11

Are Customs Records Consistent Across Countries?

Paper Authors: C. J. Krizan (U.S. Department of Labor), James Tybout (The Pennsylvania State University, National Bureau of Economic Research [NBER], and CESifo), Zi Wang (Shanghai University of Finance and Economics), and Yingyan Zhao (George Washington University)

Presenter: James Tybout (The Pennsylvania State University, NBER, and CESifo)
Moderator: Andreas Moxnes (University of Oslo)

The workshop paper by C. J. Krizan, James Tybout, Zi Wang, and Yingyan Zhao studies firm-to-firm datasets to understand their accuracy for formal hypothesis testing. Accurate data sources are vital for making informed policy choices.

James Tybout, professor of economics at The Pennsylvania State University and member of the National Bureau of Economic Research and CESifo, introduced the workshop paper by motivating the importance of accuracy in measurement of trade statistics and customs data. Researchers have used customs records to understand international trade from the perspective of the firm. This includes understanding a firm’s market entry and exit decisions and market share development. These datasets are also used to study the patterns of diffusion of technology across global value chains (GVCs) and multinational enterprise (MNE) networks. The literature on these topics relies on micropatterns in the data to understand these complex issues. Some examples of these studied patterns are firm-to-firm connections and the duration of relationships in GVCs; the frequencies and patterns in shipments, product classifications, and industries; and the value of traded goods.

According to Tybout, the work presented contributes to a literature on the accuracy and reliability of customs data and improves researchers’ understanding about the dimensions under which the data perform the best. Tybout and his coauthors compared the export records of shipments leaving Colombia with the import records collected when the shipments enter the United States. The analysis was done at the aggregate, industry, firm, and transaction levels. He highlighted that, as expected, the results are more accurate at the aggregate and industry levels versus the firm or transaction levels. He and his

coauthors offer alternative explanations for the observed discrepancies, possible implications for research in the field, and some suggestions for improvements in tracking records.

While the paper studies only the Columbia–United States customs data discrepancies, these discrepancies are also true for trade across other U.S. trade partners. According to a 1996 U.S. Census report, the discrepancy between U.S. and Australian trade is around 4–7 percent. Orsini and dos Santos (2015) show that the gap for Brazil and the United States is between 11 percent and 17 percent, while the 2012 U.S. Census estimates the gap between the United States and China to be somewhere between 22 percent and 48 percent. Kellenberg and Levinson (2019) find large variation across 126 countries.

LITERATURE REVIEW

Tybout started by explaining how this work fits in the related literature:

• Record matching: The literature on record matching shows that it is a powerful tool (see Christen [2012] for an overview).
• The Longitudinal Firm Trade Transactions Database (LFTTD): This database, developed in Bernard et al. (2009) and further curated in Barresse et al. (2017), Kamal and Monarch (2018), and Kamal and Ouyang (2020), contains the U.S. Customs data.
• Trade reconciliation studies: This literature, which includes U.S. Census (1996, 2012), Orsini and dos Santos (2015), Fisman and Wei (2004), Mishra et al. (2008), Stoyanov (2012), Ferrantino et al. (2012), Javorcik and Narciso (2017), and Kellenberg and Levinson (2019), seeks to understand the variation and discrepancies in trade statistics; however, none perform this analysis at the transaction level.
• Matched customs records studies: A very large body of literature uses matched customs records—Eaton et al. (2014), Blum et al. (2010, 2018), Bernard and Dhingra (2015), Dragusanu (2014), Kamal and Sundaram (2014), Sugita et al. (2019), Monarch and Schmidt-Eisenlohr (2017), Bernard et al. (2018b, 2018c), Carballo et al. (2018), Monarch (2019), and Helper and Munasib (see Chapter 7 of this volume).

STYLIZED FACTS AND TRENDS

Tybout displayed the reported exports from Colombia and reported imports from the United States from 2007 to 2013, as well as the percentage difference between the two measurements (see Figure 11-1). Prior to 2011, Colombia reported higher trade than the United States, reaching about 12 percent in 2009. Following 2011, the United States increasingly reported more trade than

Columbia, reaching about 15 percent in 2013. Three industries—ceramics, nonknitted apparel, and knitted apparel—account for the majority of Colombia–U.S. trade. The industry-level analysis shows similar discrepancies between Columbian exports and U.S. imports. This suggests that the measurement issue is on a larger scale than a single industry.

According to Tybout, U.S. Census experts suggest that it is common for U.S. importers to split transactions for administrative purposes, which may cause concerns of double counting in the U.S. data. The U.S. LFTTD data suggest that there are about 8 percent more transactions reported by the United States than by Colombia. At the same time, the LFTTD reports 12 percent higher value for the United States. Because there are reporting discrepancies for both the number of transactions and the size, or value, of those transactions, it is unlikely that the discrepancy is just a matter of double counting by one country.

Another potential reason for record-matching discrepancies may be entrepot trade, Tybout explained, a situation in which goods from one country (e.g., Columbia) are first shipped to another country (e.g., Panama). The second country is then recorded as the destination, although the goods are, in reality, being routed to their final destination (e.g., the United States). If there are intermediate stops in the shipment of goods, exporters may not know the destination country and importers may not know the country of origin. Disentangling this effect is difficult. This problem has been explored in Ganapati et al. (2021), which reports that 80 percent of trade is indirect. Tybout and his

coauthors checked for this discrepancy by comparing the reported country of origin and the shipping country.

As Tybout explained, while he and his coauthors found evidence that entrepot trade is increasing and may contribute to the observed discrepancies, entrepot trade did not explain the growing gap in the aggregate trade series. The value of trade by shipping country and country of origin track very closely over the period in question.

METHODOLOGY

Tybout then described the work in the paper, which focuses on the types of misreporting and not the underlying incentives. He and his coauthors explored whether customs records accurately characterize firm-to-firm trade patterns, and if the records match the correct firms in a relationship and their accuracy on the transaction level. They wanted to know whether particular firms or industries are driving the trends in misreporting. Finally, their paper explores the implications of the misreporting patterns for using this data for other questions.

Tybout went on to explain that the two methods for matching importers and exporters in the customs data include an importer identifier (the name and address of the firm) and an exporter identifier (the manufacturer ID [MID]—a string of name and address—for the U.S. data, and NIT—a form of tax ID—for the Colombian data). The authors used the importer identifier since there is less data loss in the matching process. The exporter identifier (MID and a pseudo-MID), is examined briefly but its use was not the preferred approach. Once firms are matched, the authors move to linking the transactions between the pairs.

The authors used a three-step process to merge the LFTTD and Colombian data:

• First, a standard data-cleaning process, developed for the creation of the LFTTD data by the U.S. Census Bureau, was conducted.
• Second, the first matching stage links the two datasets on the firm level using only the names and addresses. In order to reduce computation time, the information on transactions was not used. The result is a list of importer–exporter pairs in which both agents listed the other as a partner in a period, with some noise.
• Third, firm pairs were matched on their transactions by using the industry, product, value, and date as unique identifiers. The resulting dataset contains a list of transactions for each trading pair in a roughly consistent manner.

RESULTS

Tybout continued with a discussion of the results report in this workshop paper. The LFTTD database identifies 9,400 U.S. importers, but only 2,500 of the

importing firms are matched between the two databases. However, the 2,500 firms account for about 85 percent of the value of exports and about 73 percent of all transactions, providing robust coverage of a large amount of the trade. Looking at the data from the Colombian side, the 2,500 firms account for 95 percent of the transactions and about 97 percent of the value of trade. These numbers are the aggregate results. At the transaction level, only about 27 percent are matched in the LFTTD data and about 28 percent in the Colombian data. In terms of trade value, about 48 percent in the LFTTD data and about 52 percent in the Colombian data are matched. This means that the matched transactions of the matched firms are less than 30 percent of the total transactions and about 50 percent of the value of exports to the United States.

Over time, Tybout said, the value match is consistently higher than the transaction match. The Great Trade Collapse, which forced smaller exporters to exit, increased the match rate for both value and transactions because smaller exporters were less likely to match in the data.

Match rates also vary with industry type. The authors focused on the three largest categories of Columbian exports to the United States: ceramic products (HS 69),¹ nonknitted apparel (HS 62), and knitted apparel (HS 61). The match rate is highest for the ceramics industry, though it is still below 50 percent. Tybout offered a possible explanation, stating that these industries, especially the apparel industries, have many small firms with high turnover, so they are less likely to match. These trends are confirmed on the value level, but match rates are slightly higher across the board.

Matching also depends on whether firms are related or at arm’s length, explained Tybout. Using the LFTTD data, related parties are more likely to match on transactions. Related parties are less likely to match on the volume dimension. This pattern is reminiscent of Bernard and colleagues (2006), who found a similar pattern when studying transfer pricing.

Tybout offered two other, disconnected trends:

• Larger-value transactions are more likely to match than smaller-value shipments.
• The wholesale retailers match better than nonwholesale retailers when matching on transactions. The two groups are similar when matching only on value.

IMPLICATIONS OF RESULTS

Tybout stated that these discrepancies mean that research using firm-to-firm network GVCs could find imaginary links in networks and may be mismeasuring the duration of relationships. Researchers may also be mismeasuring entry, search, and learning costs associated with analysis of

___________________

¹ HS refers to the harmonized system codes, a standardized numerical method of classifying traded products used by customs authorities around the world.

exporters, in addition to mismeasuring interaction effects in analysis of technology diffusion.

Government statistical agencies should be concerned about these discrepancies in the data because these data are used to enforce commercial policy. If the data are poor, bad actors abroad will not be easily identified. There is a clear benefit to increasing the accuracy of firm-to-firm and transaction-level data.

BETTER IDENTIFIERS

Tybout presented some possible alternative identifiers for transaction-level data. Shipment invoice numbers are issued by the seller to the buyer, and these invoices list characteristics such as product and price. This is the main documentation of a sale between a firm pair. U.S. import declarations require an importer to verify and attest to records accuracy, but there may not be a similar requirement of information on the export declaration in foreign countries.

Another proposed identifier Tybout discussed is the bills of lading (BOL) number, which is issued by shipment carriers. BOLs establish a receipt of the product and contain evidence of title of ownership of the products, although these records refer to multiple invoices and identify a container rather than a shipment, and may include multiple containers. In the United States, the BOL can be used to identify the importer, but the foreign customs records may or may not identify an importer.

Tybout went on to say that the two proposed identifiers require global coordination between governments, which may not be feasible. An alternate identifier from the private sector would be Dun and Bradstreet numbers, or other similar firm-level identifiers. While there are major issues with these identifiers, collecting data on subsamples could still be valuable and provide a sample of true and false matches to compare with other datasets. This would open the possibility of matching algorithms based on a training sample.

DISCUSSION

Discussant: Jeronimo Carballo (University of Colorado Boulder)

Jeronimo Carballo, assistant professor of economics at the University of Colorado Boulder, began by motivating why the work is important to research in international trade and GVCs, which both play a defining role in the world economy. This literature depends on the use of these firm-to-firm databases, so accurate measurement is very important. This paper is the first to document the accuracy of firm-to-firm transaction data. The authors explore the consistency of data along dimensions, taking their study beyond trade value by examining transaction and firm links. This difficult task has much value. The main takeaway is that firm-to-firm datasets using names and addresses to match must be used with caution, but datasets that match on tax ID may still be valuable.

Carballo first noted that these discrepancies exist for other countries and data sources beyond the U.S.–Colombian relationship. For example, Italian figures are 10 percent higher in Eurostat than the figures published by the Organisation for Economic Co-operation and Development, and Canada reports $20 billion more exports to the United States than the United States reports as imports. The discrepancy is apparent in the data on trade for many countries and the United States, both positive and negative.

Carballo wondered if there is anything unique about the United States relative to the rest of the world. U.S. firms are identified using employer identification number (EIN), a tax identification number. A downside to this identifier is that firms can, and often do, have several EINs, making matching difficult. The Census Bureau has firm-level EINs only for income and payroll tax filings, while a firm may use a separate EIN or completely different identifier, such as a Social Security number or some foreign identifier, for trade shipments. He wondered how many of these discrepancies can be explained by firms having multiple EINs, and suggested that looking into differences between single-unit and multiunit firms may offer some insight into this concern. MNEs can have complex ownership structures, exacerbating the concerns about multiple EINs and obscuring which affiliate is reported in each database. To explore this further, the authors may be able to use Compustat-FirmID to understand if this problem is a larger concern for MNEs.

The LFTTD database contains information on two firms in each import: the ultimate consignee (the firm that takes final control of the good) and the importer. For about 20 percent of transactions in the United States, the importer is different from the ultimate consignee, and understanding what entity is reported in the foreign database is important, as this may offer an area from which missing links can be recovered.

Carballo also suggested that the authors explore the port of entry to test whether entrepot trade is the source of the discrepancy. The West Coast of the United States may be more sensitive to this effect than the East Coast. He wondered if the authors have data at the port level that would allow this style of analysis, namely if the Colombian data have the same accuracy as the LFTTD. In a similar vein, he wondered if the results vary by mode of transportation as transshipment is more common for maritime trade.

Carballo also wondered if discrepancies vary based on the mode of transportation and whether relaxing the temporal dimension would increase matching rates. Maritime trade may potentially take several days, while air freight may occur in one. Further, shipments may wait at a warehouse or port, and this would cause delays that may impact the time linages.

Finally, Carballo offered several minor comments and questions:

• Why did the authors drop the agriculture and mineral products?
• The cutoff is different for exports and imports in the United States; are the patterns similar for each?

• What are the consequences of matching at the transaction level using a many-to-many approach?

Tybout responded by saying that he and his coauthors dropped agriculture and mineral trade because the numbers failed to make sense. Trade in oil will sometimes appear as an export from the United States to the United States, and there is a central coffee exporting firm in Colombia, so the unique identifier is lost. He also responded that they tested the temporal dimension, but it did not appear to be very important. In the future, they will explore mode of transportation and ports of entry.

C. J. Krizan, director of data analytics at the U.S. Department of Labor and a coauthor of the presented work, commented that he was at the Census Bureau for 15 years and was very aware of the concerns with the EIN as an identifier. The Census Bureau does try to build a unique firm identifier that collects all the EINs a firm may have. In this project, the authors attempted to match on this unique firm identifier, but this did not have a large impact on the results.

One audience member asked about differences in the definition of a transaction, especially in the textile industry. Tybout responded that the authors aggregated all transactions in a firm pair, and they still did not recover a high-quality match. Krizan added that splitting shipments is common for U.S. firms; this may impact the match rates.