Policy Research Working Paper 10112 Trade Fraud and Non-Tariff Measures Hiau Looi Kee Alessandro Nicita Development Economics Development Research Group June 2022 Policy Research Working Paper 10112 Abstract Similar to tariffs, non-tariff measures may induce trade variations. It presents a theoretical model and empirical evi- fraud when they are restrictive. This paper examines dence showing that discrepancies increase with ad valorem whether discrepancies observed in the official trade statistics equivalents, consistent with the trade fraud due to traders of importing and exporting countries are partly due to trade intentionally mis-declaring countries of origin or misclassi- fraud from evading border non-tariff measures. To capture fying products in order to evade border non-tariff measures. the restrictiveness of non-tariff measures, the paper estimates The results are driven by homogeneous products and the the ad valorem equivalent with importer-exporter-product trade between developed and developing countries. This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at hlkee@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Trade Fraud and Non-Tari¤ Measures Hiau Looi Keey Alessandro Nicitaz World Bank United Nations Key Words: trade discrepancies, non-tari¤ measures, ad valorem equivalent of NTMs, tari¤ evasion, trade fraud JEL Classi…cation Numbers: F10, F14 We thank Pol Antras, Erhan Artuc, Chad Bown, Dave Donaldson, Davin Chor, Ana Fernandes, Michael Ferrantino, Penny Goldberg, Russell Hillberry, Aaditya Mattoo, Fernando Perro, Bob Rijkers, and the seminar participants of the IMF/WB/WTO Joint Trade Workshop, National University of Singapore, SAIS and George Washington University for their comments. Research for this paper has in part been supported by the World Bank’ s Multidonor Trust Fund for Trade and Development. The authors declare that they have no relevant or material …nancial interests that relate to the research described in this paper. The results and opinions presented in this paper do not represent the views of our institutions, the Executive Directors, or the countries they represent. y Development Research Group, The World Bank, Washington, DC 20433, USA; Tel.: (202) 473-4155; Fax: (202)522-1159; e-mail: hlkee@worldbank.org. z United Nations Conference on Trade and Development. Email: Alessandro.Nicita@un.org. “...buyers do not allow any ‘Made in Bangladesh’label on the [garment] product ...they [requested] label these very products as ‘Made in India’... ” –Prabir De, “Non-Tari¤ Measure Study: Bangladesh, India and Nepal,”2016 “Five individuals and two domestic honey-processing companies have been charged with federal crimes... The charges assert that the Chinese-origin honey was mis- declared as other commodities upon importation into the United States and trans- shipped through other countries to evade anti-dumping duties.” –World Customs Organization, Illicit Trade Report, 2012 “Professional Fraud Facilitators... provide a complete fraud package to EU im- porters and Chinese exporters, with delivery of goods to EU port, change of con- tainer in transit country and documented new origin for goods (new B/L, in- voices, origin certi…cates).” - European Anti-Fraud O¢ ce, Customs Origin Fraud and Professional Fraud Facilitators, 2014 1 Introduction In international trade, importing and exporting countries often record starkly di¤erent cus- toms statistics for the same shipments, mostly with imports signi…cantly larger than exports. For example, in 2018, the US recorded US$563 billion in overall imports from China, while China recorded US$480 billion in exports to the US.1 The result is a trade discrepancy of nearly US$83 billion, or 15 percent. Not only is the presence of such trade discrepancies puz- zling, it also distorts the true picture and may lead to public misperception and ill-informed policy decisions. At a detailed product level, the discrepancies in customs statistics are even 1 These …gures are from national statistics of the US and China. 1 more staggering and widespread across countries and products, both in terms of values and quantities. Figure 1 presents the distribution of trade discrepancies for a sample of more than 40 importing countries.2 At the sample mean, the trade discrepancy is about 85 per- cent –suggesting that most import records are much larger than the export records, for both quantity and value. What could cause such discrepancies in the o¢ cial trade statistics? There are several contributing factors. The discrepancies could be at least in part due to the di¤erent valu- ation methods applied by customs agencies, which normally include freight and insurance costs in import values but not in export values.3 Another contributing factor is that the transshipment and re-export of products going through third countries cause a mismatch in the country of origin.4 The discrepancies could also result from trade fraud in the form of undervaluation in imports in order to evade import tari¤s,5 or the undervaluation in exports in order to evade export taxes or VAT.6 The analysis of trade discrepancies has also been increasingly used to detect tari¤ evasion, especially by the customs authorities of developing countries.7 However, even taken together, these contributing factors cannot fully explain the large discrepancies in the data, which tend to vary across products and trading partners, particularly with the recorded import quantity larger than the export quantity. The objective of this paper is to assess whether part of the large discrepancies in the o¢ cial statistics is actually trade fraud due to the avoidance of costs associated with the non-tari¤ measures (NTMs) of the importing countries.8 Such trade fraud may not only 2 In this paper, trade discrepancies are de…ned according to Equation (8) ; which measures the percentage di¤erence between the importer’ s and exporter’ s statistics. Figure 1 plots the distributions of the value discrepancies (gapvsh) and quantity discrepancies (gapsh) at the HS 6-digit product level across countries. Please refer to Table 1 for the list of importing countries. 3 Hummels and Lugovskyy (2006). 4 Ferrantino and Wang (2008). 5 Bhagwati (1964), Fisman and Wei, (2004), and Javorcik and Narciso (2008). 6 Ferrantino, Liu and Wang (2012). 7 Cantens (2015); Grigoriou, Kalizinje and Raballand (2019); Grigoriou(2019); Chalendard, Fernandes, Raballand and Rijkers (2020). 8 This paper focuses only on large discrepancies at the HS 6-digit level, in which the di¤erence between 2 reduce government revenues but may also undermine the trade regulatory framework of the importing country, and ultimately alter the competitive environment by resulting in an un- fair advantage gained by the fraudsters over their law-abiding competitors. By avoiding NTMs, trade fraud may jeopardize public health or safety, and weaken environmental pro- tections. The three examples cited above in italics highlight such fraud, whereby traders either misreport country of origin or product codes to evade NTMs. We focus on the NTMs applied at the border, hereafter referred to as Border NTMs.9 Examples of Border NTMs include customs controls, quota licensing, pre-shipment inspections, and additional fees paid at customs, among many others. Some of these NTMs are bilateral in nature, targeting speci…c products from speci…c partner countries. Compliance with these requirements is often …nancially costly and/or time-consuming,10 therefore providing incentives for traders to commit trade fraud by misrepresenting the country of origin or misclassifying product codes on their customs documents (Li and Lin, 2022). To study the role of Border NTMs in causing trade fraud, this paper …rst provides a simple theoretical model built on Swenson (2001) and Ferrentino, Liu and Wang (2012) to show that the restrictive NTMs cause importers and exporters to under-report the quantity traded on their respective customs declaration forms. However, given that the penalty and detection rates of the importing countries are generally higher than those of the exporting countries, importers will under-report less than exporters hence giving rise to the observed trade discrepancies. The theoretical model also shows that trade discrepancies increase with the ad valorem equivalent (AVE) of NTMs. Furthermore, the model shows that the e¤ect of AVE on trade discrepancies is larger when products are homogeneous and when the exporting countries are developing countries. In addition, the model rationalizes origin and the import value or volume and the export value or volume is more than 40 percent, in order to side step variations in the valuations and small statistical errors. 9 For more details see Ederington and Ruta (2016) and UNCTAD (2015). 10 See UNCTAD, World Bank (2018). 3 product classi…cation fraud whereby exporters misdeclare the country or misclassify products in their customs declaration forms. Finally, the model shows that the relationship between trade discrepancies and tari¤s depends on the e¤ect of tari¤s on the detection rate of the importing countries. The empirical section of the paper heeds the message of Goldberg and Pavcnik (2016): “Measure (trade barriers) before you estimate (trade impacts)! ”We …rst led the NTM data collection e¤orts of the UNCTAD and the World Bank which allow us to generate an NTM database of 50 importing countries at the detailed 6-digit Harmonized System (HS) level with bilateral variation across 96 trading partners.11 We then rely on product level quantity- based gravity regressions to estimate the bilateral AVE of NTMs at the HS 6-digit level. Our estimated AVEs are reasonable and in line with previous studies, with a sample mean of 4.7 percent when NTMs are present. Using the estimated AVEs, this paper then assesses whether the presence of Border NTMs contributes to large trade discrepancies as shown in the theoretical model. The results show that exporting countries that face higher AVEs have larger trade discrepancies, consistent with the hypothesis of country origin fraud in order to gain more favorable market access conditions. Similarly, the regression results also demonstrate that the products that have higher AVEs are those that have larger trade discrepancies, consistent with the hypothesis of product classi…cation fraud with the purpose of reducing the costs of compliance with Border NTMs. These results are consistent with the theoretical model and are robust to simultaneity and measurement errors in AVEs, which we address with the use of instrumental variables. Overall, this paper …nds that a 10 percentage point increase in the AVEs could lead to a 15 percent increase in trade discrepancies, while a comparable increase in tari¤s leads to only an 8 percent increase in trade discrepancies. These …ndings suggest that a signi…cant part of trade discrepancies is due to trade fraud associated 11 This is done by working with local consultants in identifying and codifying countries’laws and regulations that may a¤ect international trade. Some of these laws and regulations target speci…c products or countries. 4 with NTM avoidance schemes. Robustness checks further reveal that NTM-induced trade fraud is more common for homogeneous products, as well as for trade between developed and developing countries, as predicted by the theoretical model. The rationale of this paper to investigate trade fraud relates to the existing literature on trade discrepancies and tari¤ avoidance. In particular, this paper follows on the “missing trade” approach initially proposed by Fisman and Wei (2004) to explain discrepancies in the statistics between Hong Kong SAR, China, and China. This approach was also used by Javorcik and Narciso (2008) whose examination of the exports statistics of Germany …nd that discrepancies with the statistics of the importing country tend to be greater when tari¤s are higher. Similar results are found for India by Mishra, Topalova and Subramanian (2008). In addition, focusing on Chinese exports to the United States, Ferrantino, Liu and Wang (2012) show that Chinese exporters may under-report export values in order to avoid paying s VAT, thereby helping to explain why reported export values are lower than the China’ corresponding reported import values. Most recently, Demir and Javorcik (2020) also show import duty evasion as a margin through which …rms adjust to changes in trade policy in Turkey, while Liu, Sheng and Wang (2021) show that trade data over-reporting is positively correlated with the exchange rate spread in China. This paper adds to the literature by looking at NTMs as another source of costs at the borders, which may result in similar schemes employing misdeclaration and misclassi…cation with the purpose to reduce costs. In s imports or exports, addition, unlike previous papers that mostly focused on one country’ this paper uses detailed trade data from a wide range of importing and exporting countries, which allow us to detect customs origin fraud and product fraud. Importantly, this paper also contributes to the literature on assessing the e¤ect of trade policies, particularly non-tari¤ measures, on trade. In particular, this paper builds on the work of Kee, Nicita and Olarreaga (2009) estimating the AVE of non-tari¤ measures by 5 providing a novel method to account for possible heterogeneous e¤ects of NTMs across trading partners while employing econometric techniques in the estimation of AVEs that relate to the literature on the estimation of gravity models. Unlike previous papers, the AVE estimates of this paper are highly detailed and disaggregated at the HS 6-digit level and vary by importer and exporter. Such detailed variations could be useful for studies evaluating the impact of speci…c trade policy changes. The paper proceeds as follows. Section 2 provides theoretical motivations to de…ne trade discrepancies. Section 3 provides a data description for the paper. Section 4 presents the empirical model to estimate the AVEs of NTMs. Section 5 presents the results on the estimation of AVEs. Section 6 shows the empirical evidence for NTM-induced trade fraud, while robustness checks are presented in Section 7. Section 8 concludes. 2 Theoretical Motivations This section provides a simple theoretical model built on Ferrantino, Liu and Wang (2012), which followed Swenson (2001). Those two studies emphasize the importance of transfer pricing, export taxes such as VAT and capital controls in a¤ecting the under-reporting behavior of importers and exporters from China and the US. Di¤erent from the analysis of those two papers, we focus on a wider range of importing and exporting countries with di¤erent levels of development and institutions, where transfer pricing, VAT and capital controls may not be important. Instead, we emphasize the roles of tari¤s and NTMs in inducing importers and exporters to under-report the quantity and value of trade di¤erently, which leads to the observed trade discrepancies in the o¢ cial statistics. Similarly to the previous models, traders weigh the bene…ts of under-reporting, which are the tari¤s saved, against the costs of under-reporting, which are the potential penalties, to choose the optimal 6 level of trade fraud in terms of under-reporting. In the model of this paper, in addition to the potential tari¤ evaded, which depends on trade value, under-reporting also allows traders to evade restrictive NTMs, such as inspections and testing requirements, which generally depend on the quantity traded. While the NTMs are only imposed by the importing countries, they may a¤ect both the importing and exporting …rms. Nevertheless, di¤erences in the penalty rates and detection rates of the importing and exporting countries give rise to the observed trade discrepancies. The following model focuses on under-reporting in quantity traded. Let index k = fX; M g indicate exporting or importing …rms. Exporting and importing …rms submit respec- tive customs declaration forms for the same shipment, but may declare di¤erent quantities, QX and QM ; in order to minimize their trade costs. For the exporting …rms, their exporting costs include the importing countries’NTM related transaction costs, such as pre-shipment inspection, testing requirements, conformity certi…cates and other obligations necessary to comply with the requirements of the importing country. These costs depend on the quan- tity and are captured parsimoniously by the AVE of NTM, multiplied by the size of the shipment, P QX : While these costs provide an incentive for exporting …rms to under-report export quantities, exporting …rms may also face penalties if they are detected to have under- reported in their export custom forms. The overall penalty cost depends on the penalty X rate, ; the detection rate, sX ; the squares of the size of deviation, Q QX =Q; and the true value of the shipment, P Q:12 Each exporting …rm will therefore choose to declare QX to minimize the total exporting costs, C X : X X 2 s Q QX minC X = AV E P QX + P Q: (1) QX | {z } 2 Q NTM-induced Cost | {z } Penalty Cost 12 Here we follow Ferrantino, Liu and Wang (2012) to model the cost of exporting depending on the squares of the size of deviation. 7 Under-reporting, such that Q QX > 0; implies a lower N T M -induced cost, and also a higher penalty cost. At the equilibrium, the following …rst-order condition must hold: @C X X X Q QX = AV E P s P = 0 ==> @QX Q Q QX AV E X X = X sX > 0 if s > 0 or (2) Q QX AV E = 1 X sX : (3) Q Thus, if the penalty and the detection rates are positive, it is always optimal for the exporting …rm to under-report, such that Q QX > 0. According to equation (2) ; the size of under- s NTMs, and decreases reporting increases with the cost of satisfying the importing country’ with the penalty and detection rates of the exporting country. If the exporting country does X X not impose penalties for under-reporting, such that s = 0; which is generally the case empirically as custom administrations may not have incentives or resources to accurately monitor exports, then @C X = AV E P > 0 ==> (4) @QX QX = 0: Equation (4) suggests that, in the absence of accurate controls of export shipments or penalty charges, C X increases with QX : Hence, to minimize C X ; the optimal quantity to declare in the export customs form is 0: This is the case when the exporter commits either origin fraud by mis-declaring the country code by trans-shipping through a third country facing lower border costs associated with NTMs, or product classi…cation fraud by misclassifying the product code as to face lower compliance costs with border NTMs. In both cases, the quantity of the true origin or true product code will be 0. 8 On the importers side, the per unit NTM-induced costs may include import licensing requirement and inspection, simply represented by the AVE of NTM. In addition, importing …rms may also face a tari¤, t; on the total value of the shipment. The importing …rm will therefore choose to declare QM to minimize the total importing costs, C M : 2 M M M M M Q QM minC = t P Q + AV E P Q + s P Q: (5) QM | {z } | {z } Q Tari¤ cost NTM-induced cost | {z } Penalty cost The …rst-order condition implies that @C M M M Q QM = t P + AV E P s P = 0 ==> @QM Q Q QM t + AV E M M = M sM > 0 if s > 0 or (6) Q QM t + AV E = 1 M sM (7) Q Thus, equation (6) shows that the optimal size of under-reporting for the importers increases with the level of tari¤, or the cost of complying with N T M s, but decreases with the penalty rate or detection rate of the importing country. In other words, the declared quantity QM will deviate from its true value Q depending on the level of tari¤ and the cost of compliance with NTMs, subject to the penalty and detection rates. Given that importing country governments need to collect tari¤ revenue from imports and assess whether the imported M M products meet their NTM requirements, s are expected to be positive. Realistically, penalties are larger and checks are more careful for the importing countries than for exporting countries so we assume that: M M X X s > s : A trade discrepancy, Disc; is de…ned as the recorded import quantity minus the recorded 9 export quantity, in percentage of the actual quantity: QM QX Disc (8) Q AV E t + AV E = X sX M sM M M X X X X AV E s s t s = X sX M sM : (9) Under the assumption that penalty and detection rates are higher in the importing coun- tries, Equation (9) suggests the discrepancy increases with the costs of complying with the NTMs of the importing countries: M M X X @Disc s s = X sX M sM > 0. (10) @AV E Intuitively, more restrictive NTM requirements of the importing country, which are parsimo- niously represented by the larger AVE, increase the trade costs for both the exporting …rms (pre-shipment inspection, testing and certi…cation requirements, etc.) and the importing …rms (import licensing requirement, cargo inspection, etc.). Therefore, both the importers and the exporters have incentives to under-report their quantity on their customs declara- tion forms in order to reduce such costs. However, the penalty and detection rates in the importing countries are normally higher, due to more accurate customs procedures driven by revenue collections and regulatory concerns, unlike the exporting customs that generally have fewer reasons to precisely assess shipments against the export declarations. This results in larger under-reporting by the exporters than by the importers, which implies larger trade discrepancies between the o¢ cial statistics of the importing and exporting countries. In addition, the rate of detection, s; may depend on a few factors. First, for the importing country, products that have higher tari¤s may be subjected to more careful inspection in 10 order to verify import duties. Thus, sM increases with t : sM 0 (t) > 0: Second, it is generally easier to detect fraud related to misclassi…cation of products or mis- declaration of country of origin for di¤erentiated products than for homogeneous products. For example, customs o¢ cers can easily tell a German car from a Japanese car, but may have a hard time identifying Chinese steel rods from Brazilian steel rods. Thus, s could be higher for di¤erentiated products than homogeneous products. This could lead to the e¤ect of NTMs on discrepancies to be smaller for di¤erentiated products than for homogeneous products if sX increases more than sM . Third, developed countries may have better customs administrations and resources to in- spect products, relative to developing countries. Thus, the di¤erences between the detection rates of the importing and exporting countries will be larger when the exporting country is a developing country rather than a developed country. This implies that the e¤ect of N T M s on discrepancies will be larger when the developed countries import from the developing countries. Finally, the e¤ects of tari¤s on trade discrepancies depends on the e¤ect of tari¤s on the detection rate: @Disc 1 AV E + t M 0 = M sM 1 s (t) 7 0; given sM 0 (t) > 0: (11) @t sM Equation (11) shows that, if the detection rate is very sensitive to tari¤s, then an increase in tari¤s could lead to a rise in trade discrepancies. This is because while the importing …rms may be increasing under-reporting when facing a higher tari¤, they risk being detected and face a much larger penalty. So in equilibrium, the size of under-reporting may actually 11 decrease with a higher tari¤ which implies a larger trade discrepancy, given a …xed level of export quantity. In summary, this simple model delivers the following testable hypothesis: 1. Trade discrepancies increases with the restrictiveness of NTMs imposed by the import- ing countries 2. Origin fraud and product classi…cation fraud exist, particularly for trade between high- income importing countries and developing exporting countries 3. The e¤ects of importing country NTMs on trade discrepancies are larger for trade between high-income importing countries and developing exporting countries 4. The e¤ects of importing country NTMs on trade discrepancies are larger for homoge- neous products than for di¤erentiated products 5. The e¤ects of tari¤s on trade discrepancies could be positive 3 Data The tari¤ and NTM data used in this paper are from the UNCTAD TRAINS database. Trade data originates from the UNSD COMTRADE database. The NTM data is collected only periodically, and the data utilized in this exercise was collected between 2015 and 2019. As NTM data generally varies little across time, the data on NTM is assumed to re‡ect the existing NTMs as of 2018. Consequently, trade and tari¤ data are matched to the same year. For the purpose of this paper, we focus on Border NTMs, while retaining other types of NTMs as control variables.13 Border NTMs include costs incurred by imported goods such 13 NTMs are distinguished between border and non-border variables on the basis of the international classi…cation of non-tari¤ measures (UNCTAD, 2019) and de…ned as custom measures by Ederington and Ruta (2016) 12 as customs controls, quota licensing, pre-shipment inspections, and additional fees paid at customs, among many others.14 Table 1 presents the list of importing countries included in our sample, together with their coverage ratios of border NTMs. There are a total of 49 importing countries with 96 exporting countries. The sample is largely determined by the availability of NTM data in the UNCTAD TRAINS database. There is a wide range of variations in terms of the coverage ratios across these di¤erent importing countries. We exploit this wide range of variations across countries to estimate the AVEs at the HS 6-digit level. Trade discrepancies are constructed by matching import data with export data at the HS 6-digit level according to Equation (8) for both quantity and value, based on UNSD Comtrade HS 6-digit level import and export data. Since we do not observe the true quantity, Q; we replace it with the average value of import and export quantity and value. The quantity data were thoroughly examined, and we include only observations in which the importer and the exporter use the same quantity unit in their reporting, to make sure that the discrepancies are not due to di¤erences in the units of measurement in quantity. In addition, the quantity units remain the same within any given HS 6-digit product.15 4 Estimation of the Ad Valorem Equivalent of NTMs In the theoretical model, the impacts of NTMs on trade discrepancies work through the ad valorem equivalent (AV E ) of NTMs. To empirically test the hypothesis, we will need estimates of these AV Es; which vary by product and importer-exporter pair. There is no such detailed estimates available in o¢ cial published statistics. We will need to estimate these AV Es: The closest estimates are provided by Kee, Nicita and Olarreaga (2009), at 14 Please refer to Ederington and Ruta (2016) for details. 15 For the regression analysis, we always control for product …xed e¤ects so that we do not rely on di¤erences in the quantity units across products to estimate the coe¢ cients of interest. 13 product and importer level, without variations across exporters. In this paper, to estimate the AV E of the NTMs, we run quantity-based gravity regres- sions for each of the nearly 5,000 HS 6 digit-level products for 2018. The dependent variable of the gravity regressions is the quantity traded of an HS 6-digit product between a coun- try pair, and the right-hand variables include a dummy variable indicating the presence of Border NTMs and the bilateral applied tari¤ on the product speci…c to the country pair. In these gravity regressions, the coe¢ cients of tari¤s can be interpreted as the trade elasticities, which are used to re-scale the coe¢ cients of the NTM dummy variables to obtain the AVE of NTMs. The dependent variable is treated as a discrete, non-continuous variable, and we use a count data regression technique such as Poisson or negative binomial models for our estimations.16 4.1 Econometric Issues There are some econometric issues that need to be addressed before we present our estimation model. First is the distribution of the dependent variable. For each HS 6-digit product, there is an excessive number of bilateral country pairs that do not trade. The share of quantity imported that is positive is only 6.8 percent in the sample. This leads to the large presence of zero observations in the bilateral data set as shown in Figure 2. Moreover, for the country pairs that do trade, the volume of their trade has a very large range that leads to an extreme over-dispersion as shown in Figure 3. Overall, the ratio of the variance of the positive quantity and its mean is 4.83e+08 in our data set. The large presence of zeroes together 16 This is necessary because nearly 25 percent of the observations have “Number of items” as the unit of measurement for quantity imported, which clearly is discrete in nature. The rest of the 75 percent of the observations have “Weight in kilograms” as the unit of measurement, which is often recorded as a discrete variable. Thus, nearly 100 percent of the observations recorded quantity imported as discrete integers. Therefore, it is appropriate to use a count data regression technique, such as Poisson or negative binomial models for our estimation. Furthermore, Santos Silva and Tenreyro (2006, 2011) showed that the Poisson pseudo-maximum likelihood (PPML) estimator usually used for count data works well even when they used the value of aggregate bilateral imports, which itself is not a count variable, as the dependent variable. 14 with over-dispersion led us to use the zero-in‡ated negative binomial (ZINB) model for the main estimation speci…cation. Nevertheless, for each HS 6-digit product, in addition to the ZINB, we also run the negative binomial (NB), the zero-in‡ated Poisson (ZIP), the ordinary least squares (OLS) in log and the Poisson models, and use the Vuong test, which is a model speci…cation test, to pick the best-…tting regressions for the AVE estimates. For ZINB, we assume that the observed trade data is the realization of the mixture of two distinct distributions: one distribution governs the probability of zero trade (participation or extensive margin), and the second distribution governs the realization of positive trade (volume or intensive margin). Let Qnij be the quantity of product n imported by country i from country j; and let h (Qnij ; jX) denote the negative binomial density with mean e (X ), dispersion parameter , while includes and : Here, X is a vector of variables to explain the positive import quantity, which are standard gravity variables, Zij ; as well as tari¤s and NTMs: 8 > < 0, with probability pnij Qnij ; > : Q ~ nij ; with probability 1 pnij ~ nij = 0; 1; 2; ::: Q h (Qnij ; jX) : Therefore, the import quantity density distribution can be modeled as 8 > < pnij + (1 pnij ) h (Qnij = k; jX) ; if k = 0 pr (k ) = (12) > : (1 pnij ) h (Qnij = k; jX) ; if k = 1; 2; ::: The parameter, pnij ;which is used to increase the presence of zeros in the data set, could depend on covariate W and is commonly modeled to follow a logit function to ensure its 15 range: exp (W ) pnij = 2 (0; 1) : (13) 1 + exp (W ) In general, the same set of variables could be included in both W and X and we do not need to identify any additional variables for the participation equation (see Cameron and Trivedi, 1998). Nevertheless, we still include a common religion indicator in W, motivated by Helpman, Melitz and Rubinstein (2008) to estimate Equation (13) : The second modeling issue is related to the estimation of AVE with bilateral variations. There are two main reasons why product level trade elasticities and AVEs may have bilateral variation across exporting countries for each importing country. First, the import demand of the products from di¤erent source countries are likely to have di¤erent price elasticities empirically.17 By incorporating bilateral variation into these trade elasticities, we will be better equipped to analyze the impacts of trade negotiations or agreements with speci…c partner countries. For our current context, the bilateral di¤erences in import demand further imply that the resulting AVEs will have bilateral variations, given that these trade elasticities are used to convert the trade impacts of NTMs into AVEs. Second, some countries in our data set have bilateral variations in their NTMs to target speci…c partner countries, which naturally will lead to bilateral variations in AVEs.18 Furthermore, even if the NTMs of the importing countries are not country or product speci…c, the compliance costs of NTMs are likely to vary across exporting countries and products, which will give rise to bilateral variations in the AVEs of an importing country at a product level.19 Rather than assuming and imposing that all trading partners face the same restrictiveness due to an NTM, we allow 17 For example, when faced with a 10 percent tari¤ increase, the responsiveness of the US import demand for cars from Germany could be very di¤erent from cars from the Republic of Korea. 18 For example, Kazakhstan, the Russian Federation and Belarus created the Eurasian Customs Union in 2010 so their NTMs do not apply to member countries. 19 For example, Sri Lanka has a certi…cation requirement (A830) for more than 1,600 imported products, such as Mackerel (HS 160415), which comes from China, Thailand and Peru. These trading partners have di¤erent stages of development and sophistication when conforming with such an NTM requirement. 16 the data to speak for themselves, by adopting an econometric model that allows bilateral variations in the coe¢ cient for the NTMs.20 Our empirical strategy relies on using the market powers of the exporter and importer to estimate the size of the AVEs and the trade elasticities. This is motivated by the terms of trade models or trade theories (Bagwell and Staiger, 2011, for tari¤s, and Staiger and Sykes, 2011, for NTMs). In these theories, importing countries with market power (i.e. countries which can in‡uence international prices through their trade policies) will have terms of trade gains to impose tari¤s and NTMs. This is because they can “pass-though”some of the costs to the trading partners given that their trading partners have price sensitive supply curves. If the importing country has no market power, imposing NTMs will cause imports to decrease and the domestic price of the imports to increase to the full extent of the compliance costs. On the contrary, if the exporting country has no market power, NTMs of the importing country will have no e¤ect on imports or domestic prices. The resulting AVE depends on the market powers of the importing country and the exporting country. To this end, we interact the tari¤ and NTM of an HS6 digit product with the following two variables: the share of the exporter in the world trade of this product, and the share of the importer in the world trade of this product. Both variables capture the market powers of the exporter and the importer in the world market for each HS 6-digit product. We expect the trade impact to be lower if the exporter has a larger share in the world trade of the product, which could result in a smaller trade impact due to the presence of the importing s NTMs. However, it is also possible that the exporter could easily divert its exports country’ to other markets when faced with the burdensome NTMs of a speci…c importer, and the trade impact will be larger. Likewise, if the importing country is an important market with a large 20 Overall, we consider the estimation of bilateral variations in trade elasticities and AVEs to be crucial and necessary for improving the empirical applications of these estimates, particularly for researchers and policy makers to evaluate or renegotiate trade agreements. 17 s NTMs. market share of a product, it could be more di¢ cult to comply with the country’ Then the compliance cost for the exporting countries could be higher and that would lead to a larger trade impact. However, it is also possible that if the market share of the importer is large, then it is more di¢ cult for exporting countries to divert their products to other markets, so the trade impact of the NTMs could be lower. Overall the trade e¤ects of the market shares of the importer and the exporter in the world market due to the presence of NTMs depends on the speci…c products and the country-pair. Therefore we let the data reveal the dominant force. A similar argument can be made for tari¤s and the resulting bilateral import elasticities. The third econometric issue to be addressed is related to the endogeneity of the right- hand side variables. Speci…cally, tari¤s and NTMs are trade policies set by the importers and could be endogenous to the trading volume. Including tari¤s and NTMs in the regressions will lead to inconsistent estimates. To address this endogeneity issue, we use the simple averages of tari¤s and NTMs of the three closest neighboring countries in our sample as the instrumental variables (IVs) respectively. A recent work by Jiao and Wei (2020) provides a theoretical justi…cation based on the median voter theorem regarding why trade policies of other countries are exogenous, and hence satisfy the exclusion restriction, but may a¤ect trade policies of a country which make them reasonable IVs. In addition, the trade policies of neighboring countries are likely to be correlated with the trade policy of the importing country due to a regional trading agreement or a common cultural/history background, but the bilateral imports of the importing country from a particular trading partner should not have any direct impact on the bilateral trade policies of neighboring countries on the same trading partner and thus satisfy the exclusion restrictions.21 Please note that this could 21 For example, bilateral imports of Sri Lanka from China could be lower due to the presence of Sri Lanka’ s tari¤s and NTMs on Chinese products. However, we worry that these tari¤s and NTMs could depend on Sri Lanka’ s imports of these Chinese products due to domestic political economy and local industry lobbying considerations. To address this endogeneity issue, we use the simple average of the bilateral tari¤s of India, 18 be an issue if these countries are all part of a regional trade bloc with common external tari¤s and NTMs. However, we believe this is not an issue in our data set as we have a wide range of countries in the sample and not all countries are part of any regional trade bloc with common external trade barriers (for this reason, the European Union is treated as one country). Moreover, the domestic lobbying activities of neighboring countries are not expected to be correlated among neighboring countries.22 We will check the …rst-stage regression results to make sure that these neighboring countries’bilateral trade policies have the right signs and acceptable F-statistics to rule out the issue of weak instruments. In the second stage of the IV regressions, we replace the bilateral tari¤s with their …tted values from the …rst-stage regressions. For the bilateral border NTMs, which is a discrete dummy variable, we run a probit (selection) regression for the bilateral NTM variable in the …rst stage, and construct the inverse Mills ratio according to the …rst-stage regression to be included along side the bilateral NTMs in the second-stage regressions (see Heckman, 1979). 4.2 Econometric Models Taking into account the above econometric issues, …rst we run the …rst-stage IV regressions, and then we run the product-level quantity-based gravity regressions based on cross-section data using …ve models: the zero-in‡ated negative binomial model, the negative binomial model, the zero-in‡ated Poisson model, the Poisson model and the log-OLS model. Bangladesh and Pakistan on Chinese products as the instrument for the bilateral tari¤ of Sri Lanka on the same Chinese products. Similarly, we use the simple average of the bilateral NTMs of India, Bangladesh and Pakistan on Chinese products as the instrument for the bilateral NTMs of Sri Lanka on the same Chinese products. The reasoning is that, while the trade policies of India, Bangladesh and Pakistan could be correlated with the trade policies of Sri Lanka due to regional and cultural proximity, the bilateral imports of Sri Lanka from China of a particular product should NOT a priori directly a¤ect the trade policies of India, Bangladesh and Pakistan on these same Chinese products. In other words, we do not expect domestic political economy and local industry lobbying of Sri Lanka on Chinese products to have any direct impacts on the determinations of bilateral trade policies of India, Bangladesh and Pakistan on these Chinese products. Hence these instruments satisfy the exclusion restriction. 22 For example, the industrial structure of Sri Lanka is very di¤erent from the industrial structures of India, Bangladesh and Pakistan in terms of the products they produce and trade. 19 For the …rst-stage regression for tari¤s, we use the average tari¤ of the three closest ^nij : countries, tnij ; as an instrument for the tari¤, tnij ; to get the …tted tari¤, t ^nij = Zij ^+^t tnij ; t (14) where ^ and ^t are the least squares estimates of the coe¢ cients of the second-stage control variables, Zij ; and the average tari¤, tnij : For the …rst-stage regression for the NTMs, because NTM is a dummy variable that equals one when at least one NTM is present, the …rst-stage regression is a probit regression with a density function, f (:) ; in which we use the average presence of NTMs in the three closest countries, N T M nij ; as an instrument for N T Mnij : N T Mnij 2 f0; 1g ; NT M N T Mnij NT M 1 N T Mnij f N T Mnij jX;N T M nij = Zij + N T M nij 1 Zij + N T M nij ; NT M Zij ^ + ^ N T M nij invM illnij = NT M ; Zij ^ + ^ N T M nij NT M where we retrieve the inverse Mills ratio, invM illnij ; constructed based on ^ and ^ ; the coe¢ cients of the second-stage control variables, Zij ; and the average NTM presence, N T M nij : Finally, we also run the …rst-stage regressions for the four interaction terms that involved tari¤s and NTMs, using the interaction terms with tnij and N T M nij as the respective in- 20 struments: ^nij = Zij ! shareni t !t ^ 1 +^ 1 shareni tnij ; (15) ^nij = Zij ! sharenj t !t ^ 2 +^ 2 sharenj tnij ; shareni N T Mnij = Zij ! !t ^ 3 +^ 3 shareni N T M nij ; sharenj N T Mnij = Zij ! !t ^ 4 +^ 4 sharenj N T M nij : The main speci…cation of second-stage regression for the ZINB, the NB, the ZIP and the Poisson models is t ^ NT M M ln E (Qnij jX) = n + nij tnij + nij N T Mnij + invM illnij + Zij + "nij ; (16) t t t t nij = n + 1 shareni + 2 sharenj ; NT M NT M NT M NT M nij = n + 1 shareni + 2 sharenj : The control variables included in Zij are the standard gravity variables: the log of the gross domestic product (GDP) of the importer and the exporter, the bilateral distance between the importer and the exporter, the landlocked indicators for the importer and the exporter, the common boarder indicator, and also a dummy controlling for other types of t NT M non-border NTMs measures. The bilateral coe¢ cients of tari¤s and NTMs, nij and nij ; s share in the world are obtained by using the interaction terms based on the importer’ s share in the world market, sharenj : For log-OLS, the market, shareni and the exporter’ second-stage regression is similar to equation (16) but with the log of the quantity imported, ln (Qnij ) ; as the dependent variable. For the zero-in‡ated speci…cations, such as the ZINB and the ZIP models, we also include a logit regression to explain the presence of excessive zeros in the bilateral trade. Although it 21 is not necessary to have an extra variable to explain zeros (see Cameron and Trivedi, 1998), we still include a common religion indicator in the zero regression, motivated by Helpman, Melitz and Rubinstein (2008), in addition to all the second-stage control variables, Zij ; to form W in equation (13) : Once all …ve versions of equation (16) are estimated, we use Vuong tests to select the best t NT M …t models to retrieve ^ nij and ^ nij for the construction of the AVE. For all …ve models, ^t has the interpretation of the (instantaneous) semi-elasticity of trade with respect to a nij one percentage point increase in the tari¤: ^t @ ln E (Qnij jX) nij = : (17) @tnij t To obtain the instantaneous elasticity of trade with respect to tari¤s, we multiply ^ nij with tnij : @ ln Qnij t "nij = ^ nij tnij : @ ln tnij Note that unlike the trade elasticity in the literature (see Simonovska and Waugh, 2014 for a review), these trade elasticities are estimated at the HS 6-digit level with bilateral variations. In addition, these trade elasticities are jointly estimated with the AVE of the NTMs, which reduces the potential downward bias due to the negative correlations among tari¤s, NTMs and trade ‡ow. To construct the AVE of the NTMs, we need to …rst construct the proportionate change in quantity imported (or the expected quantity imported for the count models) due to the 22 presence of NTMs: NT M ln E (Qnij jX;N T Mnij = 1) ln E (Qnij jX;N T Mnij = 0) = ^ nij ==> (18) E (Qnij jX;N T Mnij = 1) NT M ln = ^ nij E (Qnij jX;N T Mnij = 0) E (Qnij jX;N T Mnij = 1) NT M = exp ^ nij E (Qnij jX;N T Mnij = 0) E (Qnij jX;N T Mnij = 1) E (Qnij jX;N T Mnij = 0) NT M = exp ^ nij 1: E (Qnij jX;N T Mnij = 0) We then use the semi-elasticity of trade with respect to a one percentage point increase in t the tari¤, ^ nij ; to convert the proportionate change in quantity imported due to NTMs to the ad valorem equivalent tari¤: De…nition 1 The ad valorem equivalent of NTM (AV Enij ) of product n; in importing coun- try i; from exporting country j; measures the ad valorem tari¤ that induces the same propor- tionate change in the quantity imported as the presence of N T Mnij , or23 NT M exp ^ nij 1 AV Enij = : (19) ^t nij Equation (19) above makes it clear that AVE by construction will have product-importer- exporter variation. Variation across products is obvious, because all the regressions are estimated at the HS-6 product level, so all the coe¢ cients will have product variations. The reasons we have variation across importers and exporters are twofold. First, even if NTMs 23 There are other ways to de…ne the AVE, such as the equivalent tari¤ that induces the same change in the quantity imported, or the equivalent tari¤ that induces the same rate ratio change in the quantity imported. In those cases, the formula for the AVE could be slightly di¤erent. Kee, Nicita and Olarreaga (2009) use the import demand elasticity, "; to obtain the following: NT M exp ^ 1 AV E = : " 23 are multilateral with no partner country speci…c measures (which is not true for some coun- tries in our sample), the bilateral variation in import demand elasticities at the denominator will ensure that AVEs have variation across importers and exporters. Furthermore, com- pliance costs may vary across exporting countries even if they face the same NTMs of an importing country, which are captured by the numerator of Equation (19) : Overall, the size and magnitude of the AVE of a product of an importing country from an exporting country depends on both the compliance costs and market power of both the importing country and the exporting country. Finally, we bootstrap this procedure 50 times to obtain the bootstrap standard errors of AV Enij : 5 Results for the AVE Estimations The …rst-stage regressions …t very well. The average …rst-stage F-statistics for tari¤s is 142.47, with a median of 104.01, 1st percentile of 8.46, and 99th percentile of 633.56. Sim- ilarly, the average …rst-stage Chi-squared statistics for NTMs is 442.19, with a median of 346.07, 1st percentile of 12.25, and 99th percentile of 1532.15. These results indicate that tari¤s and NTMs of neighboring countries are not weak instruments for own tari¤s and NTMs. For the second stage regressions, the NB model works best for 71 percent of the products, the ZINB for 10 percent of the products, Poisson for 2 percent, and ZIP for 1 percent. For the remaining 16 percent of products, non-linear models did not converge and therefore OLS was used. Based on these estimates, we constructed the AVE of NTMs according to equation (19). At the sample mean, when the border measures are present, the average AVE is 4.7 percent, whereas the mean tari¤ is 6.9 percent. 24 What is the relationship between NTMs and the tari¤s in our sample? Column (1) of Table (2) presents the regression result when we regress the tari¤s on the AVEs, controlling for importer, exporter and product …xed e¤ects. The negative and signi…cant coe¢ cient suggests that the restrictiveness of NTMs and tari¤s are negatively correlated. Thus, at the sample mean, NTMs and tari¤s are policy substitutes. Column (2) focuses on the variation of tari¤s and NTMs when we compare among all exporting countries controlling for import- product and exporter …xed e¤ects. The coe¢ cient on AVEs remains negative, suggesting that NTMs are more restrictive for those exporting countries that face lower tari¤s for a given importing country and product. This result is consistent with the argument that more restrictive NTMs are in place for those countries that receive tari¤ preferences. Column (3) focuses on the variation of tari¤s and NTMs when we compare among prod- ucts controlling for importer-exporter …xed e¤ects. The coe¢ cient on AVEs remains neg- ative, which is consistent with the hypothesis that NTMs are more restrictive for those products that have lower tari¤s for a given pair of importing and exporting countries. In other words, when the high income countries grant tari¤ preference to their trading partners on certain products, they may impose more restrictive NTMs on other products from the same exporting countries. Despite the complicated estimation procedure with more than 700,000 estimates, our elasticity estimates are quite reasonable. First, at the sample mean, the average trade elasticity for HS 6-digit products is about -4.7, which is a bit higher than Simonovska and s (2014) structural estimates. Given that our estimates are at a more disaggregate Waugh’ level, the higher average elasticity is to be expected. In addition, we controlled for NTMs in the estimation of these trade elasticities based on tari¤s. Given that tari¤s and NTMs are negatively correlated in our sample and are both negatively correlated with trade ‡ows, including NTMs in the trade ‡ows regressions to estimate trade elasticity would minimize 25 the bias of the coe¢ cient on tari¤s and lead to higher elasticity estimates. In addition, as a robustness check, we split the HS 6-digit products into homogeneous s classi…cation. As expected, the homoge- versus di¤erentiated products according to Rauch’ neous products are signi…cantly more elastic than the di¤erentiated products. At the sample mean, the trade elasticity for the homogeneous products is -6.2, while the trade elasticity for the di¤erentiated products is -4.2, and the di¤erence is statistically signi…cant even after we controlled for the importer-exporter pair …xed e¤ects.24 Finally, instead of using the market shares of importing and exporting countries to mea- sure market powers in Equation 16, we use the logs of the GDPs of the importing and exporting countries as a robustness check. This is because larger countries may have more market power to in‡uence world prices according to the terms of trade theory. One problem with using the log of GDP is that it does not vary by products. In other words, we are assuming that larger countries have more market power for all products, which may not be the case empirically and may introduce bias in the estimations. This is why we prefer our original market share variables which measure the market power of each country for each product. Nevertheless, results based on interacting the log of GDP of the importing and exporting countries are quite consistent with our original results. The correlation between the AVE estimated using market shares and the AVE estimated using GDPs is 0.36, while the correlation between the import demand elasticity estimated using market shares and the import demand elasticity estimated using GDPs is 0.58. Regressing the AVE estimated using GDPs on the AVE estimated using market shares results in a R-square of 0.74. So both sets of estimates are positively correlated. Overall we are con…dent that the results based on our original market share estimations are robust and consistent. 24 Regression results are available upon request. 26 6 Results on NTM-Induced Trade Fraud 6.1 Least Squares Downward Bias and Instrumental Variable Es- timations With the estimated AVEs we verify the …ve testable hypothesis of the theoretical model. Table 3 presents the least squares results when we regress trade discrepancies on the AVE of Border NTMs and tari¤s. The dependent variables for Columns (1) to (3) are the quantity discrepancies, while the dependent variables for Columns (4) to (6) are the value discrep- ancies. Both sets of discrepancies are constructed according to equation (8) : We also tried other ways of measuring the dependent variables, but the results are very similar. For all the regressions in this table, we control for importer-, exporter- and product-…xed e¤ects, and we cluster the standard error by importer-product, which is the level of aggregation of most NTM variables. Clustering is particularly necessary, as the AVEs are estimated with errors. Columns (1) and (4) regress trade discrepancies on the estimated AVEs of border NTMs. Columns (2) and (5) include bilateral tari¤s in the regressions. Columns (3) and (6) sum up the AVEs and tari¤s together as one variable and include the presence of the other NTMs as control variables in the regressions. In all speci…cations, the least squares estimated coe¢ cients on the AVEs are positive and are statistically signi…cant at the 1 percent level. However, there are reasons to believe that these least squares estimates for the AVEs have downward bias. First, a primary concern is that corruption at the border could cause trade discrepancies and at the same time be paired with a lack of enforcement on NTM regulations. This may result in both larger discrepancies and less restrictive AVEs, therefore biasing the LS estimation towards zero. Moreover, a large part of the trade discrepancies could be due to idiosyncratic and random recording errors and not NTM-induced trade 27 fraud. In this scenario, it will be harder to …nd a systematic relationship between the trade discrepancies and the NTMs. This will further push the least squares estimates toward zero. Finally, there may be a concern that the AVEs are contaminated with estimation errors that could lead to systematic downward bias in the least squares results presented in Table 3. We address all these concerns with the use of instrumental variables, which are correlated with the AVEs but not necessarily correlated with trade. For importing country i; product n and exporting country j; we use the average AVE of exporting country j of the product n in the non-i markets, AV E n~ij ; as the instrument for AV Enij : This is a valid instrument, because it captures the compliance cost of exporting country j for product n; facing similar NTM measures in the other markets. Thus, potentially it could be positively correlated with AV Enij ; and yet it should not a¤ect the trade discrepancies between country i and j on product n:25 Table 4 presents the …rst-stage regression results for the IV estimations. Columns (1), (2), (4) and (5) are for AV Enij ; corresponding to Columns (1), (2), (4) and (5) of Table 3. Likewise, Columns (3) and (6) are for AV Enij + T arif f; corresponding to Columns (3) and (6) of Table 3. All standard errors are clustered by importer-product. Across all speci…cations, the coe¢ cients on the average AVE of exporting country, AV E n~ij ; are positive and statistically signi…cant. Together with the high …rst-stage F-statistics, they con…rm that AV E n~ij is a valid instrument. Table 5 presents the second-stage regression results. Similar to Table 3, the dependent variables for Columns (1) to (3) are the quantity discrepancies, while the dependent variables for Columns (4) to (6) are the value discrepancies, and we cluster the standard errors by importer-product. Compared to Table 3, the results in Table 5 are much larger, con…rming the downward bias in the previous least squares estimates. The coe¢ cients for the AVEs are 25 We also tried using the trade weighted average AVE of exporting country j of product n in non-i markets as the instrument. The results are very similar. 28 consistently higher and statistically signi…cant across all columns. Overall, at the sample mean, a 10 percentage point increase in the AVEs leads to a 15 percent increase in trade discrepancies, while a comparable increase in tari¤s leads to only an 8 percent increase in trade discrepancies. This suggests that a signi…cant part of the trade discrepancies results from NTM avoidance schemes, as predicted by the theoretical model. These results are consistent with the testable hypothesis 1, which is that trade discrepancies increases with the restrictiveness of NTMs. Moreover, the positive relationship between trade discrepancies and tari¤s also provides support for testable hypothesis 5 of the model. 6.2 Sources of Trade Fraud: Origin Fraud and Product Classi…- cation Fraud What can explain the positive relationship between trade discrepancies and AVEs? It could be that …rms mis-declared the country of origin of their products in order to avoid the burdensome NTMs. This would be consistent with origin fraud. If this is the case, we expect that, given a speci…c importing country and product, exporting countries that have higher AVEs are also those that have larger trade discrepancies with the importing country. Tables 6 and 7 test the origin fraud hypothesis with least squares and IV regressions, respectively. First-stage regressions are presented in Table 8. Instead of controlling for a full set of …xed e¤ects, we control for importer-product …xed e¤ects and exporter …xed e¤ects. Thus, we rely on the variation of the AVEs of the exporters in the same importing market with the same product to identify the coe¢ cient of the AVEs. Similar to the previous tables, the dependent variables for Columns (1) to (3) are the quantity discrepancies, while the dependent variables for Columns (4) to (6) are the value discrepancies, and we cluster the standard errors by importer-product. Both least squares and IV results show that, within the importer-product pairs, the exporters that face higher AVEs are those that have larger trade discrepancies with 29 the importers. This is consistent with origin fraud, whereby the …rms exporting products coming from countries subject to high AVEs tend to mis-declare the country of origin to bene…t from a lower AVE. The other type of trade fraud is product classi…cation fraud, whereby products are mis- classi…ed as other products in order to evade or reduce the costs associated with NTMs compliance. This type of fraud results in, given a speci…c importing-exporting countries pair, products that have higher AVEs are also those that have larger trade discrepancies. Tables 9 and 10 test the product classi…cation fraud hypothesis with least squares and IV regressions, respectively. First-stage regressions are presented in Table 11. Instead of con- trolling for a full set of …xed e¤ects, we now control for importer-exporter …xed e¤ects and product …xed e¤ects. In other words, we rely on the variation in the AVEs across products within the importer-exporter pairs to identify the coe¢ cient on the AVEs. Similar to the previous tables, the dependent variables for Columns (1) to (3) are the quantity discrep- ancies, while the dependent variables for Columns (4) to (6) are the value discrepancies, and we cluster the standard errors by importer-product. Both tables show that, within the importer-exporter pairs, products that face higher AVEs are those that have larger trade dis- crepancies. This is consistent with product classi…cation fraud, whereby …rms mis-declared products to lower costs associated with NTMs. Overall, evidence presented in Tables 3 to 10 are consistent with NTM-induced trade fraud, when …rms either mis-declared their country of origin or their product codes in order to avoid the burdensome NTMs of the importing countries. This provides evidence supporting the second testable hypothesis of the theoretical model. In addition, comparing the point estimates presented in Tables 7 and 10 suggests that product classi…cation fraud is more sensitive to restrictive NTMs, relative to country of origin fraud. In other words, facing restrictive NTMs, more traders may choose to misreport product classi…cation codes than 30 the country of origin. This is consistent with the anecdotal evidence presented at the start of the paper in italics. 7 Robustness Checks The regression results presented above show that larger trade discrepancies are associated with more restrictive NTMs. To further validate this, we split the sample according to the product characteristics and the country characteristics. 7.1 Homogeneous versus Di¤erentiated Products Homogeneous products are easier for …rms to either mis-declare the product codes or country of origin in order to evade restrictive NTMs. For example, it is easier to declare aluminum tubes (HS 760810) from China as aluminum tubes from Brazil, given that aluminum tubes are similar regardless of their origin. In contrast, it is harder to declare cars (HS 870390) from Japan as cars from Germany, given that cars are heterogeneous (Toyota vs Volkswagen). Likewise, it is easier to declare aluminum tubes as aluminum rods (HS 760410), than to declare cars as buses (HS 870210). Therefore we expect the e¤ects of NTMs on trade discrepancies to be larger for homogeneous products. Table 12 presents the second-stage IV regression results for origin and product classi…ca- tion fraud, with a sub-sample of homogeneous and di¤erentiated products. We instrument the AVEs the same way as before while controlling for tari¤, and the presence of other NTMs. The …rst-stage results are available upon request. Columns (1) to (6) are for origin fraud, while Columns (7)-(12) are for product classi…cation fraud. Columns (1)-(3) and (7)-(9) con- sist of a sub-sample of di¤erentiated products, while Columns (4)-(6) and (10)-(12) are the results for homogeneous products. The e¤ects of NTMs on trade discrepancies remain pos- 31 itive and statistically signi…cant for all these speci…cations, with the point estimates larger for homogeneous than di¤erentiated products.26 These results support testable hypothesis 4 of the model. This is also consistent with the …ndings of Javorcik and Narciso (2008), who claim that quantity measures are more relevant for the mis-declaration of homogeneous goods since homogenous goods are often sold/measured by weight.27 7.2 High-Income Countries versus Developing Countries Countries with higher income often have better infrastructure and institutions to enforce rules of law, including NTMs, and at the same time have better customs enforcement to accurately record trading transactions. We expect the e¤ects of NTMs on trade discrepancies to be smaller for trade between high-income importing and exporting countries, than between high-income importing country and developing exporting countries. Table 13 presents the second-stage IV regression results for a sub-sample of high-income importing countries. Similar to Table 12, Columns (1) to (6) are for origin fraud, while Columns (7)-(12) are for product classi…cation fraud. Columns (1)-(3) and (7)-(9) are the results when the exporting countries are developing countries. In these columns, all the AVE variables remain positive and highly signi…cant. In contrast, Columns (4)-(6) and (10)-(12) are the results when the exporting countries are also high-income countries. None of the AVE variables are statistically signi…cant. Together these results suggest that trade fraud mainly exist between high-income importing countries and developing exporting countries. These results support testable hypothesis 3 of the theoretical model. Overall, these results are consistent with our hypothesis that NTM-induced trade fraud 26 Results based on regressions for a combined sample of homogeneous and di¤erentiated products cannot reject the hypothesis that the e¤ects of NTMs on trade discrepancies are lower for di¤erentiated products than for homogeneous products, for both product fraud and origin fraud, controlling for tari¤s. 27 When we restrict the sample of homogeneous products which are measured by weight for Table 12, the results are consistently larger and stronger, con…rming the …nding of Javorcik and Narciso (2008). 32 is particularly relevant for homogeneous products and for the trade between developed and developing countries. This is consistent with the theoretical model that it is easier for …rms to mis-declare product codes or countries of origin to evade restrictive NTMs when products are homogeneous, and when customs administrations are less developed. 8 Concluding Remarks This paper examines whether some of the discrepancies in the o¢ cial statistics of importing and exporting countries result from trade fraud in which traders deliberately mis-declare the country of origin or product code in order to avoid costly or burdensome NTMs at the borders. Our approach utilizes the detailed product-level trade statistics of a wide range of importing and exporting countries. We …rst provide a theoretical model to link the re- strictiveness of NTMs to trade fraud and to rationalize the presence of origin and product classi…cation fraud. We then estimate the ad valorem equivalents of the NTMs of the im- porting countries, with variations at the product and exporting country levels and relate the estimated ad valorem equivalents to the existing trade discrepancies, in terms of the quantity and value discrepancies. Our …ndings indicate that trade ‡ows where larger discrepancies are observed are those that face more restrictive NTMs, suggesting the misclassi…cation of prod- uct codes or the misrepresentation of the country of origin in order to reduce or evade the compliance costs associated with Border NTMs. We …nd that NTM-induced trade fraud is more pronounced for homogeneous products and for trade between developed and developing countries. One additional contribution of this paper is the quanti…cation of the costs associated with Border NTMs, …lling an important void in the trade policy literature which thus far has focused primarily on tari¤s. Given that tari¤s have been steadily declining for the 33 past few decades, recent trade negotiations and agreements mostly focus on NTMs. From a methodological point of view, this paper presents a novel method for estimating the ad- valorem equivalent of NTMs. In addition, our …nding that AVEs and tari¤s are substitutes suggests that it is important to include NTMs in trade policy analysis. Finally, the general results of this paper indicate that NTMs at the border add important costs to trade and that traders seek to avoid these costs by resorting to misclassi…cation of products and/or misreporting quantities. Avoidance of NTMs at the border not only undermines the trade policy of the importing country, for example by misreporting country of origin, but also undermines regulations that seek to guarantee consumers’ health and environmental protection. References [1] Bagwell, K. and Staiger, R. (2011). “What Do Trade Negotiators Negotiate About? Empirical Evidence from the World Trade Organization,” American Economic Review 101: 1238–1273. [2] Bhagwati, J. (1964). “On the Underinvoicing of Imports,” Bulletin of the Oxford Uni- versity Institute of Economics & Statistics 27(4): 389–397. [3] Broda, C., and Weinstein, D. (2006). “Globalization and the Gains from Variety,”Quar- terly Journal of Economics 121(2): 541-585. [4] Cameron, C. and Trivedi, P. (2013), Regression Analysis of Count Data, 2nd edition, Econometric Society Monograph No.53, Cambridge University Press. [5] Cantens, T. (2015). “Mirror Analysis and Revenue Frauds,” World Customs Organiza- tion Research Paper No. 35. 34 [6] Chalendard, C., Fernandes, A., Raballand, G., and Rijkers, B. (2020). “Technology (Ab)use and Corruption in Customs,”World Bank mimeo. s [7] Demir, B., and Javorcik, B. (2020). “Trade policy changes, tax evasion and Benford’ law,” Journal of Development Economics 144, 102456. [8] Ederington, J. and Ruta, M. (2016). "Non-Tari¤ Measures and the World Trading System". World Bank Policy Research Working Paper 7661. [9] Ferrantino, M., and Wang, Z. (2008). “Accounting for Discrepancies in Bilateral Trade: The Case of China, Hong Kong and the United States,” China Economic Review 19: 502-520. [10] Ferrantino, M., Liu, X., and Wang, Z. (2012). “Evasion behaviors of exporters and importers: Evidence from the U.S.–China trade data discrepancy,” Journal of Interna- tional Economics 86: 141–157. [11] Fisman, R. and Wei, S. (2004). “Tax Rates and Tax Evasion: Evidence from "Missing" Imports in China,” Journal of Political Economy, vol. 112, no. 2. [12] Goldberg, P., and Pavcnik, N. (2016). “The E¤ects of Trade Policy,” Handbook of Com- mercial Policy, edited by Kyle Bagwell and Robert Staiger. [13] Grigoriou, C. (2019). “Mirror analysis as a support for risk management and valuation: a practical study.” World Customs Journal 13(2): 91-113. [14] Grigoriou C., Kalizinje, F., and Raballand, G. (2019). “How helpful are mirror statistics for Customs reform? Lessons from a decade of operational use.”World Customs Journal 13(2): 105-114. 35 [15] Heckman, J. J. (1979). “Sample Selection as a Speci…cation Error,”Econometrica 47(1): 153–161. [16] Helpman, E., Melitz, M., and Rubinstein, Y. (2008). “Estimating Trade Flows: Trading Partners and Trading Volumes,” Quarterly Journal of Economics 123(2): 441-487. [17] Hillberry, R., and Hummels, D. (2013). “Trade elasticity parameters for a computable general equilibrium model.” Handbook of Computable General Equilibrium Modeling 1: 1213-1269. [18] Hummels, D., and Lugovskyy, V. (2006). “Usable Data? Matched Partner Trade Sta- tistics as a Measure of Transportation Costs” Review of International Economics 14(1). [19] Javorcik, B. and Narciso, G. (2008). “Di¤erentiated Products and Evasion of Import Tari¤s,” Journal of International Economics 76: 208-222. [20] Jiao, Y., and Wei, S. (2020). “Natural Barriers and Policy Barriers,” NBER working paper 28171. [21] Kee, H. L., and Nicita, A. (2021). “Discrete Trade Reforms with Heterogeneous Elastic- ities: Assessing the Impacts of a No-Deal Brexit,”World Bank Policy Research Working Paper 8195. [22] Kee, H. L., Nicita, A., and Olarreaga, M. (2008). “Import demand elasticities and trade distortions,” Review of Economics and Statistics 90(4): 666–682. [23] Kee, H. L., Nicita, A., and Olarreaga, M. (2009). “Estimating Trade Restrictiveness Indices,” Economic Journal 119: 172–199. 36 [24] Kee, H. L., Neagu, C., and Nicita, A. (2013). “Is Protectionism on the Rise? Assessing national trade policies during the crisis of 2008,” Review of Economics and Statistics 95(1): 342–346. [25] Lee, J., and Swagel, P. (1997). “Trade barriers and trade ‡ows across countries and industries,” Review of Economics and Statistics 79 (3): 372-382. [26] Li, Y., and Lin, F. (2022). “Beyond tari¤ evasion: bypass e¤ect of FTAs to circumvent technical barriers,” Review of World Economics. [27] Liu, R., Sheng, L., and Wang, J. (2021). “Faking Trade for Capital Control Evasion: Evidence from Dual Exchange Rate Arbitrage in China,”working paper. [28] Martin, W., and Pham, C. (2020). “Estimating the gravity model when zero trade ‡ows are frequent and economically determined,” Applied Economics 52(26): 2766-2779. [29] Mishra, P., Subramanian, A., and Topalova, P. (2008). “Tari¤s, enforcement, and cus- toms evasion: Evidence from India,”Journal of Public Economics 92 (1011), 1907–1925. [30] Santos Silva, J.M.C. and Tenreyro, S. (2006). “The Log of Gravity.” Review of Eco- nomics and Statistics 88 (4): 641–658. [31] Santos Silva, J.M.C. and Tenreyro , S. (2011). “Further simulation evidence on the performance of the Poisson pseudo-maximum likelihood estimator.” Economics Letters 112(2): 220–222. [32] Sequeira, S. (2016). “Corruption, Trade Costs and Gains from Tari¤ Liberalization: Evidence from Southern Africa,” American Economic Review 106(10): 3029-3063. [33] Simonovska, I., and Waugh, M. (2014). “The elasticity of trade: Estimates and evi- dence,” Journal of International Economics 92(1): 34-50. 37 [34] Staiger, R., and Sykes, A. (2011). “International Trade, National Treatment, and Do- mestic Regulation,” Journal of Legal Studies 40: 149-203. [35] Swenson, D. (2001). “Tax reforms and evidence of transfer pricing,” National Tax Jour- nal 54 (1), 7–25. [36] UNCTAD (2015). International Classi…cation of Non-Tari¤ Measures. United Nations publication. [37] UNCTAD, and World Bank (2018). The unseen impact of non-tari¤ measures: insights from a new database. [38] World Integrated Trade Solutions (WITS)-Users Manual (2011). 38 Figure 1: Sample Distribution of Trade Discrepancies .0 2 .0 15 .0 1 .0 05 0 -2 0 0 -1 0 0 0 1 00 200 x k d e ns ity g a p v sh k de ns ity g a p s h Figure 2: Proportion of Positive Quantity Kernel density estimate 10 Density 5 0 0 .1 .2 .3 .4 (mean) q_freq kernel = epanechnikov, bandwidth = 0.0073 Figure 3: Over Dispersion: Ratio of Variance over Mean of Positive Quantity Kernel density estimate 1.000e-07 Density 5.000e-08 0 0 10000000 20000000 30000000 ratio kernel = epanechnikov, bandwidth = 1.2e+06 39 Table 1: Coverage Ratio of Non-Tari¤ Measures by Importing Countries Importing Country Percent Importing Country Percent ARE 99.22 JPN 67.08 ARG 51.38 KAZ 62.80 AUS 79.32 KOR 94.39 BGD 98.20 LKA 89.87 BOL 90.11 MAR 37.30 BRA 88.27 MEX 49.14 CAN 56.54 MYS 72.48 CHE 60.62 NGA 99.00 CHL 57.43 NZL 86.21 CHN 71.86 PAK 56.47 CIV 94.78 PER 89.13 CMR 71.90 PHL 99.01 COL 97.25 PRY 82.37 CRI 59.45 RUS 69.53 CUB 98.03 SAU 38.77 DZA 64.74 SEN 98.36 ECU 75.31 SGP 84.98 ETH 97.87 THA 90.05 EUN 60.98 TUN 46.83 GHA 97.07 TUR 86.34 GTM 86.63 URY 73.84 HKG 84.33 USA 85.44 HND 96.19 VEN 92.22 IDN 78.95 VNM 84.25 IND 79.19 Data Source: UNCTAD TRAINS. Coverage Ratio measures the percentage of HS 6 digit products that are subjected to at least one border NTMs in the importing country. 40 Table 2: Second Stage IV Regression, Dependent Variable: Tari¤s (1) (2) (3) AVEborder -0.125*** -0.0740*** -0.0874*** (0.0309) (0.0137) (0.0284) Observations 253,965 248,215 253,818 Importer Fixed E¤ects Yes Exporter Fixed E¤ects Yes Yes Product Fixed E¤ects Yes Yes Importer-Product Fixed E¤ects Yes Importer-Exporter Fixed E¤ects Yes Notes: Standard errors are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively; AVEnij is instrumented using the average AVE of exporting country j of the product n in the non-i markets. Table 3: Downward Bias Least Squares Regressions Dependent variables Quantity Discrepancies Value Discrepancies (1) (2) (3) (4) (5) (6) AVE of NTM-border 0.403*** 0.383*** 0.360*** 0.348*** (0.0663) (0.0688) (0.0650) (0.0675) Tari¤ 0.829*** 0.707*** (0.0413) (0.0421) AVE of NTM-border+Tari¤ 0.736*** 0.636*** (0.0346) (0.0353) SPS/TBT NTMs -3.769*** -3.277*** (0.991) (0.965) Other NTMs -3.999*** -4.022*** (0.961) (0.955) Constant 95.83*** 89.31*** 93.25*** 98.70*** 93.09*** 96.72*** (0.326) (0.464) (0.760) (0.319) (0.465) (0.749) Observations 256,711 250,432 250,432 251,934 245,615 245,615 R-squared 0.164 0.166 0.166 0.163 0.164 0.165 Importer Fixed E¤ects Yes Yes Yes Yes Yes Yes Exporter Fixed E¤ects Yes Yes Yes Yes Yes Yes Product Fixed E¤ects Yes Yes Yes Yes Yes Yes Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. 41 Table 4: First Stage Instrumental Variable Regressions Dependent variables AVE AVE+Tari¤ AVE AVE+Tari¤ (1) (2) (3) (4) (5) (6) Average AVE of exporter 0.634*** 0.608*** 0.650*** 0.626*** (0.030) (0.031) (0.030) (0.024) Tari¤ 0.025*** 0.024*** (0.004) (0.004) Average AVE of exporter 0.986*** 0.982*** +Tari¤ (0.018) (0.018) SPS/TBT NTMs 6.907*** 6.829*** (0.369) (0.368) Other NTMs -0.321 -0.271 (0.445) (0.450) Importer Fixed E¤ects Yes Yes Yes Yes Yes Yes Exporter Fixed E¤ects Yes Yes Yes Yes Yes Yes Product Fixed E¤ects Yes Yes Yes Yes Yes Yes F-Statistics 441.13 392.79 2938.74 472.38 419.60 2850.57 Observations 256,711 250,432 153,084 251,934 245,615 150,582 Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. Table 5: Second Stage Instrumental Variable Regressions Dependent variables Quantity Discrepancies Value Discrepancies (1) (2) (3) (4) (5) (6) AVEborder 1.257*** 1.456*** 1.354*** 1.537*** (0.218) (0.234) (0.208) (0.222) tari¤ 0.805*** 0.682*** (0.0427) (0.0436) avetar 0.890*** 0.798*** (0.0525) (0.0532) SPS/TBT NTMs -9.221*** -8.764*** (1.302) (1.272) Other NTMs -3.621*** -3.598*** (1.267) (1.256) Observations 256,711 250,432 153,084 251,934 245,615 150,582 Importer Fixed E¤ects Yes Yes Yes Yes Yes Yes Exporter Fixed E¤ects Yes Yes Yes Yes Yes Yes Product Fixed E¤ects Yes Yes Yes Yes Yes Yes Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. 42 Table 6: Origin Frauds: Downward Bias Least Squares Regressions Dependent variables Quantity Discrepancies Value Discrepancies (1) (2) (3) (4) (5) (6) AVEborder 0.691*** 0.638*** 0.571*** 0.523*** (0.104) (0.107) (0.102) (0.105) Tari¤ 2.026*** 2.007*** (0.0669) (0.0699) AVEborder+Tari¤ 1.706*** 1.670*** (0.0559) (0.0576) SPS/TBT NTMs -16.02*** -24.97*** (4.489) (4.501) Other NTMs 14.97** 13.84** (5.843) (5.566) Constant 95.41*** 79.59*** 83.19*** 98.40*** 82.63*** 91.84*** (0.185) (0.544) (3.301) (0.182) (0.570) (3.124) Observations 251,184 244,689 244,689 246,373 239,854 239,854 R-squared 0.280 0.284 0.283 0.280 0.284 0.284 Exporter FE Yes Yes Yes Yes Yes Yes Importer-Product FE Yes Yes Yes Yes Yes Yes Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. Table 7: Origin Frauds: Second Stage IV Regressions Dependent variables Quantity Discrepancies Value Discrepancies (1) (2) (3) (4) (5) (6) AVEborder 1.413*** 1.672*** 1.367*** 1.600*** (0.250) (0.269) (0.237) (0.252) Tari¤ 2.007*** 1.990*** (0.0668) (0.0698) AVEborder + Tari¤ 2.272*** 2.306*** (0.102) (0.102) SPS/TBT NTMs -18.06** -29.03*** (8.601) (8.398) Other NTMs 7.159 15.92 (11.12) (10.50) Observations 251,184 244,689 146,609 246,373 239,854 144,111 Exporter Fixed E¤ects Yes Yes Yes Yes Yes Yes Importer-Product Fixed E¤ects Yes Yes Yes Yes Yes Yes Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. 43 Table 8: Origin Frauds: First Stage Instrumental Variable Regressions Dependent variables AVE AVE+Tari¤ AVE AVE+Tari¤ (3) (4) (5) (6) (7) (8) Average AVE of exporter 0.597*** 0.571*** 0.613*** 0.591*** (0.028) (0.029) (0.028) (0.029) Tari¤ 0.021*** 0.019*** (0.001) (0.001) Average AVE of exporter 0.899*** 0.900*** +Tari¤ (0.067) (0.017) SPS/TBT NTMs 4.601*** 4.300*** (0.941) (0.940) Other NTMs 1.466 1.464 (1.034) (0.992) Exporter FE Yes Yes Yes Yes Yes Yes Importer-Product FE Yes Yes Yes Yes Yes Yes F-Statistics 444.17 388.29 2422.37 465.06 409.57 2871.97 Observations 251,184 244,689 146,609 246,373 239,854 144,111 Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. Table 9: Product Classi…cation Frauds: Downward Bias Least Squares Regressions Dependent variables Quantity Discrepancies Value Discrepancies (1) (2) (3) (4) (5) (6) AVEborder 0.342*** 0.381*** 0.320*** 0.360*** (0.0654) (0.0670) (0.0633) (0.0649) Tari¤ 0.131*** -0.0141 (0.0459) (0.0476) AVEborder+Tari¤ 0.215*** 0.103*** (0.0363) (0.0371) SPS/TBT NTMs -2.048** -1.303 (0.973) (0.939) Other NTMs -3.620*** -3.550*** (0.937) (0.925) Constant 95.90*** 94.70*** 97.01*** 98.73*** 98.69*** 100.4*** (0.318) (0.478) (0.756) (0.309) (0.482) (0.741) Observations 256,565 250,285 250,285 251,785 245,466 245,466 R-squared 0.227 0.226 0.226 0.233 0.232 0.232 Product FE Yes Yes Yes Yes Yes Yes Importer-Exporter FE Yes Yes Yes Yes Yes Yes Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. 44 Table 10: Product Classi…cation Frauds: Second Stage IV Regressions Dependent variables Quantity Discrepancies Value Discrepancies (1) (2) (3) (4) (5) (6) AVEborder 1.868*** 1.977*** 1.965*** 2.051*** (0.223) (0.239) (0.213) (0.227) Tari¤ 0.0973** -0.0479 (0.0471) (0.0487) AVEborder + Tari¤ 0.260*** 0.134** (0.0511) (0.0550) SPS/TBT NTMs -3.561*** -2.668** (1.223) (1.190) Other NTMs -3.105*** -3.063*** (1.178) (1.159) Observations 256,565 250,285 152,919 251,785 245,466 150,418 Product FE Yes Yes Yes Yes Yes Yes Importer-Exporter FE Yes Yes Yes Yes Yes Yes Notes: Robust standard errors in parentheses are clustered by importer-product; *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively. Table 11: Product Classi…cation Frauds: First Stage Instrumental Variable Regressions Dependent variables AVE AVE+Tari¤ AVE AVE+Tari¤ (3) (6) Average AVE of exporter 0.636*** 0.610*** 0.650*** 0.625*** (0.030) (0.031) (0.030) (0.031) Tari¤ 0.023*** 0.023*** (0.005) (0.005) Average AVE of exporter 0.990*** 0.984*** +Tari¤ (0.021) (0.022) SPS NTMs 6.877*** 6.828*** (0.370) (0.369) Other NTMs -0.285 -0.235 (0.461) (0.464) Product FE Yes Yes Yes Yes Yes Yes Importer-Exporter FE Yes Yes Yes Yes Yes Yes F-Statistics 445.38 396.27 2147.43 472.86 419.99 2088.07 Observations 256,565 250,285 152,919 251,785 245,466 150,418 Notes: Standard errors are clustered by importer-product , ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively 45 Table 12: Second Stage IV Regressions for Homogeneous and Di¤erentiated Products, Dependent Variable: Quantity Discrep- ancies Origin Frauds Product Classi…cation Frauds (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) AV E b o rd e r 1.681*** 2.030*** 2.823*** 2.554*** 1.993*** 2.148*** 2.375*** 2.228*** (0.335) (0.368) (0.852) (0.854) (0.295) (0.320) (0.632) (0.655) ta ri¤ 2.143*** 1.266*** 0.0751 0.246 (0.0748) (0.297) (0.0588) (0.154) ave ta r 2.272*** 1.951*** 0.283*** 0.327* (0.112) (0.545) (0.0689) (0.189) SPS/T BT N T M -8.025 -21.63 -3.652** 2.846 (11.98) (16.90) (1.479) (5.472) O th e r N T M 8.341 48.75** -6.023*** 7.765 46 (17.47) (22.18) (1.424) (5.350) Obs 182,481 177,530 108,125 11,213 10,820 6,350 184,666 179,869 110,887 11,887 11,507 7,042 F ix e d E ¤e c ts: Im p o rte r-P ro d u c t Yes Yes Yes Yes Yes Yes E x p o rte r Yes Yes Yes Yes Yes Yes Im p o rte r-E x p o rte r Yes Yes Yes Yes Yes Yes P ro d u c t Yes Yes Yes Yes Yes Yes Notes: Columns (1)-(3) and (7)-(9) consist of trading of di¤erentiated products. Columns (4)-(6) and (10)-(12) consist of trading of homogeneous products. Standard errors are clustered by importer-product. *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively Table 13: Second Stage IV Regressions for High-Income Importing Countries, Dependent Variable: Quantity Discrepancies Origin Frauds Product Classi…cation Frauds (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) AVEborder 1.530*** 1.819*** 1.409 1.544 1.947*** 2.038*** 1.427 1.752 (0.250) (0.268) (1.398) (1.511) (0.227) (0.242) (1.085) (1.226) tari¤ 2.029*** -0.224 0.0731 -1.054*** (0.0692) (0.355) (0.0474) (0.302) avetar 2.321*** 0.0342 0.225*** -0.423 (0.104) (0.517) (0.0511) (0.407) SPS/TBT NTMs -19.29** 0 -4.316*** 11.90** (8.626) (0) (1.262) (5.750) Other NTMs 5.471 0 -3.145*** -4.569 47 (11.19) (0) (1.210) (6.672) Observations 216,921 210,940 130,235 32,687 32,160 14,580 222,594 216,837 136,785 33,647 33,120 15,574 Fixed E¤ects Importer-Product Yes Yes Yes Yes Yes Yes Exporter Yes Yes Yes Yes Yes Yes Importer-Exporter Yes Yes Yes Yes Yes Yes Product Yes Yes Yes Yes Yes Yes Notes: Unless otherwise speci…ed, samples consist of high-income importing countries only. Columns (1)-(3) and (7)-(9) consist of trading between high-income importing countries and developing exporting countries. Columns (4)-(6) and (10)-(12) consist of trading between high-income importing countries and high-income exporting countries. Standard errors are clustered by importer-product. *, ** and *** indicate that coe¢ cients are signi…cant at 90%, 95% and 99%, respectively