Concording U.S. Harmonized System Categories Over Time

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Concording U.S. Harmonized System Categories Over Time Justin R. Pierce and Peter K. Schott 2012-16 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Concording U.S. Harmonized System Categories Over Time (cid:3) Justin R. Pierce y Board of Governors of the Federal Reserve System Peter K. Schott z Yale School of Management & NBER January 2012 Abstract Monitoring changes to product classi(cid:133)cation systems is an important component of a wide range of empirical research. In this paper we develop an algorithm for concording periodic revisions to the ten-digit HarmonizedSystem(HS)codesusedbyU.S.statisticalagenciestocategorizeinternationaltradesince1989. We use this algorithm to construct the (cid:133)rst comprehensive concordance of HS codes over time, and show how this concordance can be extended to incorporate future revisions. We then characterize the extent of HS-code changes since 1989 and discuss how controlling for these revisions is critical for understanding the growth of U.S. trade. Lastly, we highlight the general applicability of the algorithm to other national and international product classi(cid:133)cation systems. Keywords: Internationaltrade;product classi(cid:133)cation JEL classi(cid:133)cation: F1 (cid:3)We thank Julie Linden of the Yale University Social Sciences Library for generous help in securing the publicly available U.S. trade data. We thank Kitjawat Tacharoen and Matt Flagge for research assistance. We thank Alvin Venning, Carol Ann Aristone, James Kristo⁄and Mendel Gayle of the U.S. Census Bureau for many enlightening conversations. SchottthankstheNationalScienceFoundation(SES-0241474andSES-0550190)forresearchsupport. Pierce thanks the U.S. Census Bureau where he was employed for a large portion of this project. We also thank the editor for helpful comments. The analysis and conclusions set forth in this paper are those of the authors and do not indicate concurrence by the Board of Governors, other members of the research sta⁄or the National Science Foundation. y20th and C ST NW, Washington, DC 20551, U.S.A. Email: justin.r.pierce@frb.gov. z135 Prospect Street, New Haven, CT 06520, U.S.A. Email: peter.schott@yale.edu.

2 1. Introduction Empirical researchers including Bernard, Redding and Schott (2010, 2011), Bernard, Jensen, Redding, and Schott (2009), Goldberg, Khandelwal, Pavcnik, and Topalova (2010) and Pierce (2011), increasingly use product-level data to study trends in exports, imports and domestic production. These data have been particularly useful for examining the extent to which (cid:133)rms(cid:146)growth in output or trade is due to (cid:147)intensive(cid:148)versus (cid:147)extensive(cid:148)margins, i.e., the degree to which growth takes place within surviving products or via product adding and dropping. At the same time, national statistical agencies frequently update product classi(cid:133)cation systems to incorporate new goods, drop obsolete categories and harmonize their systems with other countries. Absent a proper concordance, it can be di¢ cult for researchers to distinguish true product-switching from spurious changes to product mix associated with product reclassi(cid:133)cations. In this article we present an algorithm for constructing a concordance among revisions of the Harmonized System (HS) product codes used to track U.S. exports and imports over time. HS codes have been used by the U.S. Census Bureau since 1989 and are updated frequently. Our algorithm matches revised codes to synthetic, time-invariant identi(cid:133)ers that follow (cid:147)families(cid:148)of related products. We use our algorithm to construct the (cid:133)rst comprehensive concordance of U.S. HS codes over time, covering the period 1989 to 2009. In an electronic appendix, we provide the Stata code used to build the concordance, thereby allowing other researchers the means to customize it or to extend it to incorporate future revisions of HS categories. Our concordance reveals that changes in HS codes are frequent and widespread, and that they a⁄ect product categories representing a substantial portion of trade value. Indeed, of the 16,836 (8,859) import (export) codes active in 2004, 7,503 (2,929) underwent revision between 1989 and 2004(cid:150)the years examined in Bernard, Jensen, Redding and Schott (2009). Furthermore, these revised codes represent 59 and 43 percent of import and export value in 2004, respectively. The prevalence and importance of product code changes in U.S. trade underscore the need for HS code concordances in the analysis of trade (cid:135)ows. Using our concordance to control for changes to product categories over time, for example, Bernard, Jensen, Redding, and Schott (2009) show that most of the year-to-year change in U.S. trade (cid:150)as well as adjustments to (cid:147)shocks(cid:148)such as the 1997 Asian (cid:133)nancial crisis (cid:150)occur along the intensive margin. The algorithm is general enough to be used to create concordances of virtually any national or international product classi(cid:133)cation system over time. This includes other international trade product classi(cid:133)cation systems such as the European Union(cid:146)s Combined Nomenclature or the Tari⁄ Schedule of Japan. Moreover, the algorithm can be employed to construct concordances over time for a variety of national or international production-based product classi(cid:133)cation systems such as the North American Industry Classi(cid:133)cation System (NAICS), International Standard Industrial Classi(cid:133)cation (ISIC) or the statistical classi(cid:133)cation of economic activities in the European Union(NACE). Theremainderofthearticleisorganizedasfollows. Section2providesabriefdescriptionofU.S. HS codes. Section 3 describes the data used to construct our concordance and Section 4 outlines the concordance algorithm. Section 5 describes the properties of a 1989 to 2004 HS-over-time concordance created using the algorithm from Section 4. Section 6 shows the e⁄ect of using the HS-over-time concordance on the measurement of product-adding and dropping using year-overyear decompositions of U.S. exports as in Bernard, Jensen, Redding, and Schott (2009). Section 7 describes the general applicability of the algorithm to other product classi(cid:133)cation systems. An electronic appendix on our personal websites provides concordance (cid:133)les in .csv format, as well as the Stata code used to generate the concordances.

3 2. Brief Description of HS Codes U.S. HS codes are based on the Harmonized System established by the World Customs Organization (WCO). The WCO assigns 6-digit codes for general categories, and countries adopting the system then de(cid:133)ne their own codes to capture commodities at more detailed levels. In the United States, the most detailed level of disaggregation is ten digits. In this article, we refer to ten-digit codes as (cid:147)product(cid:148)or (cid:147)goods(cid:148)categories. U.S. export codes(cid:150)technically referred to as Schedule B codes(cid:150)are administered by the United States Census Bureau (Census). U.S. import codes(cid:150)technically referred to as Harmonized Tari⁄ System (HTS) codes(cid:150)are administered by the U.S. International Trade Commission (USITC). We refer to HTS and Schedule B codes together as (cid:147)HS Codes(cid:148)throughout this article. Changes to U.S. export or import product codes can occur via three routes: changes by the WCO to the o¢ cial list of international six-digit pre(cid:133)xes; U.S. legislation that a⁄ects U.S. eightdigit codes (imports only); and changes by the Committee for Statistical Annotation of Tari⁄ Schedules (known as the (cid:147)484(f) Committee(cid:148)) to statistical ten-digit codes. HS codes are updated for several reasons. The WCO, for example, makes adjustment to the HS to re(cid:135)ect developments in technology and changes in trade patterns. In addition, the 484(f) Committee may split a single HS code into several new codes in order to report import or export data at a more detailed level. Similarly, producers may petition one of the o¢ cial bodies noted above for code changes to obtain a higher pro(cid:133)le for the goods they export or import. A large number of changes in 10-digit U.S. HS codes can be attributed to the WCO(cid:146)s revisions of 6-digit HS categories. The WCO has made three major revisions to the HS in 1996, 2002, 2007, with another revision planned for 2012. Each of these revisions resulted in hundreds of 6-digit HS categories being deleted, while hundreds of other 6-digit HS categories were added. The e⁄ect of the WCO(cid:146)s revisions on the number of U.S. HS changes is apparent in Table 1, where a large number of HS changes are concentrated in WCO revision years. 3. Data Each year, Census publishes documents outlining the HS codes that have become (cid:147)obsolete(cid:148) and the (cid:147)new(cid:148)codes that will take their place. We refer to these documents as Census(cid:146)(cid:147)obsoletenew(cid:148)(cid:133)les. For exports, HS code changes take e⁄ect annually in January; for imports, they can occur within as well as across years. Obsolete-new (cid:133)les for years before 1997 are available only in hard copy and were transcribed into electronic form as part of the construction of our concordance. These (cid:133)les as well as electronic versions of subsequent (cid:133)les were obtained from Mayumi Hairston EscalanteatCensus. Themostrecentobsolete-new(cid:133)lesarecurrentlypostedontheCensuswebsite. We use the terms (cid:147)simple(cid:148)and (cid:147)complex(cid:148)to describe the two basic changes to HS codes that can occur in a obsolete-new (cid:133)le. Simple changes make no adjustments to the actual items covered by a particular code, they just swap one ten-digit code for another. There are several possible reasons for a one-to-one renumbering, including: 1. To align the Schedule B and HTS codes where Census (cid:133)nds their descriptions are the same; 2. To di⁄erentiate the Schedule B and HTS codes where Census has found them to be di⁄erent; 3. To correct errors by reclassifying a commodity under a di⁄erent subheading; 4. To maintain the level of statistical detail after a revision of the 6- or 8-digit codes; and

4 5. To accommodate a new numbering pattern, usually the result of another code being broken out. In contrast to simple changes, complex changes alter the mix of items captured by a particular code. For these changes, the items formerly encompassed by one or more (cid:147)obsolete(cid:148)codes are distributed to one or more (cid:147)new(cid:148)codes. In 2002, for example, various types of waste oil, which previously were grouped with the fresh oils to which they were most similar, were given their own HS codes. As a result, the (now obsolete) former fresh oil product categories were linked to the new waste oil categories from which they emerged. Some new-obsolete (cid:133)les contain (cid:147)blanket(cid:148) mappings, our term for mappings that include codes ending in a series of X(cid:146)s, e.g., 8486XXXXXX. These observations are dropped from our concordance, as we are unable to determine the speci(cid:133)c HS codes to which they refer. Foreachsetofobsolete-newmappingsinaparticularobsolete-new(cid:133)le, weconstructasynthetic HS code which we refer to as a (cid:147)setyear(cid:148)(setyr in our Stata code). This synthetic code records both the count of the change since the (cid:133)rst change in 1989 and an identi(cid:133)er for when it takes place. Formally, for exports, it is de(cid:133)ned as the count of the particular mapping plus the four-digit year in which the change occurs divided by 10,000. For imports, it is the count of the particular mapping plus six-digit year-month in which the change occurs divided by 1,000,000. The very (cid:133)rst setyears for exports and imports, for example, are equal to 1.1989 and 1.198906. Table 1 summarizes the number of obsolete-new mappings in the raw data for export and import codes, respectively. Results for export codes are displayed in the left panel while those for import codes are displayed in the middle and right panels. The (cid:133)rst column of each panel notes the year-month in which the noted changes take place. The second and third columns report the total number of retired and replacement codes encompassed by the number of sets reported in column four. Note that the number of sets in column four of each panel is smaller than the numbers of HS codes in columns two and three because multiple codes are often involved in a particular change (i.e., a particular set). The (cid:133)fth column reports the number of changes that are (cid:147)simple(cid:148)in the sense outlined above. As indicated in the table, HS codes are updated unevenly in the sense that some years (e.g., 2002) encompass substantially more changes than others (e.g., 2000). 4. An Algorithm for Creating an HS Concordance Concording HS codes over time is complicated by the existence of chains of HS-code changes across months and years, which we refer to as (cid:147)family trees(cid:148). There are two basic types of family tree. We refer to the (cid:133)rst case, displayed in Figure 1, generically as a (cid:147)growing family tree(cid:148). In this case, code a from period t may become obsolete and be mapped to new codes b and c in period t+1. Then, in period t+2, codes b and c may become obsolete and be mapped to new codes e and f, and g and h, respectively. Our concordance of the period t to period t+2 HS codes assigns a common synthetic code to all HS codes in a growing family tree. Such an assignment may result in potentially many more HS codes being mapped to a given synthetic code in the (cid:133)nal year of the concordance than in the (cid:133)rst year. In 1997, for example, 7802000000 is mapped to 7802000030 and 7802000060. In a 1996 to 1997 concordance, we would assign a single synthetic HS code to all of these actual HS codes. For this reason, it may be useful for some analyses to restrict a concordance to a narrower set of years than the 1989 to 2009 concordance provided below. The second type of family tree, which we refer to generically as a (cid:147)shrinking family tree(cid:148), is displayed in Figure 2. In this case, codes a and b, and c and d, from period t separately become obsoleteandmappedtocodeseandf,respectively,inperiodt+1. Then,inperiodt+2,codeseand

5 Table 1: HS Code Changes by Year-Month Exports Imports Date Obsolete New Sets Simple Date Obsolete New Sets Simple Date Obsolete New Sets Simple 1989_01 234 310 157 92 1989_06 2 12 2 0 1999_01 81 88 53 16 1990_01 156 201 96 60 1989_07 112 196 91 27 1999_07 54 70 33 5 1991_01 186 313 131 34 1990_01 346 724 295 15 2000_01 16 29 13 0 1992_01 37 60 29 9 1990_05 16 20 16 12 2000_03 11 30 11 0 1993_01 64 126 60 19 1990_07 133 256 119 25 2000_04 10 17 7 0 1994_01 109 181 77 25 1990_08 38 49 30 17 2000_07 6 13 6 1 1995_01 137 205 113 63 1990_10 70 121 47 6 2000_12 24 45 24 3 1996_01 787 1,071 532 349 1991_01 69 194 45 0 2001_01 119 113 55 1 1997_01 216 232 145 107 1991_02 15 24 15 6 2001_07 19 25 9 3 1998_01 128 138 101 76 1991_05 11 20 11 2 2002_01 1,122 1,542 874 595 1999_01 23 29 22 17 1991_07 247 393 190 77 2002_07 86 84 66 49 2000_01 6 15 6 0 1992_01 85 138 50 0 2002_08 5 10 5 0 2001_01 16 9 7 0 1992_05 28 29 28 27 2003_01 26 44 20 0 2002_01 717 1,031 531 323 1992_07 117 194 109 42 2003_02 1 2 1 0 2003_01 97 87 81 74 1993_01 135 218 74 7 2003_04 5 4 4 3 2004_01 11 14 10 5 1993_02 42 51 42 33 2003_07 45 67 37 11 2005_01 43 82 38 8 1993_06 3 5 2 0 2004_01 46 38 23 2 2006_01 3 4 2 0 1993_07 7 8 7 6 2004_02 5 7 4 0 2007_01 1,140 1,030 821 631 1993_08 33 53 25 0 2004_04 4 4 2 0 2008_01 64 68 65 61 1993_11 8 10 2 0 2004_07 44 87 37 1 2009_01 15 15 11 4 1993_12 1 2 1 0 2005_01 42 72 39 11 1994_01 667 1,082 468 176 2005_07 32 45 26 9 1994_04 13 43 13 0 2005_11 4 8 4 0 1994_06 66 112 47 0 2006_01 19 38 19 0 1995_01 1,933 2,187 1,162 555 2006_03 2 2 2 2 1995_07 38 73 31 0 2006_04 4 5 4 3 1995_09 77 168 33 12 2006_06 49 58 9 0 1996_01 1,164 1,485 798 523 2006_07 63 59 35 0 1996_06 5 8 5 4 2007_01 2,026 1,896 1543 1,220 1996_07 4 12 4 0 2007_07 25 35 16 3 1996_11 18 31 18 3 2008_01 19 39 13 0 1997_01 148 198 107 66 2008_04 12 8 6 0 1997_02 11 11 11 11 2008_07 15 34 15 0 1997_06 18 33 18 3 2008_10 12 26 12 0 1997_07 231 319 190 89 2009_01 42 61 28 1 1997_08 55 65 33 1 2009_07 20 39 20 3 1998_01 52 85 47 18 1998_03 4 8 2 0 1998_04 3 3 3 3 1998_07 6 8 6 4 1998_08 9 23 9 0 Notes: Table reports changes to export (left panel) and import (middle and right panel) HS codes in noted year month. Obsolete is number of codes retired from prior year. New is number of codes replacing these retirements. Sets is a count of the overall number of obsolete new matches. Simple refers to re numberings of individual codes. f become obsolete and are assigned to new code g. In this case, the number of HS codes mapped to the family(cid:146)s common synthetic code declines over time. In 1997, for example, 8506800010 and 8506800050 are mapped to 8506800000. In a 1996 to 1997 concordance, we would assign a single synthetic HS code to all of these actual HS codes. The algorithm we develop for concording HS codes between arbitrary beginning and ending year-months accounts for both types of family trees, as well as combinations of the two types. Though speci(cid:133)c details about how the algorithm is implemented can be determined by examining the Stata code in the electronic Appendix, the basic steps are as follows: 1. Read in raw obsolete-new mappings; 2. Assign a single setyear to each obsolete-new mapping appearing in the raw (cid:133)les; 3. Choose a beginning and end year for the concordance; 4. Identify family trees extending between the beginning and end years of the concordance; and

6 Figure 1: Growing Family Tree Figure 2: Shrinking Family Tree

7 5. Assign all members of a family tree the minimum setyear among family members within the time-frame of the concordance. Note that the part of the setyear after the decimal point identi(cid:133)es the year in which the family tree starts (i.e., period t in Figures 1 and 2 above). In the Stata code below, a separate variable (named effyr) identi(cid:133)es the year in which a particular obsolete-new mapping occurs. For example, in 1998 export code 8531800035 from 1997 is mapped to code 8531804000. Then, in 2002, codes 8531804000 and 8527908015 from 2001aremappedinto8527908600. Thesetyrforthefamilyis1404.1998. Theintegerpartof this setyr indicates that the (cid:133)rst mapping in the family, from 8531800035 to 8531804000, is the 1404th mapping since 1989. The part after the decimal point indicates it occurs in 1998. The effyr for the two mappings are 1998 and 2002, respectively. Step four is accomplished by successively merging subsequent obsolete-new mappings to all periods(cid:146)obsolete-new mappings between the beginning and end years of the concordance. To bridge codes used from 1989 onwards, for example, the chained (cid:133)le is constructed as follows. First, merge the new codes in the 1990 (cid:133)le to the obsolete codes in 1991 (cid:133)le, dropping any codes that are uniqueto1991. Second, mergetheobsoletecodesinthe1992(cid:133)letothenewcodesinthepreviously merged 1990-1991 (cid:133)le, again dropping any codes unique to 1992. This procedure is then repeated until reaching the desired end year of the concordance. Note that this successive merging has to be done starting with every year-month between the beginning and ending year-month because chains can begin in any year-month, and they would be missed otherwise given the dropping just mentioned. After these chains are created, they are appended into a single (cid:133)le and added to all obsolete-new mappings that are not parts of a chain. 5. A 1989-to-2004 Concordance This section describes a 1989 to 2004 concordance constructed using the algorithm described above, which was employed in Bernard, Jensen, Redding, and Schott (2009). The (cid:133)rst and second columns of Table 2 summarize total U.S. exports in 1989 and 2004 and the total number of HS product categories exported in those two years, respectively. Columns three and four provide analogous detail with respect to U.S. imports. As indicated in the table, (nominal) exports more than double while (nominal) imports more than triple over the (cid:133)fteen-year interval. The number of preconcordance export and import HS codes observed in each year of data grows 13 percent and 21 percent, respectively. Table 2: Trade in 1989 and 2004 Exports Imports Value Codes Value Codes 1989 354 7,853 468 13,941 2004 818 8,859 1,460 16,836 Notes: Export and import values in billions of U.S. dollars. Number of codes refers to number of original ten digit HS categories in the raw trade data. Table3reportstwodecompositionsofexportandimportcodes. The(cid:133)rstthreerowsoftheTable show how many of the original HS codes in each year survive versus being replaced by synthetic codes. The remaining rows in the table decompose the actual plus synthetic codes that remain after the concordance into those which are common across years and those which are idiosyncratic to a particular year.

8 Table 3: Distribution of HS Codes in Matched 1989 to 2004 Trade Data Exports Imports 1989 2004 1989 2004 Original HS codes 7,853 100 8,859 100 13,941 100 16,836 100 Not replaced by synthetic codes 5,936 76 5,930 67 9,383 67 9,333 55 Replaced by synthetic codes 1,917 24 2,929 33 4,558 33 7,503 45 Actual + synthetic codes after concordance 7,162 91 7,157 81 12,527 90 12,534 74 Actual codes 5,936 76 5,930 67 9,383 67 9,333 55 Common to both years 5,904 75 5,904 67 9,047 65 9,047 54 Appear in only one year 32 0 26 0 336 2 286 2 Synthetic codes 1,226 16 1,227 14 3,144 23 3,201 19 Common to both years 1,221 16 1,221 14 3,057 22 3,057 18 Appear in only one year 5 0 6 0 87 1 144 1 Notes:TabledecomposesthenumberoforiginalHS codesineachyearintothosereplacedby asynthetic code versus not, and total surviving HS plus synthetic codes in each year into noted sub groups. All replacements arewithrespect toa1989to2004concordance. Evencolumnsdisplay valuesas apercent of first row in preceding column. Of the 7,853 original HS codes appearing in the 1989 U.S. export data, for example, 1,917 are replaced by synthetic codes. Since the same synthetic code is often assigned to more than one original code, the resulting concorded dataset contains 7,162 actual plus synthetic codes. Of these, 5,936 and 1,226 are actual and synthetic, respectively. Each of these totals, in turn, can be broken down into actual codes which are common to both 1989 and 2004 (5,904), synthetic codes that are common to both 1989 and 2004 (1,221), actual codes unique to 1989 (32) and synthetic codes that are unique to 1989 (5). These breakdowns reveal that the number of actual and synthetic export and import goods actually added and dropped between 1989 and 2004 is relatively small. The values of U.S. exports and imports associated with each of the cells in Table 3 are reported in Table 4. As indicated below, synthetic codes account for the majority of import value in both 1989 and 2004. Table 4: Distribution of Value in Matched 1989 to 2004 Trade Data Exports Imports 1989 2004 1989 2004 Original HS codes 353,765 100 817,936 100 468,012 100 1,460,160 100 Not replaced by synthetic codes 222,293 63 467,854 57 196,051 42 600,941 41 Replaced by synthetic codes 131,472 37 350,082 43 271,961 58 859,219 59 Actual + synthetic codes after concordance 353,765 100 817,936 100 468,012 100 1,460,160 100 Actual codes 222,293 63 467,855 57 196,051 42 600,942 41 Common to both years 204,570 58 448,183 55 193,451 41 588,628 40 Appear in only one year 17,723 5 19,672 2 2,600 1 12,314 1 Synthetic codes 131,472 37 350,082 43 271,962 58 859,219 59 Common to both years 131,405 37 347,416 42 270,859 58 855,029 59 Appear in only one year 67 0 2,666 0 1,103 0 4,190 0 Notes: Table decomposes U.S export and import value according to whether HS codes are original or synthetic.Allreplacementsarewithrespect toa1989to2004concordance. Valuesareinmillions ofU.S. dollars. Even columns display values as a percent of first row in preceding column. Tables 3 and 4 also underscore the prevalence of changes in HS codes over time. As of 2004, 45 percent of import products and 33 percent of export products had been involved in an HS code change since 1989. Moreover, trade in products with code changes accounted for 59 percent of the value of U.S. imports and 43 percent of the value of U.S. exports in 2004. We note that two features of Census(cid:146)new-obsolete mappings complicate the identi(cid:133)cation of

9 new product introductions (e.g., iPods). First, new HS codes always emerge from predecessor HS codes. Second, new HS codes(cid:146)emergence may take place an unknown period of time after an underlying good has been introduced. Statistical agencies may wait to establish a new HS category until it reaches a certain size or until manufactures apply su¢ cient lobbying. 6. The E⁄ect of the Concordance on Measurement of Product Adding and Dropping In this section we illustrate the importance of controlling for HS code reclassi(cid:133)cations when measuring product adding and dropping in U.S. export data. In Table 5 below, we present the value and share of U.S. exports associated with product adding and dropping, both with and without controlling for changes in HS codes over time. The top portion of the table reports results with unadjusted HS codes and the bottom portion reports results after controlling for HS code reclassi(cid:133)cations using our concordance We report these results for two-year periods between 1993 and 2003 as in Bernard, Jensen, Redding, and Schott (2009). The(cid:133)guresreportedinTable5weregeneratedusingpublicly-availableproduct-levelU.S.export data. At this level of data aggregation, product adding refers to an instance in which the U.S. does not export a product p in the beginning year of the period, but does export that product in the end year. Similarly, product dropping refers to an instance in which the U.S. did export a product in the beginning year, but did not export that product in the end year. Table 5: Value of Exports Associated with Product Adding and Product Dropping: With and Without Concordance 1993 19941994 1995 1995 1996 1996 1997 1997 19981998 19991999 20002000 2001 2001 20022002 2003 Added products 11,934 63,662 108,544 15,735 25,009 4,338 1,484 4,593 92,395 4,587 Added products (% Beginning year exports) 2.6% 12.4% 18.6% 2.5% 3.6% 0.6% 0.2% 0.6% 12.6% 0.7% Dropped products 11,028 52,010 102,890 16,547 24,907 4,114 1,954 4,920 101,289 5,357 Dropped products (% Beginning year exports) 2.4% 10.1% 17.6% 2.7% 3.6% 0.6% 0.3% 0.6% 13.9% 0.8% Dropped products 360 53 963 713 522 220 477 208 683 420 Dropped products (% Beginning year exports) 0.1% 0.0% 0.2% 0.1% 0.1% 0.0% 0.1% 0.0% 0.1% 0.1% Added products 276 15 900 26 2,172 2,573 6 1,937 44 41 Added products (% Beginning year exports) 0.1% 0.0% 0.2% 0.0% 0.3% 0.4% 0.0% 0.2% 0.0% 0.0% ecnadrocnoc oN ecnadrocnoc htiW Notes: Table displays the value ofU.S. exports associated with added and dropped products over two year time periods where products are definedboth withoutand withthe HS over timeconcordance. Rows for "AddedProducts" and"Dropped Products"are measuredin Millions of U.S.Dollars. Additionalrowsreportthevalueassociatedwithaddedanddroppedproductsasashareofthetotalvalueofexportsinthebeginning year of each two year period. As can be seen in the table, the value of exports associated with product adding and dropping is greatly overstated in the (cid:147)no concordance(cid:148)case with unadjusted HS codes. The reason for this overstatement is intuitive(cid:150)some of the products that appeared and disappeared during each twoyear period were due to changes in HS codes, rather than the U.S. starting or stopping exporting those products. This phenomenon is particularly pronounced in time periods with many HS code changes such as 1995-1996 and 2001-2002. In the period from 1995-1996, for example, export data with unadjusted HS codes indicate that product adding (dropping) equaled 19 percent (18 percent) of the value of 1995 exports. After using the concordance, the shares of 1995 exports associated with product adding and dropping were 0.2 percent each.

10 This example illustrates the importance of properly controlling for changes in HS codes in research examining product-adding and dropping. Indeed, accounting for these changes in HS codes contributed to Bernard, Jensen, Redding, and Schott(cid:146)s (2009) (cid:133)nding that most of the yearto-year changes in U.S. trade values occurred along the intensive margin associated with surviving products, rather than the extensive margin associated with product-adding and dropping. 7. Applicability of the Algorithm to Other National and International Product Classi(cid:133)cation Systems The algorithm described in this article can be used to create a concordance for any product classi(cid:133)cation system over time so long as the associated statistical agency periodically makes available mappings of obsolete and new codes. Given this information, the process of assigning product codes to families will be identical to that described above, and it should be fairly simple to adapt our Stata code to cover any idiosyncrasies. For example, the algorithm could be applied to other international trade product classi(cid:133)cation systems such as the European Union(cid:146)s Combined Nomenclature (CN) codes. Changes to the CN are published annually in the L-series of the O¢ cial Journal of the European Communities. Application of our method would permit evaluation of the EU(cid:146)s product-level exports and imports on a consistent basis over time. Moreover, it is possible to apply the algorithm to more aggregated levels of international trade product classi(cid:133)cation systems, such as the 6-digit HS codes de(cid:133)ned by the WCO. Our algorithm can also be applied to track changes in production-based industry classi(cid:133)cation systems such as NAICS (North America) or NACE (EU). The U.S. Census Bureau, for example, publishes correspondence tables for the various revisions to NAICS, and these can be used to identify (cid:147)families(cid:148)of industry codes over time. The analogous information for NACE is published by Eurostat with each NACE revision. 8. Conclusion Controlling for changes in product codes over time is critical in the growing body of research examining (cid:133)rms(cid:146)product-mix choices. In this article, we present a concordance algorithm that can be used to track changes in product codes and generate time-consistent (cid:147)synthetic(cid:148)codes. We use this algorithm to generate the (cid:133)rst complete concordance of changes in U.S. HS codes over time. We also describe the prevalence of changes in HS codes over time, underscoring the importance of controlling for these changes in empirical research. Lastly, we provide an electronic appendix containing the (cid:133)nal concordance (cid:133)les, as well as Stata code that can be used to customize this and other product code concordances.

11 References Bernard, A.B., Redding, S.J., and Schott, P.K. (2010). Multi-Product Firms and Product- Switching. American Economic Review, 100, 70-97 Bernard, A.B., Redding, S.J., and Schott, P.K. (2011). Multi-Product Firms and Trade Liberalization. Quarterly Journal of Economics, 126, 1271-1318. Bernard, A.B., Jensen, J.B., Redding, S.J., and Schott, P.K (2009). The Margins of U.S. Trade (Long Version). NBER Working Paper 14662. Goldberg, P.K., Khandelwal, A.K., Pavcnik, N., and Topalova, P. (2010). Imported Intermediate Inputs and Domestic Product Growth: Evidence from India. Quarterly Journal of Economics, 125, 1727-1767. Pierce, J.R. (2011). Plant-Level Responses to Antidumping Duties: Evidence from U.S. Manufacturers. Journal of International Economics, 85, 222-233.

12 A Appendix This appendix describes the (cid:133)les contained in the electronic appendix available online at: http://www.som.yale.edu/faculty/pks4/sub_international.htm. All (cid:133)les are contained in a zip folder with (cid:133)lename hs_concordance_20101020.zip. A1. Stata Programs That Create the HS-Over-Time Concordances The (cid:133)les hts.do and schedule_b.do contain our algorithm for creating import and export HS concordances, respectively, forarbitrarybeginningandendingyear-monthsbetween1989and2009. Those comfortable with Stata programming should (cid:133)nd these (cid:133)les relatively easy to manipulate. Those unfamiliar with Stata programming can instead use one of the output (cid:133)les described below. A2. A Stata Program To Match the HS-Over-Time Concordances to U.S. Trade Data The (cid:133)le trade_merge.do is a Stata program that matches our HS-over-time concordances to publicly available U.S. trade data. Researchers may (cid:133)nd this example useful when employing the concordances in their own research. In addition, this Stata program produces some of the output (cid:133)les described below. A3. A File Tracking (cid:147)Raw(cid:148)HS Code Changes Each Stata program requires as an input a data (cid:133)le containing the raw obsolete-new mappings discussed in the main text. These input (cid:133)les are named sch_b_concordances_20100522_02.dta and hts_concordances_20100522_02.dta, respectively, where 20100522 is the user-de(cid:133)ned version date. The basic structure of these input (cid:133)les resembles the raw obsolete-new (cid:133)les; i.e., each set of obsolete HS codes is followed by the new set of HS codes into which they map. In this sense, researchers who wish to examine a simple record of changes to HS codes, as reported in the o¢ cial obsolete-new releases may (cid:133)nd these (cid:133)les useful. The (cid:133)les contain the following variables: obsolete: old HS codes that become obsolete as of e⁄ective date; (cid:15) new: new HS codes replacing the obsolete codes; (cid:15) setyr: synthetic code to which new and obsolete codes belong, as de(cid:133)ned in main text; and (cid:15) e⁄yr: date the mapping is e⁄ective. (cid:15) A4. Concordances for Changes in Import and Export HS Codes from 1989 to 2009 The Stata programs described above produce the output (cid:133)les that can be used to concord HS codes in U.S. import and export data Speci(cid:133)cally, the code produces output (cid:133)les: sch_b_concordances_20100522_BEG_END.dta, and (cid:15) hts_concordances_20100522_BEG_END.dta (cid:15) where BEG and END re(cid:135)ect beginning and end years (exports: 1989_2009) or year-months (imports: 198906_200907), respectively. These concordances include the same variables as the input (cid:133)les, but with setyr and e⁄yr standardized across family trees, as described in Section 4 above. Variables in the concordance output (cid:133)les include:

13 obsolete: old HS codes that become obsolete as of e⁄ective date; (cid:15) new: new HS codes replacing the obsolete codes; (cid:15) setyr: synthetic code to which new and obsolete codes belong, as de(cid:133)ned in main text; and (cid:15) e⁄yr: year (export) or year-month (import) in which the particular obsolete-new mapping (cid:15) (cid:133)rst appears in the raw data. A5. Simple Versions of the Concordances for Changes in Import and Export HS Codes from 1989 to 2009 The (cid:133)les simple_hts_198906_200907.dta and simple_schedule_b_1989_2009.dta provide the setyear for all HS codes that have experienced changes between 1989 and 2009 for imports and exports, respectively. The (cid:133)les have a simple two-column format where the (cid:133)rst column reports the HS code that has experienced a change between 1989 and 2009 and the second column provides the setyear for that HS code. Researchers can merge this (cid:133)le by HS code with product-level trade data and easily assign a setyear to any HS codes that have been changed. HS codes not appearing in these output (cid:133)les are consistent across all years of the data. In almost every case, this simple concordance is one-to-one, in the sense that each HS code maps to a single setyear. However, six (two) HTS (Schedule B) codes were listed as obsolete in one year and then "reappeared" as new codes in a later year with a di⁄erent setyear. Each of these HS codes, therefore, has two setyears. The dates given in the setyear indicate the years in which they became active. These duplicate HS codes are: HTS - 2905492000, 5112196010, 5112196020, 5112196040, 5112196050, 7304390040; Schedule B - 481190900, 9027501000. A6. A Record of HS Codes Associated with Each (cid:147)setyear(cid:148)Synthetic HS Code The (cid:133)les setyr_x_1989_2009.dta and setyr_m_1989_2009.dta, provide a record of every HS codeassociatedwitheverysetyearthatappearsinthe1989-2009concordeddata. The(cid:133)rstcolumn of each (cid:133)le lists the setyears, sorted from low to high. Each additional column lists the actual HS codes appearing in a particular year of the trade data that should be replace by the setyear. These actual HS codes also are sorted from low to high in each year. To concord U.S. trade data from 1989 to 2009, one would just replace all codes listed in the table with the synthetic setyear, and then collapse the data according to these setyears. HS codes not appearing in these output (cid:133)les are consistent across all years of the data.