* -.05 -.81** 1.6** .37** Film [0.49] [0.07] (0.01) [0] (0.02) [0.01] (0.05) [0.04] (0.06) [0.16] (0.04) [0.05] .02 -.09** 0 -.27** -.01 Men’s pants [0.75] [0.03] (0.01) [0.02] (0.01) [0.001] (0) [0.03] (0.03) [0.02] (0.01) [0] 0 -.07** .42 0 Microwave ovens [0.58] [0.58] (0) [0.13] (0.01) [0.08] na (0.3) [0.13] (0) [0.17] Kitchen equip. .15** -.15** 0 .1** .07** (misc.) [0.87] [0.51] (0.01) [0.14] (0.01) [0.05] (0) [-0.001] (0.03) [-0.002] (0.01) [0.04] .47** -.08** -.55** .02 Olive oil [0.73] na (0.02) [0.55] (0.01) [0.01] na (0.06) [0.07] (0.01) [0.1] -.06 -.09 .05 -.41** -.27** Telephones [0.16] [0.11] (0.03) [0.02] (0.05) [0.002] (0.08) [0] (0.06) [0.06] (0.04) [0.01] .08** -.08** 0 -.36** .07** Stoves [0.74] [0.53] (0.01) [0.003] (0.01) [0.07] (0) [0.31] (0.03) [0.05] (0.01) [0.07] .09** -.04* -.04* .8** -.01* Tomatoes [0.18] na (0.01) [0.03] (0.01) [0.05] (0.01) [0.02] (0.04) [0.06] (0.004) [0.004]

.05* .02 .02 -.24** -.05* Televisions [0.77] [0.72] (0.01) [0.35] (0.01) [0.01] (0.02) [0.002] (0.07) [0.11] (0.01) [0] .03* -.05* -.03* .37** -.05* VCRs [0.42] [0.37] (0.01) [0.08] (0.01) [0.06] (0.01) [0.002] (0.05) [0.01] (0.01) [0.01] .004 -.05* 0 -.14 .07** Video cameras [0.29] [0.11] (0.01) [0.003] (0.01) [0.04] (0) [0.003] (0.12) [0.02] (0.01) [0.01] .03** -.06** 0 .32** -.04** Watches [0.76] none (0,004) [0.17] (0.01) [0.08] (0) [0] (0.02) [0.36] (0.003) [0.25] .21** -.10* 0 .28** -.04 Women’s pants [0.49] [0.20] (0.02) [0.32] (0.02) [0.02] (0) [0.04] (0.05) [0.03] (0.04) [0.14] B. By Country of Origin Absolute Country Full Model Brand Business Type Sales Flag Price Basis Endo Exits Deviations .18** -.06** .10** Euro Area [0.46] [0.34] (0.01) [0.05] (0.01) [0.002] (0.01) [0.14] [0.01] [0.002] -.01 *-0.11** .05* Canada [0.56] [0.52] (0.02) [0] (0.01) [0.15] (0.01) [0.18] [0.02] [0.2] .06 -.10** .05* China [0.75] [0.54] (0.01) [0.05] (0.01) [0.03] (0.01) [0.002] [0.01] [0.06] .04* -.06** -.22 Japan [0.56] [0.55] (0.01) [0.01] (0.01) [0.01] (0.09) [0.04] [0.26] [0] .09** -.06** -.04** Mexico [0.76] [0.74] (.004) [0.23] (0.003) [0.02] (0.004) [0.06] [0.44] [0.07] .04 -.08** -.08 United Kingdom [0.27] [0.22] (0.03) [0.004] (0.01) [0.05] (0.01) [0.02] [0.02] [0.004] Notes: Standard errors are on parenthesis. In brackets are contributions to R2 from the column variable. A * (**) denotes significance at the 5% (1%) level.

Appendix to Table 1a and 2a: number of observations (intra-company transfer prices excluded) Regular Price (CPI) Sale price (CPI) cif fob cif fob All items 6054 14,090 1018 3316 Alcoholic beverages 2959 1108 660 166 Audio players 150 290 23 58 Bananas --- 3482 --- 793 Beer 1509 3155 148 463 Calculators --- 214 --- 19 Cameras --- 199 --- 39 Computer accessories 20 110 --- 8 Drugs 7 89 --- --- Film 581 114 42 --- Men’s pants --- 128 --- 328 Microwave ovens --- 126 --- 81 Kitchen equip. (misc.) --- 157 --- 42 Olive oil --- 170 --- 58 Telephones 65 341 --- 41 Stoves 82 246 29 71 Tomatoes 513 2231 62 544 Televisions 105 820 22 308 VCRs 56 633 29 153 Video cameras 7 206 --- 37 Watches --- 134 --- 60 Women’s pants --- 103 --- 35

Appendix to Table 1a and 2a: number of observations (intra-company transfer prices only). Regular Price (CPI) Sale price (CPI) cif fob cif fob All items 546 6210 173 1259 Alcoholic beverages 230 7 92 --- Audio players 172 130 34 9 Bananas --- 3447 --- 573 Beer --- 122 --- 14 Calculators --- 36 --- --- Cameras --- 299 --- 46 Computer accessories --- 30 --- 14 Drugs --- 51 --- --- Film --- 21 --- --- Men’s pants 12 149 --- 96 Microwave ovens --- 10 --- 7 Kitchen equip. (misc.) --- 171 --- --- Olive oil --- 69 --- 12 Telephones --- 226 --- 18 Stoves 76 14 21 --- Tomatoes 6 10 --- --- Televisions 18 797 6 287 VCRs --- 283 --- 72 Video cameras --- 266 --- 93 Watches --- 49 --- --- Women’s pants 27 10 20 ---

Distribution Margins vs IPP Stickiness Correlation = -.31 0.9 0.8 Tomatoes Oliveoil 0.7 Bananas 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Probability of a Price Change nigraM noitubirtsiD Distribution Margins vs Absolute LOP Deviations Correlation = .68 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 LOP Deviations snigraM noitubirtsiD Distribution Margins vs Endogenous Exits Correlation = -.37 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Endogenous Exits snigraM noitubirtsiD Distribution Margins vs CPI Stickiness Correlation = -.42 0.9 0.8 Tomatoes 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Probability of a Price Change nigraM noitubirtsiD