Heterogeneous Firms and Import Quality: Evidence from Transaction-Level Prices

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 991 January 2010 Heterogeneous Firms and Import Quality: Evidence from Transaction-Level Prices Benjamin R. Mandel NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. References to International Finance Discussion Papers (other than an acknowledgement that the writer has had access to unpublished material) should be cleared with the author or authors.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/. This paper can be downloaded without charge from Social Science Research Network electronic library at www.ssrn.com.

Heterogeneous Firms and Import Quality: Evidence from Transaction-Level Prices Benjamin R. Mandel* Abstract A key emerging insight in international economics is that the scope for quality differentiation can help to explain patterns in export prices at the level of products or firms. In this paper, a unified theoretical framework of firm heterogeneity in cost and quality is brought to bear on an expansive data set of U.S. import transaction prices collected by the Bureau of Labor Statistics (BLS). The higher moments of the price distribution are used to identify the scope for quality differentiation at the detailed product level; highly differentiated products account for about half of U.S. import value. The product classification is then used to evaluate two claims in the nascent firm-level trade quality literature. First, the positive link between exporter capability and price is found to depend on the nature of the product: productive exporters simultaneously specialize in high-priced varieties in quality differentiated goods and low-priced varieties in more homogeneous goods. Second, a novel time series test documents firm sorting into export markets according to output quality. Keywords: Quality differentiation, heterogeneous firms, firm sorting JEL classification: F12, F41 *Division of International Finance, Board of Governors of the Federal Reserve System, Washington, D.C. 20551 U.S.A. Email: Benjamin.R.Mandel@frb.gov. I especially thank Robert Feenstra for guidance and constructive suggestions, David Hummels and James Harrigan for insightful discussant comments, as well as Robert McClelland and Rozi Ulics at BLS. I gratefully acknowledge helpful feedback by seminar participants at BLS, FRBNY, Federal Reserve Board, Georgetown, Johns Hopkins SAIS, UC Davis, UCSD IRPS, UVA, UW Madison, and the 2008 conferences of EIIT and ITFA. The views in this paper are solely the responsibility of the author and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or any other person associated with the Federal Reserve System.

1 Introduction The price levels of internationally traded goods are largely driven by marginal costs of production, re(cid:135)ecting the underlying productivity di⁄erences across exporting (cid:133)rms, and product quality, the heterogeneity in desirable characteristics across competing varieties. A key emerging insight in international economics is that the scope for quality di⁄erentiation can help to explain observed patterns in average export prices at the level of products or (cid:133)rms.1 That is, in contrast to industry models of (cid:133)rm heterogeneity in which high productivity (cid:133)rms set lower prices,2 a theory in which higher productivity (cid:133)rms set higher prices for higher quality goods helps to match broad empirical facts. The tension between the lack of comprehensive measures of unobserved product quality and their usefulness in describing trade patterns has given rise to two common assertions, that: (i) (cid:133)rms sort into export markets according to their output quality, and (ii) capable exporters set high prices for high quality outputs. This paper evaluates these claims using newly developed methods and (cid:133)nds support for (i) and a modi(cid:133)ed version of (ii). In addition to improving our understanding of trade patterns, quality(cid:146)s role in price setting has important implications for how we actually measure prices. Mis-measured or ignored compositional changes create bias in prices, as discussed at length in the index numberliterature,3 andthereismountingevidencethatthedynamicsofqualitydi⁄erentiated goods prices are di⁄erent relative to those of more homogeneous goods.4 Therefore, there is strong impetus to be able to quantify the scope of quality di⁄erentiation across products and exporting countries, and to account for their respective dynamics. In this paper, a 1Qualitydi⁄erentiationissuggestedbySchott(2004)andHummelsandKlenow(2005)toexplainincreasingunitvaluesinexporterincome, invokedbyBaldwinandHarrigan(2007)toexplainincreasingunitvalue prices in exporter distance, by Kugler and Verhoogen (2008) to explain plant-level size-price correlations, andbyManovaandZhang(2009)andCrozet,HeadandMayer(2009)toanalyzecorrelationsofexportprice with several destination characteristics. 2Inmodelsofthatsort,(cid:133)rmsdi⁄eracrossanarrayofproductivitylevelsandcompeteinamonopolistically competitive industry. See Melitz (2003). Herein I refer to the absence of quality di⁄erentiation with the descriptor: (cid:145)cost.(cid:146) 3For example, a price increase for an upgraded model of automobile may be re(cid:135)ecting its more desirable featuresrelativetoitspredecessor. Alternatively, ashiftindemandintohigherpricedgoodswouldincrease an average (non-quality-adjusted) price even if each underlying price went unchanged. The distinction and large potential di⁄erences between constant quality and non-quality-adjusted price indexes is discussed and illustrated in Alterman (1991). 4For instance, see Auer and Chaney(cid:146)s (2007) discussion of quality and exchange rate pass-through where varieties of di⁄erent quality (within an industry) have di⁄erent price sensitivity to exchanage rates.

2 straightforward departure from existing (cid:145)cost(cid:146)models of (cid:133)rm heterogeneity is modeled to allow for an endogenous array of quality types, and this uni(cid:133)ed framework is brought to bear on an expansive data set of transaction-level U.S. import and export prices collected by the International Price Program (IPP) of the Bureau of Labor Statistics (BLS). The focus is on the distribution of transactions prices and (cid:133)rm productivity within narrowly de(cid:133)ned product groups to make inference as to the scope of unobserved product quality for the majority of U.S. trade, and to explore the higher moments of the price distribution. Indexes of exporter capability and quality are then constructed to measure their sensitivity to changes in the macroeconomy, across countries and over time. The point of departure for these empirical exercises is a classi(cid:133)cation scheme of products according to their scope for quality di⁄erentiation5 which, in turn, is derived from the upper moments of the U.S. import price distribution. In section 1, I provide a descriptive analysis of U.S. import prices within detailed harmonized system 10-digit categories. I (cid:133)nd signi(cid:133)cant clustering of prices within products and skewness that is highly industry-speci(cid:133)c. I then extend a benchmark model of heterogeneous (cid:133)rms to show that these patterns are consistent with a theoretical framework in which (cid:133)rms endogenously choose their level of quality and sort into export markets accordingly. In section 2, I use the theoretical setup(cid:146)s sharp predictions about the higher moments of the price level distribution to identify differences in the scope for quality di⁄erentiation across product groups. This contrasts with prior studies of product or sectoral quality measurement, such as Hallak and Schott (2009), Khandelwal (2009), Baldwin and Ito (2008), Johnson (2008), Harrigan & Barrows (2006), Feenstra (1988) and Aw and Roberts (1986) in that it directly exploits intra-product pricing patterns without relying on inference from average prices at any level of aggregation. Moreover, it is argued that the identi(cid:133)cation of quality ladder length is robust to more general model speci(cid:133)cations with variable markups across (cid:133)rms. Employing measured import price skewness, and controlling for other factors a⁄ecting the distribution of prices, (cid:145)cost(cid:146)industrieswithlowscopeforqualitydi⁄erentiationaredistinguishedfrom(cid:145)quality(cid:146)industrieswith a high scope for quality di⁄erentiation. I (cid:133)nd that quality industries account for roughly half of U.S. import value. 5Notethedistinctionbetweenverticalqualitydi⁄erentiationandthestandardassumptioninmonopolistic competitionmodelsofdi⁄erentiatedvarieties(i.e.,horizontaldi⁄erentiation). Inthehorizontalcase,varieties are distinct but provide equal value to the consumer(cid:146)s utility. In the vertical case, as will be made explicit in the model below, varieties enter asymmetrically into utility.

3 Given the model and resulting classi(cid:133)cation scheme, it is possible to discern from price data alone the high productivity/high quality exporters from the low. The theory suggests that within an industry, each exporting country(cid:146)s productivity level and specialization in qualitycharacteristics canbe ascertainedbyits locationinthe U.S. import price distribution (i.e., a relatively high price in a long-quality ladder industry denotes both high quality and capability). In section 3, I use quantile regression techniques to identify country export productivity and (cid:133)nd that, within a given sector, more productive countries tend to sell (cid:145)quality(cid:146)productsathigherpricesand(cid:145)cost(cid:146)productsatlowerpricesonworldmarkets. This result re(cid:133)nes the income-quality nexus suggested by Schott (2004), Hummels and Klenow (2005) Hallak (2006) and Choi, Hummels and Xiang (2006), in that country export quality is not a monolith: high unit value prices in wealthier (more productive) countries belie specialization in low priced varieties in less-quality-di⁄erentiated sectors. Finally, the paper provides a novel time series test of (cid:133)rm sorting by output quality. The model has stark implications for the average prices of imports in response to any shock that alters the composition of (cid:133)rms participating in trade. I describe and measure the implications of quality sorting for real-exchange rate pass-through; index number techniques are used to identify the relative price and quality of entering and exiting (cid:133)rms, and hence changes in composition due to the extensive margin. Consistent with the predicted ordering of (cid:133)rms in the model, I (cid:133)nd that pass-through is systematically higher in unit values (i.e., not controlling for composition) than in constant-quality prices. The remaining sections are organized as follows. The next section describes patterns in the price distribution of U.S. imports, which motivate the model of endogenous (cid:133)rm quality choice presented in section 2. Then, the cross-section and time series of the price and quality distributions in U.S. imports are detailed in sections 3 and 4, respectively. Section 5 concludes.

4 1 The IPP Import Price Data TheIPPdata,whichconsistoftransaction-level(cid:145)at-the-dock(cid:146)pricesforapproximately40,000 imported and exported items per month,6 provides a large breadth of coverage for roughly the entire range of U.S. goods import industries over the period 1994-2006. Transaction prices are surveyed from U.S. importers for uniquely de(cid:133)ned items; IPP sta⁄take a detailed description of each item and respondents are asked to provide its unit price on a monthly basis going forward. A panel of this size and diversity makes it an extremely useful tool in the investigation of international price-setting.7 Previously, studies have used national micro-data to investigate the quality composition of aggregate price de(cid:135)ators. For example, for the micro-data underlying the CPI, Bils (2004) quanti(cid:133)es the di⁄erence between price and quality growth by examining the point at which one product is substituted for another and the explicit quality adjustment made by the BLS. Bils(cid:146)s work is motivated by earlier estimates of quality bias in the CPI by the Boskin Commission (1996) and by Moulton and Moses (1997). Since the IPP constructs a matched model index for imports, which is quality-adjusted by construction,8 the empirical methods used by Bils are not applicable and new means of inference must be devised to obtain information about aggregate quality from individual prices. I begin by analyzing the price distribution for very disaggregate product groups, which o⁄ers some clues to that end. In terms of classi(cid:133)cation, we consider the disaggregate, (cid:145)narrowly-de(cid:133)ned(cid:146)product group as a Harmonized System 10-digit (HS10) category. The IPP uses a similar de(cid:133)nition called (cid:145)classi(cid:133)cation(cid:146),whichgroupstogethersmaller,similarHS10products;9 sinceinroughlyhalfof cases, classi(cid:133)cation groups are identical to HS10 groups (with the majority of the remainder containing two HS10 groups), I will refer to them synonymously. Examples of HS10 groups in U.S. imports are: 6The IPP collects prices for roughly 20,000 imported items and 20,000 exported items per month. Over the course of the sample, approximately 60,000 imported items were observed. 7The IPP data has also been used for the measurement of the frequency of price changes as well as exchange rate pass through. For detailed descriptions of the data, see: BLS (2009); Nakamura & Steinsson (2009); Gopinath & Rigobon (2008); and Gopinath, Itskhoki & Rigobon (2007). 8The IPP tracks unique varieties of product groups over time and then aggregates the varieties using (cid:133)xedweights. Therefore,thecompositionofitemcharacteristicsisheldliterallyconstantintheaggregation of each period(cid:146)s index. 9ThereasonthattheBLSusesslightlymoreaggregategroupsisthatnoteverygroup(particularlysmaller ones) are represented in the IPP sample.

5 1. Portable digital automatic data processing machine, not more than 10 kg, w/ CPU, keyboard & display 2. Cucumbers, gherkins; Entry 12/1-end of February; fresh or chilled The(cid:133)rstexampleissimplythedescriptionascribedtolaptopcomputers, whichcomposes the majority of the more aggregate HS6 category: handheld computers. In the IPP sample, a laptop price observation would be for a particular, precisely de(cid:133)ned model and brand imported in a particular unit of measure, from a given country in a given month.10 For laptops, itisclearthatmostdi⁄erentiationacrosslaptopstakesplacewithintheHS10group; thatis, di⁄erentiationinscreensize, memoryandprocessorspeedwillallbere(cid:135)ectedinprice di⁄erences at the item level. In contrast, the category of winter cucumbers and gherkins is relatively homogeneous, and there are only six sub-categories de(cid:133)ned by the USDA denoting cucumber coloration, formation and size. I will show that the price distribution is indeed quite di⁄erent between these types of products, which can be used to distinguish their scope of quality di⁄erentiation. The general contours of the sample of import prices at the HS10 level are as follows. The entire 12 year sample contains over 66,000 items (speci(cid:133)c varieties) categorized into over 7,000 classi(cid:133)cation groups.11 Certain product groups (e.g. HS84 and HS85 which contain the sizable machinery and electric equipment categories) are relatively large, though the average number of items per classi(cid:133)cation group does not vary wildly across categories. Figure1showsanillustrativeexampleof (log)importpriceswithinaparticularclassi(cid:133)cation group over time. Each blue dot is a monthly import price for a speci(cid:133)cally de(cid:133)ned item,12 10Although the frequency of the sample is monthly, actual price observations may not be available each period and are imputed by the BLS; this either means that the item price remains unchanged or is imputed linearlybasedongrouppricechanges. Foratypicalitem,pricesarerelativelystablewithoccasionalchanges. For a typical classif group, certain items are observed only sporadically while others enter or exit over the course of the data span; this creates several issues for aggregation which will be addressed separately for the uses of the data described below. For our purposes it will often be convenient to leave imputed price values in the sample, as the BLS does when it computes import price de(cid:135)ators. Leaving in imputations contrasts with other uses of the IPP data, such as the examination of the frequency of price changes. Whereas in frequency calculations imputations can cause biased statistics, imputed values actually add stability to the sample over time when computing the distribution at low levels of aggregation. 11Duetochangingclassi(cid:133)cationcategoriesovertime,thenumberofgroupsinanygivenmonthislessthan 7,000. 12In the analysis of classi(cid:133)cation group price distribution, the reported units of sale are also controlled for; I distinguish between goods sold by the ton versus those sold by container which would be priced non-comparably. Also removed from the sample are those items not priced in dollars.

6 and the long, straight series of dots re(cid:135)ect the fact that most items do not change prices very frequently. For the product in Figure 1, the item prices tend to diminish over time, but the distribution in any given month looks fairly similar at any given point along the horizontal axis. Generally, withinHS10groups (andinagivenmonth) thereis acleardistinctionbetween the high and low priced goods, typically manifested in a tight group of prices around the medianandafewoutlyinghigherorlowerpriceditems. Theclusteringofpricesisconsistent with what we would expect from a power law-type distribution of (cid:133)rm productivity in that thelargemassof(cid:133)rmswithsimilar(relativelylow)productivitysetspriceswhicharesimilar. Studies of detailed (cid:133)rm-level data, as in Bernard and Jensen (1999) and related works, have shownthat thesize distributionof (cid:133)rms, whichmoves intandemwith(cid:133)rm-level productivity in most models, looks something of this sort with a tight cluster of small (cid:133)rms and fewer, much larger (cid:133)rms. That right skewed shape has motivated the calibration assumption in quantitative trade models that the productivity distribution of (cid:133)rms is Pareto. Returning to our example in Figure 1, the high price outliers are exactly the opposite of what we would expect from a Pareto distribution of (cid:133)rm productivity since the large, high productivity (cid:133)rms should be setting prices that are low re(cid:135)ecting their cost advantages. One potential explanation is that more productive (cid:133)rms elect to produce higher quality, more costly goods whose prices are high outliers. Aggregating across HS10 products and time within 16 broad sectors,13 Table 1 presents some of the moments of the import and export price distributions. The statistics are calculated monthly by HS10 group and then averaged over months and HS10 groups within a sector using sales weights. For U.S. imports, we con(cid:133)rm the signi(cid:133)cant clustering of prices across sectors, with the average kurtosis found to be greater than 3 (i.e., leptokurtic, with a more acute peak than a standard normally distributed variable). The skewness statistic, ontheotherhand, appearstohavealargeproduct-orindustry-speci(cid:133)ccomponent. For industries we might imagine to have a lower degree of product di⁄erentiation, such as wood and mineral products, skewness is negative (i.e., skewed left; with low outliers). For 13Sector categorization is de(cid:133)ned by HS2: Animal and animal products (1-5); Vegetable products (6-15); Foodstu⁄s (16-24); Mineral products (25-27); Chemicals & allied industries (28-38); Plastics & rubber (39- 40);Rawhides,skins&leather(41-43);Wood&woodproducts(44-49);Textiles(50-63);Footwear/Headgear (64-67);Stone/glass(68-71);Metals(72-83);Mechanicalandcomputers(84);Electricmachinery(85);Transportation (86-89); Miscellaneous (90-96).

7 industries with a higher degree of product di⁄erentiation and value added, such as textiles, electricmachineryandcomputers, skewnessispositive(i.e., skewedright; withhighoutliers). For U.S. exports, the skewness of prices tends to be lower overall but the ordering of sectors is remarkably similar, with primary goods and commodities more negatively skewed and higher value added manufacturing and technology sectors more positively skewed.14 Based on the observed distribution of price levels in the data, the skewness statistic is indicative of the type of di⁄erentiation in each industry. In the less quality-di⁄erentiated industries, one might expect a higher degree of cost and price competition, as goods are less de(cid:133)ned by their characteristics. With more competition in cost, only the high productivity (cid:133)rmoutliersareabletobreakfromtheclusterofmedianpricesineachcategorytoo⁄erlower prices, thus skewness is negative. On the other hand, with a higher degree of di⁄erentiation in characteristics, it is only those (cid:133)rms that are productive enough to bear the costs of innovation and more intricate processes that are able to produce high quality, high priced goods, causing skewness to be positive. The overall skewness of prices is positive, which is re(cid:135)ected in the higher incidence of high outliers relative to low outliers.15 This is consistent with empirical work such as Baldwin and Harrigan (2007), which suggests that on average, unit value prices are biased upward by quality. Here the data suggest that underlying the average positive skewness of U.S. import prices is an array of industries whose price distribution (and hence unit values) re(cid:135)ect the cost advantages of high productivity (cid:133)rms in two distinct ways. Also of interest is the observation that variable skewness across industries is caused not only by an increase in high priced outliers, but a decrease in low priced outliers. Aside from quality, another interpretationoflow-priceoutliersisthenewentryontoworldmarketsoflowcostproducers, sometimes referred to as a (cid:145)Wal-mart e⁄ect.(cid:146) In the following sections, I attempt to (jointly) quantify these e⁄ects for U.S. imports and (cid:133)nd that there are very important product- and country-speci(cid:133)c components that determine a good(cid:146)s likelihood of being a high or low outlier in the IPP sample. To guide those empirical exercises, in the next section I specify a model 14It is not obvious a priori that export price distributions should be similar to import price distributions: U.S. exporters compete in a di⁄erent set of products and within products U.S. exporters may optimally specializeonadistinctportionofthequalityladder(assuggestedbyKhandelwal(2008)andSchott(2008)). Moreover, there may exist non-monotonicities in export participation (as in Hallak and Sivadasan (2008)) that would a⁄ect the price distribution of exports versus imports. 15High outliers are de(cid:133)ned as prices greater than one standard deviation above the median price. Low outliers are de(cid:133)ned as prices less than one standard deviation below the median price.

8 of (cid:133)rm heterogeneity in productivity and quality. 2 A Model of Endogenous Quality Choice Themodelusestheprobabilisticframeworkof (cid:133)rmheterogeneitywithmonopolisticallycompetitive (cid:133)rms distinguished in their productivity, as in Melitz (2003) and several subsequent works. Firms are di⁄erentiated along an array of productivities (indexed by ’(!)) where each (cid:133)rm produces a unique variety !. Production of each variety in every HS10 product groupissubjecttopositiveconsumerdemandwhich, inturn, isbasedonaconstantelasticity of substitution sub-utility function: (cid:27) X = [x(!)z(!)] (cid:27) (cid:0)(cid:27) 1 d! (cid:27) (cid:0) 1 (cid:26)Z! (cid:10) (cid:27) 2 where x(!) is the quantity of variety !, and z(!) is the (cid:145)weight(cid:146)attributed to the unique characteristics of that variety. Of course, consumers only care about the composite good x(!)z(!), which de(cid:133)nes the (cid:145)quality-adjusted(cid:146)quantity d(!) = x(!)z(!). Consumers compare quality-adjusted varieties and their respective quality-adjusted prices in deciding their purchase allocation, where quality-adjusted price q(!) is de(cid:133)ned as the sticker price p(!) normalized by quality: q(!) = p(!)=z(!). This speci(cid:133)cation of utility gives rise to the standard demand and expenditure (r(!)) functions in quality-adjusted terms: q(!) (cid:27) (cid:0) d(!) = D (1) Q (cid:18) (cid:19) q(!) 1 (cid:27) (cid:0) r(!) = R (2) Q (cid:18) (cid:19) where D is a composite quantity of the di⁄erentiated product and R aggregate expenditure on that product. The aggregate quality-adjusted price index is: 1 1 (cid:27) Q = q(!)1 (cid:27) (cid:0) (3) (cid:0) (cid:18)Z! (cid:19) Productionaccountsexplicitlyforthecostsofquality, andeach(cid:133)rmemployslaborinputs

9 L(!): (cid:17)z(!)a L(!) = f +x(!) (cid:20)z(!)b + (4) ’(!) (cid:26) (cid:27) Labor consists of a (cid:133)xed cost f and two variable costs which are increasing in the level of quality. Inordertohaveaninteriorsolutionforqualitylevel,theaveragecostofqualitymust have a well-de(cid:133)ned minimum. As I will show, a su¢ cient condition for a non-degenerate distribution of quality is b < 1 < a. What these unit cost curvature assumptions imply is that there are increasing returns to quality in z(!)b, and decreasing returns to quality in z(!)a. For intuition, consider (cid:17)z(!)a to be a (cid:145)process(cid:146)component and (cid:20)z(!)b to be a (cid:145)monitoring(cid:146)component. Process costs, in addition to increasing in quality at an increasing rate, are lower for more productive (higher ’) (cid:133)rms. These can be understood as something analogous to an input material cost: a graphite tennis racquet costs more to produce than a wooden one, and a more productive (cid:133)rm requires fewer people to assemble it. Units of monitoring cost, on the other hand, increase in quality at a decreasing rate and are the same across (cid:133)rms. One might imagine this as quality control infrastructure: adding a supervisor to inspect the tennis racquets for visible (cid:135)aws decreases the incidence of defects (i.e., increases quality), though his or her cost is no greater for a graphite racquet than a wooden one.16 This monitoring aspect of production can alternatively be interpreted as a reduced form of the O-ring production function proposed by Kremer (1993).17 In this case, (cid:133)rms that choose to produce more complex/ higher quality products have an incrementally higher wage bill due to monitoring costs. Given this technology, the Home (cid:133)rm(cid:146)s pro(cid:133)t function is: (cid:25)(!) = x(!)p(!) wL(!) (cid:0) (cid:17)z(!)a = x(!) p(!) w (cid:20)z(!)b + wf (cid:0) ’(!) (cid:0) (cid:26) (cid:20) (cid:21)(cid:27) 16A more speci(cid:133)c interpretation of monitoring costs could be as the costs of quality-di⁄erentiated inputs (whosequalityleveliscomplementarytothatoftheoutput), asinKuglerandVerhoogen(2008). Here, the particular reason for increasing returns to quality is not as relevant, so I will proceed with the more general speci(cid:133)cation. 17In Kremer (1993), production is separated into a set of tasks undertaken by workers of varying skill. The equilibrium outcome is that workers of the same skill are matched together in (cid:133)rms, with higher skill (cid:133)rms paying a higher wage bill. Moreover, when technology choice is endogenous (i.e., (cid:133)rms choose the number of tasks, and hence complexity, associated with production), higher skill (cid:133)rms choose more complex production processes. Verhoogen (2007) employs an O-ring production function to explicitly show returns to skilled and unskilled labor.

10 where w is the wage rate, considered exogenous by the (cid:133)rm. The expression for pro(cid:133)t can also be written in the quality-adjusted notation described above: (cid:17)z(!)a 1 (cid:25)(!) = d(!) q(!) w (cid:20)z(!)b 1 + (cid:0) wf (5) (cid:0) (cid:0) ’(!) (cid:0) (cid:26) (cid:20) (cid:21)(cid:27) Quality and quality-adjusted prices are chosen simultaneously by the (cid:133)rm, which allows us to separate the (cid:133)rm problem into two parts. For the quality component, (5) is maximized by the (cid:133)rm with respect to z (z ); given that the expression for pro(cid:133)t is in terms of quality- (cid:3) adjusted quantity and price, it is immediately clear that maximizing pro(cid:133)ts is equivalent to minimizing the term in square brackets, the average cost of quality: (cid:17)z(!)a 1 min (cid:20)z(!)b 1 + (cid:0) (cid:0) z ’(!) (cid:26) (cid:27) 1 = ) z(’) = (cid:18) z ’a (cid:0) b (6) 1 where (cid:18) z = ( ( a 1 (cid:0) 1 b) ) (cid:20) (cid:17) a (cid:0) b is a positive constant. The result is an expression for the quality (cid:0) characteristi(cid:16)cs chose(cid:17)n by each (cid:133)rm as a function of its productivity draw, where the quality of each variety is increasing in the level of (cid:133)rm productivity. The pricing rule is derived by maximizing pro(cid:133)ts (5) with respect to the quality-adjusted price, subject to the consumer demand relation (1). This setup leads to the standard constant markup pricing rule for the (cid:133)rm, only now the marginal cost term contains a measure of the good(cid:146)s quality. Quality-adjusted price q is expressed in terms of z, w and the (cid:133)rm-speci(cid:133)c productivity parameter: (cid:27) (cid:17)z(!)a 1 q[z(’);’] = w (cid:20)z(!)b 1 + (cid:0) (7) (cid:0) (cid:27) 1 ’(!) (cid:18) (cid:0) (cid:19) (cid:20) (cid:21) Substituting (6) into (7), the quality-adjusted price can be expressed purely as a function of the wage and productivity parameter: b 1 q(’) = (cid:18) q w’a(cid:0) (cid:0) b (8) b 1 a 1 where(cid:18) q = (cid:27) (cid:27) 1 (cid:20) ( ( a 1 (cid:0) 1 b) ) (cid:20) (cid:17) a(cid:0) (cid:0) b +(cid:17) ( ( a 1 (cid:0) 1 b) ) (cid:20) (cid:17) a(cid:0) (cid:0) b is a positive constant. The quality- (cid:26) (cid:0) (cid:18) (cid:0) (cid:0) (cid:19)(cid:27) adjusted pric(cid:0)e dim(cid:1)inis(cid:16)hes in p(cid:17)roductiv(cid:16)ity, sim(cid:17)ilar to the prices in Melitz (2003), which are a

11 CES markup of w=’. Quality-inclusive prices can be reconstituted by multiplying quality-adjusted prices by quality: b p(’) = (cid:18) p w’a (cid:0) b (9) b a where (cid:18) p = (cid:27) (cid:27) 1 (cid:20) ( ( a 1 (cid:0) 1 b) ) (cid:20) (cid:17) a (cid:0) b +(cid:17) ( ( a 1 (cid:0) 1 b) ) (cid:20) (cid:17) a (cid:0) b is a positive constant. (cid:26) (cid:0) (cid:18) (cid:0) (cid:0) (cid:19)(cid:27) (cid:16) (cid:17) (cid:16) (cid:17) (cid:0) (cid:1) Equation (9) is particularly interesting since prices are not constrained to be a negative function of (cid:133)rm productivity. If b > 0, more productive (cid:133)rms choose quality that is a b (cid:0) su¢ ciently high so as to increase price relative to low productivity (cid:133)rms: this is what will be referred to as a quality industry. If b < 0, more productive (cid:133)rms choose higher quality, a b (cid:0) but cost advantages still lead them to set lower prices relative to low productivity (cid:133)rms: this is what will be referred to as a cost industry. 2.1 Distinguishing Cost vs. Quality Industries: U.S. Imports In this section, the observed IPP import price moments are matched with those predicted by the model in order to estimate the parameters underlying the scope of quality di⁄erentiation across products. The intuition behind this identi(cid:133)cation is that the productivity (’) distribution of (cid:133)rms, usually assumed to be skewed right (i.e., with high outliers), maps very di⁄erently into prices depending on the level of sectoral quality di⁄erentiation. As in standard models, in more homogeneous sectors high productivity maps directly into low costs and price, so the distribution is skewed left (i.e., with low outliers). In contrast, one might expect high productivity (cid:133)rms to be more adept at producing quality characteristics, as demonstrated by the equilibrium levels of quality in the model, so in quality di⁄erentiated sectors high productivity maps into higher quality, cost and prices; thus, the price distribution is also skewed right. Switching to discrete notation, we have expressions for quality, quality-adjusted price and quality-inclusive price by item i, HS-10 product j, country c, and month t:

12 1 Quality: z = (cid:18)z ’a b (10) ijct j ij(cid:0)ct b 1 Adjusted Price : q = (cid:18)q w ’a(cid:0)b (11) ijct j jct ij(cid:0)ct b Inclusive Price : p = (cid:18)p w ’a b (12) ijct j jct ij(cid:0)ct where (cid:18)z, (cid:18)q and (cid:18)p are positive product-speci(cid:133)c constants. j j j Taking logs of the quality-inclusive price (12) yields: b lnp = ln(cid:18)p +lnw + j ln’ (13) ijct j jct a b ijct (cid:18) j (cid:0) j(cid:19) As noted above, b is the slope of the productivity-price schedule, which I assume is a b (cid:0) HS10-speci(cid:133)c:forpositivevalues, thecostsof producing(ahighernumberof)qualitycharacteristics outweigh the cost savings of being farther right along the productivity distribution. Since (cid:133)rm-level productivity data are not available across a broad array of products and countries to estimate (13) directly, I proceed by using approximations of the distribution of (cid:133)rm productivity and wage to try to identify the sign and magnitude of b . From (13), the a b (cid:0) second and third moments of the price distribution can also be expressed as the following, derived in the Appendix: b 2 j Var (lnp ) = Var (lnw )+ Var (ln’ ) (14) jt ijct jt jct a b jt ijct (cid:18) j (cid:0) j(cid:19) Skew (lnp ) Skew (lnw ) b 3 Skew (ln’ ) jt ijct jt jct j jt ijct = + (15) [Var jt (lnp ijct )] (cid:0)2 3 [Var jt (lnw jct )] (cid:0) 3 2 (cid:18) a j (cid:0) b j(cid:19) [Var jt (ln’ ijct )] (cid:0) 3 2 Theleft-handsidevariablesof(15)canbemeasuredattheHS10levelusingtheIPPdata,and the distribution of industry-level wage across countries is calculated using annual industry data from the ILO Yearbook of Labor Statistics.18 Since (cid:133)rm productivity measures are not available by product and source country, as a proxy for the (cid:133)rm productivity distribution I exploit another statistic used by the BLS in the construction of its aggregate international prices. Speci(cid:133)cally, the IPP uses (cid:133)rm-level export sales weights to aggregate within HS10 groups for its U.S. export price indexes. To the extent that the size distribution of (cid:133)rm-level 18ILO Yearbook wage data are available at the SIC 4-digit industry level, a courser level of aggregation.

13 sales corresponds to (cid:133)rm productivity, these weights provide a handy approximation of the U.S. export productivity for thousands of disaggregate products, and I apply these weights uniformly to foreign exporters.19 Figure2portraysselectedpercentilesfortheproduct-level skewness of transactionprices, exporter wages and (cid:133)rm export size. In the top-left panel, the import price skewness of the median HS10 product is roughly zero (i.e., prices for that HS10 are symmetrically distributed), with a substantial number of both positive and negative skewness products, This is consistent with the wide range of skewness statistics by sector in Table 1. Wages, on the other hand, shown in the top-left panel, are almost all left skewed, likely re(cid:135)ecting the high incidence of trade among high income countries, with low-wage exceptions. Finally, in accord with prior (cid:133)rm-level studies, U.S. export sales skewness by product shown in the bottom panel is predominantly positive.20 Denoting the dollar export weights y , the estimating equation for (15) is: ijct Skew (lnp ) = (cid:11) +(cid:11) Skew (lnw )+ (cid:11) Skew (lny )+" (16) jt ijkt 0 1 jt jkt 2;hs6 jt ijkt jt (cid:3) hs6 (cid:3) X 3 3 where (cid:11) is the point estimate for Varjt(lnyijkt) 2 bj within a given HS6 cate- 2;hs6 Varjt(lnpjkt) aj bj (cid:26) (cid:0) (cid:27) gory.21 (cid:16) (cid:17) (cid:16) (cid:17) 2.1.1 Variable Markups and (16) What would be the implication for (16) if the (cid:133)rm(cid:146)s competitive environment also factored into its price-setting decision? For instance, if the underlying consumer demand for imports was of the translog functional form instead of CES, then the markup charged by (cid:133)rms of 19NotethatthelevelofU.S.(cid:133)rmsize(asaproxyforproductivity)isnotbeingappliedtoforeignexporters, but rather the skewness of the size distribution. If the true underlying distribution of (cid:133)rm size in the U.S. andabroadisinthepowerlawfamily(suchasaPareto)thenthemeasuredskewnesswillbescaleinvariant; applying these measures to foreign exporters invokes the weaker assumption that the shape of the U.S. (cid:133)rm size distribution is the same as that in the rest of the world. 20Thehighnumberofproductswithclosetosymmetricdistributionsmayre(cid:135)ectthesmallnumberof(cid:133)rms that the BLS samples within certain HS10 product groups. Figure 2 and the regression estimates below are not substantially di⁄erent if the minimum number of (cid:133)rms per product included is increased. 21Technically it is possible to estimate the quality cost parameters at the classi(cid:133)cation/HS10 level (as opposed to HS6). However, since the BLS classi(cid:133)cation and HS10 codes are identical at the HS6 level, for ease of interpretation I use the slightly more aggregate codes.

14 di⁄ering productivity would no longer be the same proportion of marginal cost. As a result, the pricing equation (9) would be di⁄erent and observed skewness of prices would re(cid:135)ect variable markups in addition to quality di⁄erences. For the identi(cid:133)cation of the quality scope, this is a problem if the price distribution interacts with variable markups distinctly in low versus high scope industries. Consider the relationship between markups and productivity when consumers allocate consumptionacrossvarietiesaccordingtothetranslogexpenditurefunction. Inthatsetting, moreproductive(cid:133)rmsarenotonlylarger,withhighermarketshare, theyalsochargeahigher percentage markup over their marginal cost. In an industry with high scope for quality di⁄erentiation, indeed the high-priced, high-quality varieties also have higher markups; part of the observed price skewness in the industry is potentially driven by markups. On the otherhand, inanindustrywithlowscope, theproductive(cid:133)rmswithrelativelyhighshareand high markups have relatively low prices. In those industries, the observed price skewness is attenuated by variable markups and is less negative as a result. In sum, variable markups of the type described introduce an upward level-shift in measured skewness. The fact that skewness is higher in both high and low scope industries suggests that the bias in (cid:11) due to 2 variable markups is of the second-order: it depends not on the degree of markups but on the di⁄erence of that degree across industries. 2.1.2 Results The OLS regression of (16) is pooled across all product categories for which price, wage and sizedataareavailable:41,633product-timepairobservations. Alsoincludedareyeardummy variables, to control for secular trends in the respective distributions.22 This speci(cid:133)cation yields estimates for approximately 1,100 HS6 categories, approximately 450 of which are statistically distinguishable from zero.23 The resulting estimates for the largest products by volume in cost industries b < 0 and quality industries b > 0 are shown in Table 2. a b a b (cid:0) (cid:0) 22Seasonality does not appea (cid:0) r to be a (cid:1) very important driver of varia (cid:0) tion in th (cid:1) e import price distribution. Despite monthly import sales being quite volatile due to lumpiness, prices from month-to-month are quite rigid. Figure 1 illusrates how a change in the average price for an HS10 is not necessarily driven by a corresponding change in the price distribution. Moreover, IPP imputation techniques tend to reinforce the stability of the distribution at higher frequencies. 23Robust standard errors, clustering HS10 estimates within HS6 groups, barely alter the number of precisely estimated scope measures.

15 At (cid:133)rst glance, the quality industries in the top panel seem to conform to our prior notions of products with a high degree of quality di⁄erentiation (e.g., passengers cars, clothing and wine). Likewise, the bottom panel includes goods with a (cid:145)cost story(cid:146)such as machine parts and accessories, and metal furniture. On the other hand, both panels contain products that are not so intuitively categorized, such as transport motor vehicles and parlour games in the cost panel, and lique(cid:133)ed butane in the quality panel. There are several reasons not to get bogged down in constructing stories to explain the level of quality di⁄erentiation across product groups. First, category de(cid:133)nitions at the HS6 or HS10 level are somewhat arbitrary. For example, lique(cid:133)ed butane(cid:146)s largest HS category encompasses grades of the hydrocarbon between zero and 80 percent purity, which by de(cid:133)nition is highly di⁄erentiated. Alternatively, if a new categorization scheme emerges including a separate group for a particular laptop brand with 2GB SD RAM and 80GB hard disk drive, we would not observe much quality di⁄erentiation within that product despite the fact that the broader product class has a lot of quality heterogeneity. What is important is that there is a consistent way to characterize what is going on within each product group, however de(cid:133)ned. Second, the estimates should be interpreted in the context of U.S. import demand. Anarrowscopeforqualitydi⁄erentiationintransportvehiclesre(cid:135)ectsthefactthat the U.S. imports a narrow range of these goods; therefore the measure should be interpreted as the scope of quality within an HS6 category conditional on an international transaction taking place. Third, we must be careful to discern between horizontal and vertical (quality) di⁄erentiation, where the former is di⁄erentiation that occurs across product features that cannot be ordered. For example it would be di¢ cult to place ice cream as either a cost or a quality industry; while there is alot of horizontal di⁄erentiationin(cid:135)avors, it is not clearwhetherthere is a broad array of quality di⁄erences among similar (cid:135)avors coming from di⁄erent producers. To check the extent that the measured quality scope corresponds to product-level horizontal di⁄erentiation, I compare b with previous estimates by Broda and Weinstein (2006) of a b (cid:0) intra-product elasticity of substitution ((cid:27)). In the CES framework above, (cid:27) indexes consumers(cid:146)willingness to substitute among quality-adjusted varieties, so in this context it can be strictly interpreted as an index of horizontal di⁄erentiation. In Table 2, there is no clear pattern between b and (cid:27), and on average a slight negative relationship (i.e., cost industries a b (cid:0) have a low (cid:27) while quality industries have a high (cid:27)). Taking the model of vertical di⁄er-

16 entiation very seriously, one might expect producers in a horizontally di⁄erentiated sector (low (cid:27)) to not need to distinguish themselves as much vertically, however I (cid:133)nd that over all estimates the correlation between b and (cid:27), albeit measured with a large degree of error a b (cid:0) from both sets of parameter estimates, is only 0.01. A more subtle point is that the sign of (cid:11) in (16) could re(cid:135)ect the magnitude of 2;hs6 the skewness of productivity rather than its sign. That is, considering a product where both prices and (cid:133)rm export size are positively skewed, a negative sign could still arise for (cid:11) if the prices are simply less positively skewed relative to other product groups of equal 2;hs6 productivity skewness. Thus positive skew prices can still be cost industries, which is less intuitive. As an indirect way of decomposing how much of the scope estimates are due to sign versus magnitude, the estimates are rerun for only the set of products with both positively skewed prices and productivity.24 I (cid:133)nd that even though there are still quite a number of signi(cid:133)cant negative estimates for (cid:11) , the average magnitude of the estimates 2;hs6 changessubstantially. Forproductswithasigni(cid:133)cantnegativeestimate(indicatingthatthey are a cost industry) in the unrestricted sample, the average size of (cid:11) is -1.70, compared 2;hs6 to 22.30 for the identical set of products in the restricted sample. The estimates for quality industries remain roughly unchanged at 4.82 in the unrestricted sample versus 3.42 in the restricted sample. This suggests that the sign of price skewness plays a very important role in determining cost versus quality industries. Finally, applying the quality scope estimates, the share of U.S. import trade accounted for by cost and quality industries is illustrated in Figure 3. In panel (a), approximately 50-60 percent of total import value is categorized as either cost or quality, with the asterisk superscript denoting quality scope estimates that are signi(cid:133)cantly di⁄erent from zero.25 Between 1993 and 2006, the share of cost industries declined from 40 percent to about 25 24For a fair comparison, the estimates are recomputed for the whole sample with wage skewness on the left-hand side (i.e., Skew (lnp ) Skew (lnw )) and then compared to the restricted sample. jt ijct jt jct (cid:0) 25The interpretation of sales value sums across cost and quality products may be dubious due to the reliance of the estimates on product category de(cid:133)nitions, as discussed above. For instance, if certain highly di⁄erentiated technology industries have very narrowly de(cid:133)ned HS10 categories and also large sales, then too large a weight would be assigned to low scope industries despite the underlying quality heterogeneity of the industry. The assumption necessary to believe the sales sums (or any cross-product analysis for that matter) is that the agency determining the breadth of product categories de(cid:133)nes them (cid:145)correctly,(cid:146)where a correct product de(cid:133)nition exactly matches the consumers(cid:146)notion of the variety characteristics within a product. e.g.,IfconsumersagreethatPDA(cid:146)saredistinctproductsfromcellulartelephones,thentheagency would be correct in de(cid:133)ning a new product code. Given this limitation, the applications in this paper will focus on the characteristics of varieties within products.

17 percent with quality industries growing from 20 percent to about 25 percent. Thus, the current proportion of imports with a large scope for quality di⁄erentiation is roughly 50 percent. Within a balanced panel of products in panel (b), the shares of cost and quality imports are more stable at 50 percent.26 In the sections that follow, the distinction between cost and quality products will be applied to further explore the cross-section and dynamic features of the import price distribution. 2.2 Distinguishing Cost vs. Quality Industries: U.S. Exports Given the availability of U.S. export transaction prices, the quality ladder measures for imports in the preceding section can be compared with those for exports. The use of export prices leads to several simpli(cid:133)cations in the empirical implementation. First, the measure of the (cid:133)rm size skewness for U.S. exporters has a direct measure. Second, given a single source country, the wage skewness term from (16) drops out, leaving the following regression of HS10 export price skewness on the HS10 (cid:133)rm size distribution: Skew (lnp ) = (cid:11) + (cid:11) Skew (lny )+" (17) jt ijct 0 1;hs6 jt ijct jt hs6 (cid:3) X As discussed in the appendix, the interpretation of (cid:11) in this speci(cid:133)cation does not con- 1 tain information on the magnitude of b but, rather, only its sign. Running (17) across a b (cid:0) the array of U.S. exports yields estimates for b in 439 HS6 industries, 324 of which are a b (cid:0) distinguishable from zero as either quality or cost industries. Table 3 displays the largest quality and cost products by sales volume. Again, with a few exceptions, we see a more or less intuitive categorization scheme, with semiconductors and automobiles among the most quality di⁄erentiated and more commoditized goods such as parts and accessories among the least quality di⁄erentiated. Based on industries where estimates are available, representing $422 billion dollars (or 53 percent) of U.S. exports in 2005, quality industries account for approximately 48 percent of trade volume. It is not obvious a priori what the relationship between import and export quality scope 26Byonlyconsideringproductstradedthroughoutthesampletimeframe,panel(b)ignoresthecreationof newcategories,wherealotoftheactionmaybeintermsofqualitycomposition,aswellassomeratherlarge reclassi(cid:133)cations in product codes over the period. What we can take away from this selection of products is the absence of large shifts in the intensive margin towards products with long quality ladders.

18 measures should be. First, product level specialization is re(cid:135)ected in the relatively limited overlap of very large import and export HS6 categories. Of the 1,098 import and 439 export scope estimates, only 215 match. Of those, intra-product specialization may, in theory, lead to vastly di⁄erent scope estimates. In the U.S. data, the sign of the import and export quality scope measures (i.e., the sign of the price-productivity schedule) correlate positively and signi(cid:133)cantly, with a logit regression coe¢ cient of 0.68 (0.27) between sign dummies for (cid:11) in (16) and (17). hs6 3 Quality Specialization by Country In the previous section, the correlation of price and productivity skewness identi(cid:133)ed the scope for quality di⁄erentiation of products. In this section, quality scope measures and import price levels are employed to rank exporting countries according to their productivity and quality levels. Relative quality is identi(cid:133)ed by picking a product and observing which countries inhabit the tails of its price distribution. Countries consistently selling in a quality industry(cid:146)s right tail are considered to be specialized in quality, with the converse holding for cost industries. In this section, quantile regression techniques are used to discern countries(cid:146) propensity to sell in the tails of the price distribution and, indeed, similar sets of countries populate the tails; a country exporting at a high price in a quality industry tends to export at a low price in a more homogeneous cost industry. In other words, the price distribution and the underlying productivity-price mapping reveals the productivity of exporting countries; given a monotone link between productivity and quality, relative quality levels are also identi(cid:133)ed. Previous studies (cid:133)nd that average prices vary systematically with country characteristics such as wealth and distance, and so this exercise can be interpreted as additionally documenting inter-product heterogeneity in country pricing patterns consistent with the model of quality choice. Onceagainthe(cid:133)rstorderconditionsofthe(cid:133)rmproblemo⁄eraconvenientstartingpoint. The item pricing equation (13) is a simple relationship between quality-inclusive price, wage and (cid:133)rm productivity. Controlling for wage, the residual of item price contains information about the exporting (cid:133)rm(cid:146)s location within the productivity distribution as well as any other country- or (cid:133)rm-speci(cid:133)c factor. Conditioning by country (as opposed to (cid:133)rm productivity

19 in the previous section) allows for analysis in levels: lnp = (cid:11) +(cid:11) lnw + (cid:11) d + (cid:11) d + (cid:11) d +" (18) ijct 0 1 jct 2c c 3j j 4t t ijct c j t X X X where (cid:11) estimates a country-speci(cid:133)c relative price, controlling for product composition 2c with HS10 dummy variables d and for time variation with year dummy variables d . Given j t the observation above of asymmetric price distributions, and the implication that quality levels di⁄er substantially across the spectrum of observations within a product, least squares estimates of the conditional mean country elasticities confound e⁄ects in the tightly clustered body of the price distribution with those in the more disperse tail. To additionally condition on location within the distribution, we separate positively and negatively skewed classi(cid:133)cation groups and then estimate quantile regressions at the 15th and 85th quantiles of each set.27 Each regression pools across products within a sector, so each sector has four estimates for each active country: the 15th quantile of the positively skewed products (i.e. the (cid:145)body(cid:146)of the distribution), the 85th quantile of the positively skewed products (i.e. the (cid:145)tail(cid:146)of the distribution), the 15th quantile (tail) of the negatively skewed products and the 85th quantile (body) of the negatively skewed products. Again, using ILO Yearbook data to construct country-product-year wage measures, monthly transaction prices are regressed on wage and (cid:133)xed e⁄ects within each of seven sectors. The country coe¢ cients can be interpreted as relative to Canada, the omitted country dummy. Table 4(a) displays the wage coe¢ cients of (18) by sector. Recall from (13) that the simple model predicts a wage elasticity of one. While not always precisely estimated, all but one of the signi(cid:133)cant coe¢ cients (denoted by an asterisk for p<.1) are positive, some with coe¢ cients quite close to one. Each sectoral regression produces four lists of country relative prices. Comparing these prices for the same country across products of di⁄erent quality scope reveals an interesting pattern. As an illustration, Table 5 presents the results for the textiles sector.28 To begin, consider the (cid:133)tted regression lines running through the 27The quantile regression does not function well for large numbers of righthand side variables nor for sparselypopulatedcells. Tocopewiththislimitation, Idividethesampleintothe16sectorslistedinTable 1 (for notational convenience, I suppress sector subscript in (18)). Then, each regression uses only the 5 largest classi(cid:133)cation groups (by observations) by sector-skewness and only includes countries with at least 75 observations over the course of the sample. Running the quantile regression command in Stata on the resulting restricted samples, convergence of the algorithm is achieved in 7 sectors (listed below). 28Note that since the products of di⁄erent skewness are produced by a di⁄erent mix of countries, the coe¢ cients shown are those for which a country produces products in industries with both positively and

20 tails of the price distribution: the 85th quantile (high prices) of the quality products and the 15th quantile (low prices) of the cost products. We observe that the relatively high prices in the quality products correspond with relatively lower prices in the cost products; Pakistan and Bangladesh are farther out in the tails while Korea, Macao and Hong Kong reside closer to the dense cluster of prices in the body. A theory of quality sorting like the one above suggests that (cid:133)rms in Pakistan and Bangladesh are more productive in textiles exports for the products included in (18), while those in East Asian countries are less so. In contrast, there is not a clearly discernible pattern, or perhaps even the opposite pattern, for the regression lines (cid:133)tted through the body of the price distribution (i.e., 15th quantile of quality products and 85th quantile of cost products) where a producer like Korea tends to have relatively low prices and a producer like Turkey tends to have relatively high prices in both types of product. Figure 4 illustrates the (cid:145)tail(cid:146)prices across all sectors,29 where each point in the scatter plot is a pair of relative price estimates for a particular country-sector combination, based on quantile regressions of the 85th quantile of the quality products and 15th quantile of the cost products (i.e., including estimates for textiles in columns 2 and 3 of Table 5). On the horizontal axis is a given country(cid:146)s relative price for its quality products in a given sector, and on the vertical axis is that country(cid:146)s relative price in cost industries for that sector. Across all sectors, the pattern in textiles is preserved, with the majority of estimates falling in the top-left or bottom-right quadrants of the grid, tracing out a downward linear trend. On one end of the spectrum, UK plastics exports have relatively high prices in quality products and relatively low prices in cost products. The quality choice theory would suggest that this re(cid:135)ects UK (cid:133)rms(cid:146)position on the tail of the productivity distribution. On the other end of the spectrum, Mexican metals exports are low-priced in quality products and high-priced in cost products, denoting lower average productivity. The high incidence of estimates in the top-left and bottom-right quadrants indicates that such sign-switching in relative prices across products occurs frequently, and the general pattern across countries and sectors is a negative relationship, with a one percent higher relative price in quality industries corresponding to a 0.28 percent lower relative price in cost industries. negativelyskewedprices. Notshownarethosecountriesinpositivelyskewedindustriesonlyorinnegatively skewed industries only. 29The (cid:133)gure illustrates the estimates for the seven sectors listed in Table 4(a).

21 Figure 5 shows the analogous illustration for the body of the price distribution. In contrast to the tail estimates, the resulting price mapping across cost and quality industries actually has a positive upward trend, with a one percent higher relative price in quality industries corresponding to a 0.45 percent higher relative price in cost industries. The incidence of relative prices of like sign is much higher than in the tail of the distribution, as illustrated by the greater number of estimates in the top-right and bottom-left quadrants. Assuming (cid:133)rms in a country-sector cell have similar productivity, the pattern in Figure 4 is consistent with the price-setting behavior outlined in the model of quality choice; productivity maps into price inversely for quality and cost industries. How can we then explain the pattern in Figure 5? One possibility is that it is simply harder to precisely measure a negative relationship between relative prices across products with di⁄erent skewness in the body of the price distribution versus the tail. This, in turn, may suggest that it is bene(cid:133)cial to use transactions prices instead of unit value average prices for this exercise, where unit values confound the within-distribution pricing behavior of exporters. Alternatively, (cid:133)rm sorting in prices might not be as strong for (cid:133)rms that are not productivity outliers. We also cannot neglect the possibility that (cid:133)rms in a country-sector do not have similar productivity. For example, if Japan has high productivity in the production of engines (positive skewness) and low productivity in the production of o¢ ce machine parts (negative skewness), then we would observe a positive relationship among Japanese export prices in the machinery sector. However, if that were the case, we would also expect that pattern to be generated by the same sector-countries in the tail of the price distribution, which we do not observe. The robustness of this pattern is checked by conditioning on additional features of the price observations. The IPP collects an array of characteristics for each price in its sample, includingwhetheritisamarketortransferprice, whatitsunitof measureisandwhenitwas discontinued.30 Additional dummy variables are added to (18) for market-based transfer price, cost-based transfer price, the various units of measure and irregular discontinuation. The last variable denotes an item that was discontinued due to reasons other than regular sample rotation, the price of which may re(cid:135)ect the idiosyncrasies of an item at the end of its life cycle. The control estimates are shown in Table 4(b) for sectors with enough observations. As above, the coe¢ cient on wage is generally positive. The estimates for 30There is also an identi(cid:133)er for whether the item price was imputed for the purpose of index construction. Omitting the imputations does not a⁄ect the price distribution statistics systematically.

22 transfer prices do not show a clear pattern, though it is interesting to note that (cid:145)marketbased(cid:146)transfer prices are often signi(cid:133)cantly di⁄erent from prices at arms-length. For the most part, and particularly for quality industries, items discontinued irregularly had lower prices than those which were not discontinued over the course of the sample. Based on the ordering of (cid:133)rms by quality level, this implies that it is the low quality varieties that tend to exit the export market. For cost industries, the relationship is somewhat ambiguous with the exiting varieties having either lower or higher prices. The resulting country estimates for the distribution tail and body are shown in Figure 6 and Figure 7, respectively. Similar toFigure 4, Figure 6illustrates a negative relationship betweenthe tail relative prices incost and quality industries, with a one percent higher price in quality industries corresponding to a 0.47 percent lower price in cost industries. In Figure 7, on the other hand, the positive relationship between the body relative prices in Figure 5 disappears. Insum, these relative price patterns suggest that the nature of specializationis particular to where producers reside in the exporter productivity distribution. The (cid:133)rms in the tail of the distribution specialize in high quality products in quality industries and low price productsincostindustries. Thisobservationisbroadlyconsistentwithsectoralcomparative advantage translating into price distinctly, depending on the scope for quality di⁄erentiation of the export. As such, it is both supportive of the theory of quality sorting and revealing of the heterogeneous pattern of specialization in quality and cost industries. 4 A Time Series Test of Quality Sorting Since the composition of quality characteristics is conditional on an export transaction taking place, patterns in composition will depend on any factor that a⁄ect the (cid:133)rm(cid:146)s export participation decision. In this section, patterns in quality composition are predicted and measured in response to real exchange rate changes. In the model, (cid:133)rms endogenously sort into foreign export markets if they are above some threshold productivity (and hence pro(cid:133)t) level.31 It is the (cid:133)rms on the extensive margin of (cid:133)rm entry, at the low end of the 31In the description above, the model is not closed to solve for the endogenously determined productivity cut-o⁄,thoughtheexactlocationofthemarginal(cid:133)rmwillnotmatterfortheresultsobtainedinthissection. It is only important to know that there is some well-de(cid:133)ned, unique equilibrium in which the (cid:133)rm on the margin of entry into the international market earns zero pro(cid:133)ts.

23 productivity distribution of active (cid:133)rms, whose survival in the market is predicated upon any change that a⁄ects the location of the threshold (cid:133)rm. For the measurement of average prices in the wake of such a change, key considerations are: who are the marginal (cid:133)rms, and do they have relatively high or low prices? In a cost industry, the marginal, least productive, (cid:133)rms have the highest price and so (cid:133)rm entry will put upward pressure on the average price of remaining (cid:133)rms. In a quality industry, the marginal, least productive, (cid:133)rms have the lowest price, and so entry will put downward pressure on the average price. I test these predictions across a wide array of disaggregate products by comparing two types of import price index: one that controls completely for import composition, a (cid:145)constant-quality(cid:146)index, and a second that allows for changes in the extensive margin, a (cid:145)quality-inclusive(cid:146)index. Exchange rate pass-through is incomplete in this simple framework due to (cid:133)rm entry and exit. With CES preferences, changes in costs pass through on a 1-for-1 basis into prices at the (cid:133)rm level; thus pass-through is complete for each individual variety. To be concrete, an adverse, exogenous change in (cid:133)rm marginal cost is re(cid:135)ected as a proportional increase in price. Accounting for the changing mass of (cid:133)rms, however, average prices re(cid:135)ect the changing composition of the import bundle. For example, a real appreciation in the foreign country causes the least productive (cid:133)rms in each industry to drop out of the market, such that in a quality industry the lowest price (cid:133)rms exit and the average price of the remaining (cid:133)rms increases. These extensive margin e⁄ects are above and beyondany other price change the (cid:133)rm will undertake in response to the shock. In a cost industry, the least productive, highest price (cid:133)rms exit the market and the average price of the remaining (cid:133)rms decreases. These compositional e⁄ects will be manifested in a quality-inclusive price index, which is constructed from unit values average prices. In contrast, aggregating transaction-level IPP pricesforexactlythesameitemsfromperiodtoperiodallowsfortheconstructionofanindex that is bereft of quality changes by de(cid:133)nition, a constant-quality index. Since the extensive margin is shut down and aggregation weights are (cid:133)xed in each period in the constant-quality index, the changes in average prices due to composition, as are observed in unit values, are absent. Thus, the sharpest predictions of the model due to quality sorting are that: (i) in quality industries, (cid:133)rms will pass through exchange rate shocks to quality-inclusive prices by more than to constant-quality prices, and (ii) in cost industries, (cid:133)rms will pass through exchange rate shocks to quality-inclusive prices by less than to constant-quality prices. ToexplorethesepredictionsfortheU.S.,Irunpass-throughregressionsforbothconstant-

24 quality prices and quality-inclusive prices of the form: 2 (cid:0) lnP = (cid:11) + (cid:11) lnRER +(cid:11) X +" (19) jt 0 1 jt 3 jt jt t=0 X where lnP is the annual import price index level for each classi(cid:133)cation group, j, in year jt t. For the measure of quality-inclusive price, I construct Tornqvist indices of annual unit value price changes using HS10-country export quantity (x = x ), sales (r = jkt i ijkt jkt p x ) and trade weights (w = r ) data provided by Feenstra et al. (2002):32 i ijkt ijkt jkt jkt P P rjkt lnPUV = w ln xjkt (20) j;t 1;t jkt rjk;t 1 (cid:0) X k xjk;t (cid:0) (cid:0) 1 ! For the measure of constant-quality price, I aggregate the Tornqvist indices directly from the individual variety prices (p ) in the IPP sample, using annual country weights to ijkt approximate item sales weights (w w ): jkt ijkt (cid:25) p lnPIPP = w ln ijkt (21) j;t 1;t jkt p (cid:0) X k X i (cid:18) ijk;t (cid:0) 1(cid:19) Cumulating (20) and (21) yields the cumulative import index levels for quality-inclusive (lnPUV)andconstant-quality(lnPIPP)importprices, respectively, byclassi(cid:133)cationgroup.33 jt jt Additionally, I compute an annual index of the real exchange rate by classi(cid:133)cation group using IFS country data for real exchange rates, and aggregate using the Tornqvist formula and import sales weights, as above. Included in (19) are current and two lags of the exchange rate index. Finally, I include a vector of controls ( X ) containing an annual jkt index of competing export prices (published by the BLS at the HS4 level), an index of tari⁄s, a measure of the Chinese import share in each classi(cid:133)cation group and a full set of classi(cid:133)cation group (cid:133)xed e⁄ects. Since the import price indexes are both (cid:145)at-the-dock(cid:146)(i.e., 32In cases where the classi(cid:133)cation group is more aggregate than HS10, country-HS10 unit value changes are calculated and then aggregated to the classi(cid:133)cation level (i.e., as opposed to (cid:133)rst summing across values and potentially non-comparable quantities in order to calculate classi(cid:133)cation level unit values). 33Since the IPP data frequency if monthly, once the monthly cumulative index is constructed at the classi(cid:133)cation level, annual averages are taken to make this series comparable to the lower frequency unit value index.

25 net of tari⁄), we expect either a zero or negative coe¢ cient on the tari⁄control.34 Bergin and Feenstra (2007) demonstrate that increasing export competition by countries with (cid:133)xed exchangeratesmaybelesseningexchangeratepass-throughin(cid:135)exibleexchangerateexporter prices, so I include the share of Chinese and Hong Kong exports in each classi(cid:133)cation group as a proxy for (cid:133)xed exchange rate export competition. Table 6 shows the results for the (cid:133)xed e⁄ects OLS regression of (19), run annually (1994- 2004) over all classi(cid:133)cation groups for which data are available.35 Additionally, the data are split by cost and quality groups based on the estimates in the method of moments exercise above.36 The results are supportive of: (i) the existence of quality sorting in the data, and (ii) the notion that quality di⁄erentiated products have higher pass-through. For sorting, the theory suggests that in cost industries quality-inclusive prices pass through currency appreciation by less than constant-quality prices. The results are consistent with this prediction, as pass-through of the real exchange rate is positive and signi(cid:133)cant at approximately 7percent37 incolumns(I)and(II),andindistinguishablefromzeroinquality-inclusiveprices in columns (III) and (IV). Though suggestive, large standard errors make it impossible to distinguish between the two coe¢ cients. The theory also suggests that in quality industries quality-inclusive prices pass through currency appreciation by more than constant-quality prices, re(cid:135)ecting the exit of the lowpriced marginal (cid:133)rms in the wake of a real exchange rate appreciation. The results in 34Theoretically the coe¢ cient should be zero for small importers and negative for large importers, so it will depend on the size of the US in each particular product market. 35Since the IPP data is a sample and the data does span the entire set of HS10 product-country groups, unitvalueswereonlycomputedforthesampledproduct-countrygroupsintheconstant-qualitypriceindexes. This ensures that the comparison of HS10 indexes is not itself contaminated by di⁄erent underlying HS10country product composition. 36This distinction is based solely on the sign of the estimated scope for quality di⁄erentiation. Cost industries are those with b <0 while quality industries are those with b >0. Using only the subset of a b a b scope estimates that are sig(cid:0)ni(cid:133)cantly di⁄erent than zero does not a⁄ect th(cid:0)e results. 37In a broad set of empirical studies, estimated average pass-through of nominal exchange rates to U.S. importpricesisapproximately0.5inthe1980(cid:146)s,decliningtoapproximately0.2inthe1990(cid:146)s(amongothers, see Goldberg and Knetter (1997), Olivei (2002), Marazzi et al. (2005), Gust et al. (2006), and Bergin and Feenstra (2007)). One explanation for the measured decline in average pass-through, posited by Campa and Goldberg (2005), attributes much of the decline to the changing composition of import bundles, from sectors with relatively high pass-through such as energy to sectors with relatively low pass-through such as manufactures. Following this line of reasoning a step further, sectoral pass-through itself is just the average elasticityacrossproductswithdisparatescopefordi⁄erentiation,re(cid:135)ectingtheunderlyingmicrofoundations of the (cid:133)rms choice of product characteristics. The seemingly low pass-through estimates here likely re(cid:135)ect the low annual frequency of the data relative to other studies. However, the results are not vastly di⁄erent from pass-through estimates in the 1990(cid:146)s of 20 percent.

26 columns (IX)-(XII) strongly support this hypothesis with estimates of pass through more than doubling in quality-inclusive prices relative to constant-quality prices, with estimates statistically distinguishable from one another at the 5 percent level. Finally, analyzing both cost and quality products jointly in columns (V)-(VIII), we see that overall qualityinclusive prices have higher pass-through than the corresponding constant-quality indexes (alsosigni(cid:133)cantlydi⁄erentfromoneanother). Thissuggests thatU.S. importprices behave, on average, like a quality industry. The second interesting pattern in Table 6 is that, within import price measures, quality industries have higher pass-through coe¢ cients than cost industries. This is suggested by the monotone increase in estimates from left to right in the sets of columns: {(I),(V),(IX)}, {(II),(VI),(X)}, etc. In quality-inclusive prices, the degree of pass-through in quality industries is signi(cid:133)cantly greater than that in cost industries. One might expect a more pronounced impact of quality di⁄erentiation on price dynamics in sectors with longer quality ladders and a greater diversity of products, such as mechanical devices and electric machinery: HS84 and HS85. These sectors compose approximately one third of the classi(cid:133)cation groups for which quality scope estimates exist and (due to their sheer size of import value) tend to have more products sampled by the IPP per classi(cid:133)cation group than others, augmenting con(cid:133)dence in the measures of sample skewness and quality scope. Indeed, the left panel in Table 7 presents stronger support of quality sorting than Table 6: in cost industries, constant-quality prices are 11 percent while quality inclusive prices are approximately zero and in quality industries constant quality prices are 6 percent compared to 43 percent in quality-inclusive industries. Finally, intherightpanelofTable7, weapplythecost/qualityclassi(cid:133)cationderivedfrom U.S. exports in order to expand the range of products for which pass-through elasticities can be computed. Above, the quality range of U.S. imports as measured by the sign of b , a b (cid:0) was shown to be positively correlated with the sign of the quality range for U.S. exports. As such, I use the export classi(cid:133)cation as an additional gauge of the quality di⁄erentiation of certain import products. The pattern of pass-through coe¢ cients is similar to to the estimates using the import-based classi(cid:133)cation: the pass-through estimate is insigni(cid:133)cant in the cost industry quality inclusive index and very high in the quality industries. Comparing pass-through in (cid:145)matched model(cid:146)prices (in which the composition of varieties is held constant) to unit values is not in itself novel. Alterman (1991) argues that composi-

27 tional e⁄ects have a large e⁄ect on pass-through estimation due to the unit value(cid:146)s imperfect measurement of price. Here, these large di⁄erence are shown to be systematic in the way that (cid:133)rms sort into export markets in quality and cost industries. 5 Conclusions It has long been recognized that average prices are imperfect measures of both price and quality. This paper takes a step forward in overcoming the di¢ culties of quality inference fromaveragepricesbydisentanglingtheirrelationshiptransactionbytransaction. Exploring the distribution of transaction prices within narrow product groups introduces a new dimension to our understanding of (cid:133)rm pricing behavior and adds texture to our observations of average trade patterns. In particular, the higher moments in prices reveal not only the scope for di⁄erentiation across products, but the countries that specialize in quality and the dynamics of pricing due to compositional changes. I (cid:133)nd evidence that an average relation between output price and exporter capability nets out highly disparate patterns of specialization: productive exporters simultaneously set high prices in quality di⁄erentiated industries and low prices in more homogeneous industries. This relationship between productivity, quality and price also emerges in the time series of average prices. Index number techniques identify (cid:133)rm entry and exit along the extensive margin after a change in the real exchange rate and the relative quality of marginal (cid:133)rms. The exit of low productivity, low quality, low price (cid:133)rms in a quality di⁄erentiated industry pushes up the average price of the surviving (cid:133)rms. Both sets of results bolster the positive relationship between productivity and quality typically asserted in models of quality choice. Finally, it is important to view the results herein in context. Perhaps the future of trade price data promises a global census of transaction quantities and prices and perfect information about product speci(cid:133)cations. In such a world, it would be feasible and straightforward to estimate the marginal value of product characteristics and there would be no need to infer quality from composition-contaminated average prices. In the meantime, this paper presents an intermediate rung on the data quality ladder.

28 References [1] Alterman, William (1991) (cid:147)Price Trends in U.S. Trade: New Data, New Insights,(cid:148) in Peter Hooper and J. David Richardson, eds. International Economic Transactions. Studies in Income and Wealth, no. 55. Chicago: Univ. of Chicago Press and NBER, 109-139. [2] Auer, Raphael and Thomas Chaney (2009): "Exchange Rate Pass-Through in a Competitive Model of Pricing-to-Market," Journal of Money, Credit and Banking, vol. 41(s1): 151-175. [3] Aw, Bee Yan & Roberts, Mark J. (1986): "Measuring Quality Change in Quota- Constrained Import Markets: The Case of U.S. Footwear," Journal of International Economics, Elsevier, vol. 21(1-2), pages 45-60, August. [4] Baldwin, Richard E. and James Harrigan (2007): "Zeros, Quality and Space: Trade Theory and Trade Evidence," NBER Working Papers 13214, National Bureau of Economic Research, Inc. [5] Baldwin, Richard E. and Tadashi Ito (2008): "Quality competition versus price competition goods: An empirical classi(cid:133)cation," HEID Working Paper No7/2008. [6] Bergin, Paul R. and Robert C. Feenstra (2007): "Pass-through of Exchange Rates and Competition Between Floaters and Fixers," NBER Working Papers 13620, National Bureau of Economic Research. [7] Bernard, Andrew B. and J. Bradford Jensen (1999): "Exceptional Exporter Performance: Cause, E⁄ect, or Both?," Journal of International Economics, 47(1), 1-25. [8] Bils, Mark (2004): "Measuring the Growth from Better and Better Goods." NBER Working Papers 10606, National Bureau of Economic Research. [9] Bureau of Labor Statistics (2009): BLS Handbook of Methods, Chapter 15. [10] Boskin M.J., E.R. Dulberger, R.J. Gordon, Z. Griliches and D.W. Jorgenson (1997): "The CPI Commission: Findings and Recommendations" American Economic Review, Vol. 87, No. 2, Papers and Proceedings of the Hundred and Fourth Annual Meeting of the American Economic Association (May), pp. 78-83.

29 [11] Broda, Christian. & David E. Weinstein (2006): "Globalization and the Gains from Variety," Quarterly Journal of Economics, MIT Press, vol. 121(2), pages 541-585, May. [12] Campa, JosØ Manuel and Linda S. Goldberg (2005): "Exchange Rate Pass-Through into Import Prices," Review of Economics and Statistics, 87(4): 679-690. [13] Choi, Yo Chul, David Hummels and Chong Xiang (2006): "Explaining Import Variety and Quality: The Role of the Income Distribution," NBER Working Papers 12531, National Bureau of Economic Research. [14] Crozet, M., K. Head and T. Mayer (2009): "Quality sorting and trade: Firm-level evidence for French wine" CEPR Discussion Paper 7295. [15] Feenstra, RobertC.(1996):"U.S.Imports, 1972-1994: DataandConcordances,"NBER Working Paper 5515. [16] Feenstra, Robert C. (1988): "Quality Change under Trade Restraints in Japanese Autos," Quarterly Journal of Economics, 103(1): 131-46. [17] Feenstra, Robert C., Romalis, John and Peter K. Schott (2002): "U.S. Imports, Exports and Tari⁄Data, 1989-2001," NBER Working Paper 9387, National Bureau of Economic Research. [18] Goldberg, Pinelope and Michael Knetter (1997): (cid:147)Goods Prices and Exchange Rates: What Have We Learned?(cid:148)Journal of Economic Literature, 35, 1243-1272. [19] Gopinath, Gita and Roberto Rigobon (2008): "Sticky Borders," Quarterly Journal of Economics, 123(2): 531-575. [20] Gopinath, Gita, Oleg Itskhoki and Roberto Rigobon (2007): "Currency Choice and Exchange Rate Pass-Through," manuscript. [21] Gust, Christopher, Sylvain Leduc and Robert J. Vigfusson (2006): (cid:147)Trade Integration, Competition, and the Decline in Exchange-Rate Pass-Through,(cid:148)Board of Governors of the Federal Reserve System International Finance Discussion Paper #864. [22] Hallak, Juan Carlos (2006): "Product Quality and the Direction of Trade," Journal of International Economics, 68, 238-265.

30 [23] Hallak, JuanCarlosandPeterK.Schott(2009): "EstimatingCross-CountryDi⁄erences in Product Quality," Working paper. [24] Hallak, Juan Carlos and Jagadeesh Sivadasan (2008): (cid:147)Productivity, Quality and Exporting Behavior Under Minimum Quality Requirements(cid:148), mimeo. [25] Harrigan, James and Geo⁄rey Barrows (2006): "Testing the Theory of Trade Policy: Evidence from the Abrupt End of the Multi(cid:133)bre Arrangement," NBER Working Papers 12579, National Bureau of Economic Research. [26] Hummels, David and Peter J. Klenow (2005): "The Variety and Quality of a Nation(cid:146)s Exports," American Economic Review, 95(3): 704-723. [27] Johnson, Robert C. (2008): "Trade and Prices with Heterogeneous Firms," mimeo. [28] Khandelwal, Amit (2009): "The Long and Short (of) Quality Ladders," forthcoming in Review of Economic Studies. [29] Kremer, Michael R. (1993): "TheO-ringTheoryof EconomicDevelopment."The Quarterly Journal of Economics, Vol. 108, No. 3, pg. 551-575, August. [30] Kugler, Maurice and Eric A. Verhoogen (2008): (cid:147)The Quality-Complementarity Hypothesis: Theory and Evidence from Colombia.(cid:148)NBER working paper #14418, Oct. 2008. [31] Manova, Kalina B. and Zhiwei Zhang (2009): "Export Prices and Heterogeneous Firm Models", mimeo, Stanford University. [32] Marazzi, Mario, Nathan Sheets, Robert J. Vigfusson, Jon Faust, Joseph E. Gagnon, Jaime Marquez, Robert F. Martin, Trevor A. Reeve and John H. Rogers (2005): (cid:147)Exchange Rate Pass-through to U.S. Import Prices: Some New Evidence,(cid:148)Board of Governors of the Federal Reserve System International Finance Discussion Paper #833. [33] Melitz, Marc J. (2003): "The Impact of Trade on Intra-Industry Reallocations and AggregateIndustryProductivity."Econometrica,Vol.71,No.6.,pp.1695-1725,November. [34] Moulton, Brent R. & Moses, Karin E. (1997): "Addressing the Quality Change Issue in the Consumer Price Index" Brookings Papers on Economic Activity 1, 305(cid:150)349.

31 [35] Olivei, G.(2002): (cid:147)ExchangeRatesandthePricesofManufacturingProductsImported into the United States,(cid:148)New England Economic Review, First Quarter, 3-18. [36] Schott, Peter K. (2004): "Across-product Versus Within-product Specialization in International Trade," Quarterly Journal of Economics, MIT Press, 119(2): 646-677. [37] Schott, Peter K. (2008): "The Relative Sophistication of Chinese Exports," Economic Policy, 53:5-49. [38] Verhoogen, Eric A. (2008): "Trade, Quality Upgrading, and Wage Inequality in the Mexican Manufacturing Sector," Quarterly Journal of Economics, 123(2): 489-530.

32 Theoretical Moments 1. Mean: Within exporting country k, the average quality-inclusive price of product j at time t is: lnp = lnp f(’) d’ jkt ijkt i !jkt 2 R b = ln(cid:18)p +lnw + ln’ f(’ ’ ! )d’ j jkt a b ijkt j 2 jkt i (cid:20) (cid:18) (cid:0) (cid:19) (cid:21) R b = ln(cid:18)p +lnw + ln’ f(’ ’ ! )d’ j jkt a b ijkt j 2 jkt (cid:18) (cid:0) (cid:19) i b R = ln(cid:18)p +lnw + ln’ j jkt a b jkt (cid:18) (cid:0) (cid:19) where ! is the range of active (cid:133)rms and ln’ is their average productivity level. jkt jkt For product j in year t, the average quality-inclusive price is a simple average across countries: 1 lnp = lnp jt jkt k k P 1 b 1 = ln(cid:18)p + lnw + ln’ j k jkt a b k jkt k (cid:18) (cid:0) (cid:19) k P b P = ln(cid:18)p +lnw + ln’ j jt a b jt (cid:18) (cid:0) (cid:19) 2. Variance: Within country k, the variance of quality-inclusive prices for product j at time t is: 2 Var (lnp ) = lnp lnp f(’) d’ jkt ijkt ijkt jkt (cid:0) i !kt 2 R (cid:2) (cid:3) b 2 = ln’ ln’ f(’) d’ a b ijkt (cid:0) jkt (cid:18) (cid:0) (cid:19)i 2 !kt b 2 R (cid:2) (cid:3) = Var (ln’ ) a b k ijkt (cid:18) (cid:0) (cid:19) Forproductj (allexportcountries), thevarianceofquality-inclusivepricesadditionally

33 depends on the variance of wages across countries: 1 2 Var (lnp ) = lnp lnp f(’) d’ jt ijkt ijkt jt k (cid:0) X k (i 2 R !kt (cid:2) (cid:3) ) 2 1 (lnw lnw )+ jkt jt = (cid:0) f(’) d’ X k k 8 < i 2 R !kt " a (cid:0) b b (ln’ ijkt (cid:0) ln’ jt ) # 9 = (cid:0) (l(cid:1)nw lnw )2+ : jkt jt ; (cid:0) 1 8 2 b 2 (ln’ ln’ )2 3 9 = > a b ijkt (cid:0) jt f(’) d’> > (cid:0) > X k k > > > <i 2 Z !kt 6 6 6 + (cid:0) 2 a (cid:1) (cid:0) b b (lnw jkt (cid:0) lnw jt ) 7 7 7 > > > = (ln’ ln’ ) 6 (cid:0) (cid:1)ijkt jt 7 > 6 (cid:0) 7 > > > 1 > 4 5 > = ( > > :lnw jkt lnw jt )2 > > ; k (cid:0) k X b 2 1 + (ln’ ln’ )2 f(’) d’ a b k ijkt (cid:0) jt (cid:18) (cid:0) (cid:19) X k "i 2 R !kt # b 2 = Var (lnw )+ Var (ln’ ) (22) jt jkt a b jt ijkt (cid:18) (cid:0) (cid:19) In the fourth line, the covariance between item productivity and country wage is assumed to be zero.

34 3. Skewness: Assuming uniform wage across countries, the skewness of quality-inclusive prices for product j is: 1 lnp lnp 3 f(’) d’ k ijkt (cid:0) jt Skew jt (lnp ijkt ) = X k (i 2 R !kt (cid:2) (cid:3) ) 3 2 1 lnp lnp 2 f(’) d’ k ijkt (cid:0) jt X k (i 2 R !kt (cid:2) (cid:3) ) 1 ln’ ln’ 3 f(’) d’ b 3 k ijkt (cid:0) jt = a (cid:0) b X k (i 2 R !kt (cid:2) (cid:3) ) b 3 a (cid:0) b ! 1 ln’ ln’ 2 f(’) d’ 2 (cid:12) (cid:12) k ijkt (cid:0) jt (cid:12) (cid:12) X k (i 2 R !kt (cid:2) (cid:3) ) 3 b = a b Skew (ln’ ) (cid:0) b jt ijkt ! a b (cid:0) (cid:12) (cid:12) (cid:12) (cid:12) This means that the skewness of prices has the same sign as the skewness of (cid:133)rm productivity if b > 0. That is, if prices are increasing in (cid:133)rm productivity and a b (cid:0) productivity has positive skewness (high outliers), then price skewness will be positive; if prices are decreasing in productivity and productivity has positive skewness, then price skewness will be negative. Interestingly, the skewness of prices always has the same magnitude as the skewness of productivity.

35 Allowing wages to di⁄er across countries, the skewness of quality-inclusive prices for product j is: 1 lnp lnp 3 f(’) d’ k ijkt (cid:0) jkt Skew jt (lnp ijkt ) = X k (i 2 R !kt (cid:2) 3 (cid:3) ) [Var j (lnp ijkt )]2 Skew (lnp ) 1 jt ijkt 3 = lnp lnp f(’) d’ [Var jt (lnp ijkt )] (cid:0)2 3 X k k (i 2 R !kt (cid:2) ijkt (cid:0) jkt (cid:3) ) 1 = (lnw lnw )3 +covariance terms jkt jt k (cid:0) k X b 3 1 + (ln’ ln’ )3 f(’) d’ (23) a b k ijkt (cid:0) jt (cid:18) (cid:0) (cid:19) X k "i 2 R !kt # Skew (lnw ) b 3 Skew (ln’ ) jt jkt jt ijkt = + (24) [Var (lnw )] 3 a b [Var (ln’ )] 3 jt jkt (cid:0)2 (cid:18) (cid:0) (cid:19) jt ijkt (cid:0)2 Again, assuming that the covariance terms of industry productivity and country wage arezero,(23)reducestoanexpressionwhichstatesthatquality-inclusivepriceskewness (non-standardized) is a function of cross-country wage skewness and average industry productivity skewness.

36 Figure 1: Monthly transaction price levels within an illustrative HS10 category.

37 2.5 2 1.5 1 0.5 0 1% 5% 10% 25% 50% 75% 90% 95% 99% -0.5 -1 -1.5 -2 Percentile ssenwekS ecirP tropmI .S.U 1 0.5 0 1% 5% 10% 25% 50% 75% 90% 95% 99% -0.5 -1 -1.5 -2 -2.5 Percentile ssenwekS egaW retropxE ngieroF 3.5 3 2.5 2 1.5 1 0.5 0 1% 5% 10% 25% 50% 75% 90% 95% 99% -0.5 -1 Percentile ssenwekS eziS mriF retropxE SU Figure 2: Cross-HS10 distributions of U.S. import price skewness, foreign wage skewness and U.S. firm-level export size skewness Notes: Skewness is measured at the HS10 level for a given month and then averaged over products and time using import sales weights.

38 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 Cost* Cost Quality Quality* No Estimate a. Aggregate imports; unbalanced 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 Cost* Cost Quality Quality* b. Subset of imports; balanced Figure 3: Estimated value share of cost and quality imports over time

39 20 15 10 5 0 -15 -10 -5 0 5 10 15 y = -0.28x + 0.47 -5 -10 Country Mean Price for Skewness>0 Products stcudorP 0<ssenwekS rof ecirP naeM yrtnuoC Mexico (Metals) UK (Plastics) Figure 4: : Estimated mean relative prices in the tail of the price distribution 15 10 5 y = 0.45x - 0.08 0 -4 -2 0 2 4 6 8 10 -5 -10 Country Mean Price for Skewness>0 Products stcudorP 0<ssenwekS rof ecirP naeM yrtnuoC Figure 5: : Estimated mean relative prices in the body of the price distribution

40 8 6 4 2 y = -0.47x + 0.49 0 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 -2 -4 -6 -8 Country Mean Price for Skewness>0 Products stcudorP 0<ssenwekS rof ecirP naeM yrtnuoC Figure 6: : Estimated mean relative prices in the tail of the price distribution 6 4 2 y = -0.08x + 0.06 0 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 -2 -4 -6 -8 Country Mean Price for Skewness>0 Products stcudorP 0<ssenwekS rof ecirP naeM yrtnuoC Figure 7: : Estimated mean relative prices in the body of the price distribution

41 U.S. Imports U.S. Exports Freq. of Freq. of Standard High Low Standard Deviation Skewness Kurtosis Outliers Outliers Deviation Skewness Wood & Wood Products 0.32 -0.37 4.57 0.14 0.12 Raw Hides, Skins, Leather 0.40 -1.25 Mineral Products 0.25 -0.33 4.98 0.16 0.10 Miscellaneous 1.92 -0.28 Animal & Animal Products 0.43 -0.11 4.48 0.10 0.14 Foodstuffs 0.61 -0.20 Raw Hides, Skins, Leather 1.03 -0.10 3.04 0.18 0.13 Vegetable Products 0.28 -0.19 Stone / Glass 1.01 -0.01 2.79 0.17 0.13 Mineral Products 0.21 -0.18 Metals 0.83 0.01 3.46 0.16 0.14 Animal & Animal Products 0.48 -0.10 Footwear / Headgear 0.61 0.03 3.43 0.18 0.14 Chemicals & Allied Ind. 0.85 -0.05 Miscellaneous 1.40 0.03 2.94 0.16 0.15 Transportation 1.33 -0.03 Vegetable Products 0.29 0.05 3.42 0.17 0.13 Mechanical & Computers 1.88 -0.02 Transportation 1.06 0.13 2.93 0.16 0.16 Metals 1.14 -0.01 Foodstuffs 0.52 0.16 3.28 0.17 0.15 Footwear / Headgear 0.59 0.00 Electric Machinery 1.58 0.28 3.29 0.17 0.12 Wood & Wood Products 0.49 0.02 Mechanical & Computers 1.60 0.29 3.71 0.18 0.12 Electric Machinery 1.81 0.09 Chemicals & Allied Ind. 2.08 0.35 2.79 0.19 0.12 Stone / Glass 0.90 0.18 Textiles 0.56 0.38 4.15 0.17 0.11 Plastics / Rubber 1.04 0.27 Plastics / Rubber 1.35 0.73 3.99 0.18 0.10 Textiles 0.36 0.64 1.18 0.15 3.43 0.17 0.13 1.49 0.00 Table 1: The intra-HS10 distribution properties of U.S. import prices, by sector Notes: Sample includes those classification groups with greater than 10 item price observations in a given monthly period. All statistics are calculated within a classification group and then aggregated over classifications and time, weighted by the dollar sales value of imports. The frequency of high price outliers is the number of products with prices greater than one standard deviation above the median price. An analogous defintion holds for the frequency of low price outliers.

42 Top 20 Largest Quality Products by Value 2005 Import Value HS6 ($bn) b/(a-b) sigma Description 870323 46.50 2.02 27.08 Passenger Vehicles, Spark-Ignition, Engine >1500 CC 852520 24.90 2.30 3.08 Transmission Apparatus Incorporating Reception Apparatus 300490 24.00 0.44 11.03 Medicaments NESOI, Measured Doses, Retail, NESOI 852990 5.14 0.36 2.97 Parts For Transmission, Radar, Radio,TV, NESOI 620342 5.11 1.65 4.57 Men'S Or Boys' Trousers, Not Knit, Cotton 901890 4.86 0.29 2.07 Instruments & Appliances For Medical Surgical Dental Vet., NESOI 850440 4.81 0.29 8.91 Static Converters; Automated Data Processor Power Supplies 848180 4.39 1.03 2.38 Other Valves And Other Appliances For Pipes, Tanks, Vats Or The Like 271112 4.24 1.11 6.25 Propane, Liquefied 852190 3.97 1.73 2.20 Video Recording/Reproduction Appliances 853710 3.81 0.82 3.82 Bases For Electric Control Or Distribution, Not Exceeding 1,000V 401110 3.74 2.44 4.88 New Pneumatic Tires Of Rubber, For Motor Cars 760110 3.73 1.20 27.85 Unwrought Aluminum, Not Alloyed 760120 3.47 4.34 9.16 Unwrought Aluminum Alloys 740311 3.24 8.28 33.70 Refined Copper Cathodes And Sections Of Cathodes 220421 3.05 1.35 4.07 Wine, From Grapes, NESOI, <2 Liters 271114 2.87 2.28 29.76 Ethylene, Propylene, Butylene 842952 2.51 1.58 18.00 Mechanical Shovels & Excavators 271113 2.34 2.74 9.55 Butanes, Liquefied 853400 2.12 0.29 9.71 Printed Circuits Top 20 Largest Cost Products by Value 2005 Import Value HS6 ($bn) b/(a-b) sigma Description 847330 27.60 -0.19 Parts & Accessories For Automated Data Processor Machines & Units 870431 10.80 -3.06 42.90 Motor Vehicles For The Transport Of Goods, Not Over 5 Metric Tons 851790 8.59 -0.28 Parts Of Electrical Apparatus For Line Telephony Or Line Telegraphy 840734 6.66 -2.77 26.62 Reciprocating Piston Engines For Vehicles, Exceeding 1, 000CC 854430 5.78 -0.23 Insulated Wiring Sets For Vehicles, Ships, Aircraft 940190 5.74 -0.32 Parts Of Seats (Excl. Medical, Barber, Dental Etc) 950490 3.58 -0.90 1.20 Other Articles For Funfair, Table Or Parlour Games 610910 3.42 -2.44 5.61 T-Shirts, Singlets, Other Vests, Knitted Or Crocheted, Of Cotton 30613 2.80 -3.36 5.25 Shrimps And Prawns, Including In Shell, Frozen 940320 2.72 -1.75 Metal Furniture, NESOI 620520 2.66 -1.04 5.24 Men's Or Boys' Shirts, Of Cotton 843149 2.63 -0.33 2.57 Parts Of Derricks, Cranes, Graders, Levellers, Scrapers Or Pile-Drivers 903289 2.61 -0.79 1.72 Automatic Regulating Or Controlling Instruments & Apparatus 950410 2.52 -1.67 Video Games Of A Kind Used With A Television Receiver 940161 2.48 -1.06 2.17 Seats With Wooden Frames, Upholstered, NESOI 847989 2.46 -0.29 21.75 Air-Coolers, Air Purifiers Of Other Machines And Mechanical Appliances 300439 2.31 -0.56 4.10 Medicaments Containing Other Hormones 847340 2.07 -0.43 Parts And Accessories Of Office Machines, NESOI 901819 1.92 -0.39 28.58 Electro-Diagnostic Apparatus Nesoi, And Parts 853669 1.75 -0.56 1.99 Plugs And Sockets, For A Voltage Not Exceeding 1,000V Table 2: Quality scope estimates, ((b/(a-b))), and a measure of horizontal differentiation, σ, for selected import products

43 Top 15 Largest Quality Products by Value (U.S. Exports) 2005 Export Value HS6 ($bn) Description 854221 22.2 Digital Monolithic Integrated Circuits 870323 12.6 Passenger Vehicles, Spark-Ignition, Engine>1,500CC 870324 9.8 Passenger Vehicles, Spark-Ignition, Engine>3,000CC 271019 9.7 Light Petroleum Distillates, NESOI 854229 6.2 Monolithic Itegrated Circuits, Other Than Digital 300210 4.1 Antisera And Other Blood Fractions 520100 3.9 Cotton, Not Carded Or Combed 841199 3.8 Gas Turbine Parts, NESOI 851750 3.7 Telecommunications Apparatus For Line Systems 847149 2.9 Digital Automated Processing Machines And Units 847150 2.9 Digital Processing Units, NESOI 840991 2.9 Spark-Ignition Internal Combustion Piston Engine Parts 760612 2.1 Aluminum Alloy Rectangular Plates, Over .2MM Thick 848180 2.1 Other Valves And Appliances For Pipes, Tanks And Vats 470321 2.0 Wood Pulp, Soda Or Sulphate, Coniferous, Bleached Top 15 Largest Cost Products by Value (U.S. Exports) 2005 Export Value HS6 ($bn) Description 847330 12.2 Parts & Accessories For Automated Data Processing Machines 870899 10.2 Parts & Accessories For Motor Vehicles, NESOI 870829 8.6 Parts & Accessories Of Bodies Of Motor Vehicles, NESOI 841191 7.8 Turbojet And Turbopropelor Parts 100590 4.9 Maize, Except Seed Corn 847989 4.6 Machines And Mechanical Appliances With Individual Function, NESOI 841112 4.5 Turbojets Of A Thrust Exceeding 25 Knots 840734 3.9 Reciprocating Piston Engines For Vehicles, Exceeding 1,000CC 870840 3.5 Gear Boxes For Motor Vehicles 851790 3.5 Parts Of Electrical Apparatus For Line Telephony Or Line Telegraphy 382200 3.5 Composite Diagnostic Or Laboratory Reagents, NESOI 392690 3.4 Plastic Articles, NESOI 840820 3.4 Compression-Ignition Internal Combustion Piston Engine 852520 3.3 Transmission Apparatus Incorporating Reception Apparatus 270112 3.1 Bituminous Coal, Not Agglomerated Table 3: Quality scope estimates for selected export products.

44 Skewness>0 Products Skewness<0 Products 15th 85th N 15th 85th N Animal & Animal Products -0.06 -0.09 3,404 0.03 0.05 5,174 Vegetable Products 0.34 * 0.64 * 2,747 0.33 * 0.62 * 1,801 Plastics / Rubber -0.04 2.63 * 5,176 -0.39 0.65 703 Textiles 0.06 0.48 * 2,633 -0.59 * -0.02 3,132 Metals 0.22 * 0.36 * 2,816 1.91 * 0.05 2,200 Mechanical & Computers 0.40 * 0.58 * 7,992 0.16 1.06 * 8,550 Transportation 0.70 * 0.09 22,531 2.47 * 2.30 * 6,504 (a) Wage elasticity of item prices, pooled across quantiles Skewness>0 Products Skewness<0 Products 15th 85th N 15th 85th N Wage Animal & Animal Products -0.70 * -0.66 3,154 0.11 0.09 * 5,591 Vegetable Products 0.56 * 0.31 * 2,384 0.29 * 0.25 * 1,529 Mechanical & Computers -0.51 -0.93 7,427 1.45 * 1.84 * 8,221 Transportation 0.01 0.07 22,483 1.41 * 0.93 * 5,997 Market Transfer Animal & Animal Products 0.21 * 0.11 Vegetable Products 0.07 0.05 0.73 * 1.04 * Mechanical & Computers 0.16 -1.43 * -1.42 * 0.12 Transportation -0.13 -0.17 * -2.61 * -1.94 * Cost Transfer Animal & Animal Products Vegetable Products 0.12 -0.11 1.32 * 0.72 * Mechanical & Computers 1.30 -1.35 * -0.18 1.30 * Transportation 0.54 * -0.01 1.04 * 1.73 * Irregular Discontintuation Animal & Animal Products -0.09 * -0.04 * 0.08 * 0.03 * Vegetable Products -0.35 * -0.16 * -0.07 * -0.02 Mechanical & Computers -0.39 * 0.02 -0.26 * 0.28 * Transportation 0.05 0.01 0.01 -0.36 * (b) Wage elasticity of item prices with additional controls Table 4: The correlation of item price and wage Notes: Shown are estimated coefficients for the wage elasticity and other controls in equation (18). An asterisk denotes significance at the 10 percent level.

45 Quality Cost Skewness>0 Skewness<0 Products Products 15th 85th 15th 85th Pakistan -1.68 1.15 -0.82 1.39 Bangladesh -1.53 1.11 -0.32 1.39 Egypt -1.13 0.47 1.13 1.33 Costa Rica -1.56 0.34 0.57 0.08 China -1.12 0.26 -0.20 1.43 India -1.18 -0.22 -1.52 -0.47 Malaysia -1.27 -0.77 0.84 0.05 Turkey -0.91 -0.78 1.96 1.62 Thailand -1.56 -0.79 0.33 1.47 Philippines -1.27 -0.88 0.17 0.04 Colombia -1.37 -0.90 0.35 0.20 Italy -0.48 -1.15 1.28 0.38 Korea -1.76 -1.68 0.98 -0.06 Macao -1.38 -1.76 1.33 -0.06 Hong Kong -1.44 -1.83 1.40 0.04 Table 5: Estimated country relative prices for exporters in the textiles sector at the 15th and 85th quantiles of the price distribution Notes: Shown are estimated coefficients for α in equation (18). Quantile regressions are estimated separately for Quality and 2c Cost products. All estimates are significant at the 10 percent level.

46 Cost Industries All Industries Quality Industries Dep. Variable: Constant-Quality Quality-Inclusive Constant-Quality Quality-Inclusive Constant-Quality Quality-Inclusive Import Price Index (Transaction-Level) (Unit Value) (Transaction-Level) (Unit Value) (Transaction-Level) (Unit Value) (I) (II) (III) (IV) (V) (VI) (VII) (VIII) (IX) (X) (XI) (XII) Real Exch. Rate 0.07 ** 0.06 ** 0.09 0.11 0.08 ** 0.08 ** 0.21 ** 0.21 ** 0.10 ** 0.10 ** 0.27 ** 0.27 ** (w/ 2 lags) (0.01) (0.01) (0.06) (0.06) (0.01) (0.01) (0.04) (0.05) (0.01) (0.01) (0.08) (0.08) Export Price 0.17 ** 0.15 ** 0.21 * 0.20 0.17 ** 0.16 ** 0.27 ** 0.27 ** 0.16 ** 0.16 ** 0.30 * 0.29 * (0.02) (0.02) (0.10) (0.11) (0.01) (0.01) (0.08) (0.08) (0.02) (0.02) (0.12) (0.12) Tariff 0.04 ** -0.10 0.00 0.03 -0.05 ** 0.06 (0.01) (0.07) (0.01) (0.06) (0.01) (0.10) China Share 0.00 ** 0.01 * 0.00 0.00 0.00 0.00 (0.00) (0.01) (0.01) (0.00) (0.00) (0.01) Observations 4,530 4,500 4,530 4,500 13,012 12,892 13,012 12,892 7,035 6,989 7,035 6,989 R-squared 0.76 0.76 0.82 0.82 0.75 0.75 0.75 0.75 0.75 0.75 0.72 0.72 Notes: Shown are estimates for (19) using ordinary least squares and classification group fixed effects. Standard errors are reported in parentheses, and stars denote the 5 and 1 percent significance levels. Table 6: Real exchange rate pass-through to constant-quality and quality-inclusive import prices (annual, 1994-2004)

47 HS 84 & 85 only Export-Based Quality Scope Measure Cost Industries All Industries Quality Industries Cost Industries All Industries Quality Industries Dep. Variable: Constant- Quality- Constant- Quality- Constant- Quality- Constant- Quality- Constant- Quality- Constant- Quality- Import Price Index Quality Inclusive Quality Inclusive Quality Inclusive Quality Inclusive Quality Inclusive Quality Inclusive (I) (II) (III) (IV) (V) (VI) (VII) (VIII) (IX) (X) (XI) (XII) Real Exch. Rate 0.11 ** -0.03 0.08 ** 0.30 ** 0.06 ** 0.43 ** 0.07 ** 0.08 0.07 ** 0.23 * 0.08 ** 0.38 * (w/ 2 lags) (0.02) (0.22) (0.01) (0.15) (0.01) (0.19) (0.01) (0.12) (0.01) (0.11) (0.01) (0.19) Export Price 0.21 ** 0.72 * 0.20 ** 0.13 0.19 ** -0.13 0.20 ** 0.08 0.18 ** 0.22 0.17 ** 0.29 (0.03) (0.32) (0.02) (0.24) (0.02) (0.31) (0.02) (0.22) (0.01) (0.19) (0.02) (0.29) Tariff 0.30 ** -2.06 ** 0.20 ** 0.79 ** 0.11 ** 1.84 ** 0.05 * -0.49 * 0.04 * -0.17 0.03 0.04 (0.06) (0.71) (0.02) (0.50) (0.04) (0.63) (0.02) (0.22) (0.01) (0.19) (0.02) (0.30) China Share 0.00 * -0.02 0.00 ** -0.01 0.00 * 0.00 0.00 * -0.01 0.00 ** 0.22 0.00 ** 0.01 (0.00) (0.02) (0.00) (0.01) (0.00) (0.02) (0.00) (0.01) (0.00) (0.19) (0.00) (0.01) Observations 902 902 3,515 3,515 2,613 2,613 2,139 2,139 4,154 4,154 2,015 2,015 R-squared 0.88 0.87 0.86 0.72 0.85 0.66 0.82 0.73 0.82 0.69 0.83 0.66 Notes: Shown are estimates for (19) using ordinary least squares and classification group fixed effects. Standard errors are reported in parentheses, and stars denote the 5 and 1 percent significance levels. Table 7: Real exchange rate pass-through to constant-quality and quality-inclusive import prices (annual, 1994-2004)