NBER WORKING PAPER SERIES MISREPORTING TRADE: TARIFF EVASION, CORRUPTION, AND AUDITING STANDARDS. Derek Kellenberg Arik Levinson

Transcription

1 NBER WORKING PAPER SERIES MISREPORTING TRADE: TARIFF EVASION, CORRUPTION, AND AUDITING STANDARDS Derek Kellenberg Arik Levinson Working Paper NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cabridge, MA Septeber 2016 The authors are grateful to Jordan Marcusse for research assistance, and to the National Science Foundation (grant # ). The views epressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Econoic Research. NBER working papers are circulated for discussion and coent purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accopanies official NBER publications by Derek Kellenberg and Arik Levinson. All rights reserved. Short sections of tet, not to eceed two paragraphs, ay be quoted without eplicit perission provided that full credit, including notice, is given to the source.

2 Misreporting Trade: Tariff Evasion, Corruption, and Auditing Standards Derek Kellenberg and Arik Levinson NBER Working Paper No Septeber 2016 JEL No. F14 ABSTRACT In official international trade statistics, annual coerce between every pair of countries is reported twice: once by the iporting country and once by the eporter. These double reports provide an opportunity for audit. In principle, the two reported trade values should differ systeatically only by transport costs, because the values reported by iporters include freight and insurance. But in practice, after controlling for distance and other standard trade costs, the reaining gaps between iporter- and eporter-reported trade vary systeatically with GDP, tariffs and taes, auditing standards, corruption, and trade agreeents, suggesting that firs intentionally isreport trade data. These isreports have iplications for trade agreeents and doestic fiscal policy, and for epirical assessents of the efficacy of those policies. Derek Kellenberg Econoics Departent University of Montana 32 Capus Dr. #5472 Missoula, MT Arik Levinson Departent of Econoics ICC 571 Georgetown University 3700 O St., NW Washington, DC and NBER al6@georgetown.edu

3 Misreporting Trade: Tariff Evasion, Corruption, and Auditing Standards I. Introduction News stories about isreported trade periodically surface in the business press. In 2012 Chinese custos officials detained 50 people for underreporting seafood iports and evading $11.6 illion in tariffs. 1 In 2014 Meican authorities disrupted a tariff evasion ring that involved 22 custos officers and 197 copanies around the world and had cost Meico $37 illion in unpaid tariffs. 2 And later that sae year, Pakistani investigators caught custos officials plotting with an iporter to underreport the value of 110 containers of tetiles. 3 Coon thees in these stories include tariff and ta evasion, corruption, regulatory enforceent, and firs and organizations spanning countries at different levels of econoic developent. In this paper we show that this type of trade isreporting is widespread and varies systeatically with country characteristics such as econoic developent, taes and tariffs, corruption, and participation in regional trade agreeents (RTAs). The iportance of tariff evasion has not gone unnoticed by econoists. Fisan and Wei (2004) eained the effect of tariff rates on reported trade between China and Hong Kong. Subsequent papers have applied their approach to other groups of countries: eports fro Gerany to 10 transition econoies, iports to India, direct eports fro China to the United States, trade between the United States and Canada, iports to Kenya, Mauritius, and Nigeria, and ost recently, Tanzanian iports fro three developing country trading partners. 4 And although global assessents are scarce, one recent study estiated that in 2012, ore than $729 billion flowed out of developing countries as the result of trade isinvoicing and tariff evasion (Kar and Spanjers 2014). Such reports ostly confir the Fisan and Wei result that higher tariff rates lead to ore tariff evasion for the saples of countries studied. Like Fisan and Wei, they all identify evasion by coparing reported product-level tariffs with product-level trade data reported by eporters and iporters between select pairs of countries. As a result, even though the cobined latino.fonews.co/latino/news/2014/10/23/eico-disantles-tetile-industry-tariff-evasion-ring. 3 tribune.co.pk/story/801951/tariff-evasion-custos-seize-dry-port-records-to-probe-rs170-illion-sca. 4 See Javorcik and Narciso (2008), Mishra et al. (2008), Ferrantino et al. (2012), Stoyanov (2012), Bouët and Roy (2012), and Epaphra (2015). 1

4 body of evidence suggests that higher tariffs lead to ore isreporting, the results are not generalizable beyond the specific sets of countries eained. We don t really know the etent to which trade isreporting occurs across a broader set of countries or goods, whether it has been increasing over tie, or whether it is ore prevalent aong rich or poor countries. More iportantly, the focus on tariff evasion disregards other country characteristics that ay reward or penalize firs for isreporting trade. Many of these, like incoe taes, corruption, accounting standards, and capital controls, do not vary across industries within a country but do vary across countries and over tie. So their effects on isreported trade cannot be identified using isolated pairs of countries. A large panel of trade data and country characteristics, like the one we have assebled for this project, is necessary to identify the effects of these country characteristics on isreported trade. We begin by describing a odel in which firs choose how uch to isreport their iports or eports. Those isreports are functions of country characteristics like tariffs, corruption, taes, and the strength of auditing and accounting standards. We then use that odel, along with a siple accounting identity, to estiate the effects of those country characteristics on the gap between total reported eports and iports aong pairs of countries, using data on annual trade aong 126 countries fro 2002 to Our approach using aggregate trade, as opposed to the industry-by-industry easures used by Fisan and Wei (2004) and others, does coe with a trade-off. The aggregate approach cannot identify trade isreporting that results fro isclassifying products, say fro high-tariff to low-tariff categories, because those isreports are netted out in the aggregate. 5 What we observe is pure isreporting: cases where firs have reported to trade authorities a value different fro the true value transacted. In this sense, our analysis should be viewed as a conservative estiate of overall trade isreporting. In return, however, our aggregated approach has several advantages. First, by eaining aggregate trade aong any countries, we are able to estiate the effects of policy-relevant 5 For eaple, suppose etal chairs are being traded but wooden chairs have a lower tariff rate. An eporter ay correctly report $100,000 in etal chairs to the eporting country while the iporter reports $100,000 in wooden chairs to the iporting country to evade the higher tariff. This would be an eaple of tariff evasion through isclassification. But in aggregate, the isreported value would be zero. Both countries have reported $100,000 in chairs. If instead the iporter reported $50,000 worth of chairs (etal or wooden), then the aggregate isreported trade gap would be $50,000. 2

5 country characteristics other than tariffs. These include auditing and accounting standards, 6 corruption, participation in RTAs, and other doestic ta rates. Second, by netting out isclassification through aggregation and focusing on pure isreporting effects, we avoid difficult epirical challenges associated with disentangling the two in industry-level analyses. Third, when we do eaine tariffs, our estiation strategy allows us to estiate their average effect on isreporting across a wide variety of iporting and eporting countries. And finally, by aggregating across industries and using country variables as proies for isreports, we account for the potential endogeneity of true unobserved trade volues and those isreports. Based on a 126-country, 11-year panel, our epirical results confir prior findings that tariffs lead to underreporting of iports. We show, however, that the results are not unifor across countries. For pairs of countries that are both RTA ebers, with correspondingly low or zero tariffs, the easured tariff elasticity of isreporting is zero, as would be epected. For country pairs that are not RTA ebers, the tariff elasticity of isreporting is large, and it is larger for high-incoe countries than for low-incoe countries. A 1 percent increase in average tariff rates leads to a 3 percent increase in relative underreporting by high-incoe iporters but a 1 percent increase in relative underreporting by low-incoe iporters. As for the country characteristics other than tariffs, we find robust evidence that stronger auditing and accounting standards decrease the underreporting of eports (relative those sae shipents reported by iporters). A one standard deviation increase in an inde of those standards in the eporting country decreases the trade gap by approiately 3.8 percent. This finding is consistent across both high- and low-incoe countries and underscores the iportance and econoic benefits of strong auditing and accounting standards. We also find that corruption plays an iportant role in the degree of isreports for both iporters and eporters. A one standard deviation decrease in an inde of eporter corruption corresponds with a 12.5 percent decrease in the reported trade gap in both high- and low-incoe countries. Corruption is also associated with ore underreporting of iports by low-incoe countries. Finally, we find soe evidence that higher doestic ta rates correspond to ore underreported trade by eporting countries. 6 A few papers, such as Javorcik and Narciso (2008) and Mishra et al. (2008), find evidence of greater tariff evasion in product categories that are ore differentiated and therefore harder for custos officials to enforce. However, the ethod only indirectly easures enforceent and has only been applied to select product categories and countries. 3

6 These results are iportant for two reasons. First, tariff revenue reains an iportant source of funds for any governents, particularly in developing nations. Revenues lost to tariff evasion result in lower governent services or higher taes on capital or labor incoe (Ki and Kose 2014) and can result in higher international aid to developing countries (Collier and Venables 2011). In general, understanding how tariffs, corruption, and auditing standards affect isreporting can aid and iprove enforceent and tariff revenue collection. Second, although recent research suggests that corruption affects the real value of trade, 7 we show that corruption increases the isreporting of that trade. So soe of corruption s effect on trade easured previously ay be attributable to increased isreporting, not real trade reductions. Siilarly, regarding the trade creation and diversion effects of RTAs, 8 our results find substantially ore isreporting aong non-rta ebers than aong RTA ebers. So soe of the trade creation easured by prior work ay be attributable to less underreporting. Before outlining our fir-level theory of trade isreporting, in the net section we describe soe aggregate statistics that otivate the analysis. II. A Snapshot of Trade Misreports In the UN Cotrade data, which contain annual industry-level trade flows aong countries, each annual value is collected and reported twice once by the eporting country and once by the iporting country. 9 This project is otivated by the fact that the two values appear to differ, significantly and systeatically. Figure 1 plots the su of all global trade as reported by iporters and eporters separately. 10 Total trade reported by iporters is larger, as epected, because the iporter reports include cost, insurance, and freight (CIF) transport costs. But the difference is tiny, suggesting little aggregate isreporting. It turns out, however, that the global aggregates ask a lot of cross-country heterogeneity. To eaine the difference between the trade reported by iporters and eporters in ore detail and as a proportion of total reported trade, define the trade reporting gap as 7 See Dutt and Traca (2010) or Thede and Gustafson (2012). 8 Eaples include Magee (2008) and Carrére (2006). 9 The Cotrade data can be found at 10 Trade data are available for ore years than we eaine here, but our analysis in later sections is liited to by the availability of data on soe country characteristics, such as accounting and auditing standards and organized crie. 4

7 V trade reporting gap = V V + V. (1) where V is the annual total trade shipped fro eporting country to iporting country, as reported by the iporter (), and V is that eact sae annual value, as reported by the eporter (). Subscripts denote the eporting and iporting countries; superscripts denote the reporting country. The trade gap in (1) can in theory range fro 1 to +1, but in practice V and should differ only by CIF trade costs, and so the trade reporting gap should be sall and never negative. Figure 2 plots a histogra of the annual values of equation (1). Each observation is the annual gap between two countries, and each pair of countries appears twice each year: once for eports fro to, and once for the reverse. The ean and ode of the observations are clustered slightly above zero, as already deonstrated by Figure 1 and as epected by the iporter reports inclusion of CIF trade costs. But the spread is rearkable. For a significant fraction of the annual observations, the iporter reports are ore than twice or less than half of the eporter reports. 11 In fact, rather than declining continuously farther fro zero, the nuber of observations with trade gaps outside ±0.75 is increasing. When the gap is 0.75, reported iports are seven ties as large as eports. Clearly, there are lots of enorous, orders-ofagnitude differences between the two reports. 12 Our goal in this project is first to eaine whether those trade gaps are all erely accidental reporting errors or whether soe represent intential isreports by iporters or eporters seeking to avoid taes or tariffs, and then to eaine the role that corruption, accounting standards, econoic developent, and trade agreeents play in eacerbating or eliinating the degree of iseports. If trade reporting gaps arise fro intentional isreports, a natural otive is tariff evasion. Figure 3 plots the trade gap in 10 deciles organized by the size of the average tariff on those trades. Surprisingly, the trade gap grows for the first four deciles. Reported iports (which ust pay the tariff) eceed reported eports by ore the higher the tariff. Less surprisingly, for the highest si deciles of the tariff, as tariffs increase, reported iports shrink relative to reported eports. That hup-shaped pattern ay arise fro two offsetting forces. The higher the tariffs, V 11 When this happens, the trade gap is outside ±⅓. 12 In what follows we test several subsaples of our data in which we eclude cases where reported eports eceed reported iports by orders of agnitude, and vice versa. 5

8 the ore incentive the iporting country has to enforce accurate reporting, and the ore incentive the iporter has to underreport iports. In what follows we try to account for this by controlling for accounting and auditing standards along with tariffs. Figure 4 presents soe evidence that the isreports are correlated with these other country characteristics. Each panel in Figure 4 plots the average trade reporting gap by quintile of a different characteristic of the eporting country. For eaple, the upper left graph in Figure 4 shows that when eporters auditing standards are rated ore highly by the World Econoic Foru, the trade reporting gap is lower. 13 Since that gap is the iporter report inus the eporter report, we interpret the pattern in Figure 4 as indicating that eporters fro well-audited countries are less likely to underreport (as opposed to iporters fro well-audited countries being less likely to overreport, which is atheatically possible but akes less intuitive sense). Siilarly, in the upper right graph in Figure 4, eporters fro countries with ore corruption are ore likely to underreport. And in the botto two graphs, eporters fro poorer countries underreport ore, and eporters fro high ta countries underreport less, though that latter pattern ay be nononotonic for reasons siilar to tariffs in Figure 3. Figure 5 plots the sae four figures by iporter characteristics, displaying soe of the sae trends, though the patterns are less stark. And clearly, none of this is anything but suggestive. Tariffs are correlated with GDP. Both are likely correlated with corruption, accounting standards, and other country characteristics that affect trade isreports, and all of those country characteristics have been changing over tie. To study how trade reporting responds to country characteristics ore systeatically, we need an epirical odel that ebodies the incentives for firs to isreport their iports and eports. III. A Fir-Level Model of Trade Misreporting Our starting point is a generalized version of the odel in Ferrantino et al. (2012). In each industry i and year t there is a representative eporting fir in country and a representative iporting fir in country. The firs privately negotiate and know the true * value of free-on-board (FOB) eports, V, but this value is not known by custos officials, as it signified by the superscript asterisk (*). This FOB value is the true value of goods when they

9 arrive at the eporting port before shipping to the iporting country. 14 We ake no assuption about the underlying arket structures, copetitive or otherwise. We siply assue that each fir chooses prices and quantities to aiize profits given the arket structure and copetition they face, resulting in a privately negotiated equilibriu value for (unobserved) * V. it Each eporting fir has to decide how uch of its eports to report to authorities, Let δ be the proportional deviation fro the true value V * that the eporter reports: V it * it V it. = (1 + δ ) V. The superscript denotes that δ is the eporter s deviation. Firs ay choose to underreport eports (δ <0), report accurately (δ =0), or overreport (δ >0). These deviations ay coe in the for of isreporting prices (as with transfer pricing) or quantities. 15 Why would firs isreport eports? Doing so has both costs and benefits. First, firs ay incur product, sales, or service taes or subsidies on the proceeds of goods eported, giving the incentives to under- or overreport those eports. Second, in an effort to stabilize currency and capital accounts, soe countries restrict capital eports or iports (Prasad and Rajan 2008). To evade those capital controls, firs ay isreport trade transactions (Patnaik et al. 2012). Third, firs in countries with ore corruption or organized crie ay have to pay bribes or face etortion if they isreport the true value of eported goods (Dutt and Traca 2010). Finally, firs in countries with stricter accounting and auditing standards ay have a harder tie isreporting trade. Each of these otivations can be represented by a function representing the net benefit of isreporting: * B = B ( δ, V : Z ), (2) where the industry and year subscripts are suppressed. The benefit (or cost) of isreporting eports is a function of the size of the isreport δ and the true value of eports V *, both of which are chosen by the eporting fir, and a set of country characteristics Z. These include 14 For now, we just assue that these trades are negotiated in US dollars. The Cotrade data we use are typically reported in national currencies and then converted to dollars by the UN using average annual trade-weighted echange rates. 15 Ferrantino et al. (2012) assue that the quantity of trade q is fied and that the two firs negotiate over price p. Either way, in the net step we take this arket-deterined value, V=p q, as given and odel the reduced for decision about what fraction of that V to isreport. In our odel, it does not atter whether δ ultiplies p or q as it has the sae effect on V. 7

10 taes, capital controls, auditing standards, and levels of corruption. For any given value of true eports, V *, firs choose isreporting δ to aiize (2), resulting in an iplicit function: δ = δ Z. (3) * ( V : ) On the iporting side, firs face all the sae benefits and costs of isreporting as eporters plus one iportant additional incentive: tariffs, τ. As with eports, we can think of iporters choosing true iports V * and isreports δ to aiize the net benefit (or cost) of isreporting: * B = B ( δ, V :, τ Z ). (4) And as with eports, for a given true value of trade, this defines an iplicit function for optial isreported iports: * ( V :, ) δ = δ τ Z. (5) Obviously we cannot estiate (3) or (5) directly. We don t observe the true value of trade V * or the isreports δ and δ. But we can use the country characteristics in Z as proy variables to estiate the effects of those characteristics on aggregate trade isreporting. So the net step is to convert (3) and (5) into equations that can be estiated, which we do using a siple accounting relationship between reported trade and the unobserved true value of trade. Observable reported eports V it in any industry (i) and year (t) ust be equal to the true unobservable trade value ties the eporter s isreport and a rando error ter ( ε we assue to be ultiplicative and lognorally distributed. Siilarly, reported iports ( 1 ) * it it it it it ), which V = V + δ ε (6) V it ust equal the true trade value, adjusted for isreports, an error ter, and CIF trade costs σ it > 1 that we assue are also ultiplicative. ( 1 ) V = V σ + δ ε (7) * it it it it it Equations (6) and (7) are siply definitions, but V *, as well as the δs and εs, reain unobserved in both. 8

11 To translate (6) and (7) into soething we can estiate epirically, we first su all industries for each year and pair of countries. Aggregate true (unobserved) trade between any two countries in year t is just V n * * t Vit i= 1 = ; aggregate CIF costs are σ t n = σ ; and i= 1 it aggregate rando reporting errors are ε eports in (6) can be rewritten t n n it t i= 1 i= 1 = ε and ε ( 1 ) * t t t t = ε. Aggregate reported it V = V + δ ε, (8) where V t n = V and i= 1 it δ t V ε (9) n * it it = δ * it i= 1 Vtε t is the average degree of isreporting, weighted by the industry-level eports and reporting errors. Siilarly, aggregate reported iports in (7) becoe ( 1 ) V = V σ + δ ε (10) * t t t t t where V t n = V and i= 1 it δ t V σ ε (11) n * it it it = δ * it i= 1 Vtσtεt is the average degree of isreporting, weighted by industry-level iports, reporting errors, and CIF costs. Equations (8) and (10) are accounting identities showing that custos officials, and therefore researchers, observe only the reported aggregate values of FOB eports, iports, V, and CIF t V, which are functions of three unobserved country characteristics: (i) the true value t of goods traded, (ii) CIF trade costs, and (iii) a weighted average of eporters and iporters intentional isreports. 9

12 The final step is to take the ratio of aggregate reported iports in equation (10) to the aggregate reported eports in equation (8), yielding V V ( 1+ ) ( 1+ δ ) ε σ δ ε t t t t = t t t. (12) Taking the log of both sides of equation (12) we get where ε t t ( ) ( ) lnvt lnvt = lnσt + ln 1+ δt ln 1+ δt + εt, (13) = ln ε ln ε is now a norally distributed error ter for shipents fro to in t year t. We call the left-hand side of (13) the log trade gap, and it fors the baseline for all of our epirical analyses. In theory, even if we had data on σ, δ, and δ t t 10 t, we could not consistently estiate the log trade gap in (13) because those weighted average isreports ( δ and δ theselves still functions of the unobserved true values of trade aong industries (see equations (9) and (11)). To address the fact that we do not observe those δ s, and that even if we could, they would be endogenous, we rely on a proy variable estiation strategy to obtain consistent estiates of iportant characteristics that influence firs decisions to isreport (Wooldridge 2010). t t ) are As a refresher, proy variables can itigate oitted variable bias like that in equation (13). Consider a variable z to be used as a proy for one of the issing δs in (13). To be a sucessful proy, z ust be redundant. That is, E(lnVt ln Vt δ, z) = E(lnVt ln Vt δ). (14) Or put differently, the epected trade gap estiated with and without the proy z is the sae, once we ve conditioned on the other right-hand variables (X), including the issing one (δ). In what follows, we describe proies for the issing δs: country characteristics that change the costs or benefits to firs of isreporting their eports or iports. These include local taes, tariffs, corruption, and accounting standards the eleents of Z and Z in equations (3) and (5). The idea is siilar to a standard instruental variables approach, where instead of having an endogenous right-hand variable, we have an unobserved endogenous right-hand variable. Proy variables, like instruents, ust be correlated with the endogenous variable and

13 uncorrelated with the dependent variable ecept echanically through the endogenous righthand variable. And the epirical strategy is even sipler than for instruents. We siply substitute proy variables for the δ s in equation (13). That leaves us unable to estiate the direct effect of isreporting on the log trade gap but does allow us to consistently estiate the effect of the proy variables on the log trade gap. As an auditing check on the etent of isreporting, that suffices. If there is no intentional isreporting, the δ s are zero, and the coefficients on the proies should be zero. In our case, the proies for the δ s are just easures of the country characteristics, Z, that deterine firs optial isreports in (3) and (5): taes and tariffs, corruption and auditing standards, and trade agreeents. The critical assuption is that these country characteristics, Z, affect only the left-hand side of (13) through their effect on the δ s. By taking the ratio of reported aggregate iports to eports, we have canceled out the true value of trade fro our estiating equation, ecept where it appears in the weighted average of those isreports. But since we don t observe the isreports anyway, only proies for isreports, we can estiate equation (13) consistently using those proies in lieu of the δ s. lnvt lnvt = β 0 + βσ 1 t + β 2Zt βz 3 t + µ t (15) Equation (15) is essentially an auditing forula. If there is no isreporting, the difference between the value of trade shipped fro to should reflect only trade costs (σ) and a ean-zero rando error (μ). That difference should not be a function of tariffs, accounting standards, taes, or corruption. The coefficients in β 2 and β 3 should be zero. Most iportantly, we can estiate equation (15). We have data on coponents of trade costs, σ, which are siply the standard eleents of a gravity odel of international trade such as distance, coon borders, and language. And we have data on easures of country characteristics, Z. We describe both in the section that follows. IV. Data and Estiation Descriptive statistics for all the data we use are in Table 1. With 178 countries over 11 years, there are 346,566 possible annual trade flows in the Cotrade data used. 16 But trade was reported in only about half of those. Unreported trade occurs either because countries do not

14 A1. 17 Our proies for CIF trade costs are the standard coponents of a typical gravity odel of trade with each other or because countries fail to report their trade. In no cases was there coerce reported by the eporter but not by the iporter, or vice versa. In what follows, our analyses include the data only for years and country pairs with reported coerce. This reduces our potential available observation set to 165,215. We also drop countries with issing audit, corruption, tariff, or ports data, and we drop cases with too few iporters, eporters, or country pairs to identify country-pair fied effects. Even without these observations, our final data set of 86,185 observations contains ore than 87 percent of the reported iports for the 165,215 potential observations for which we have data during the tie period. A detailed accounting of how we arrive at the 86,185 observations described in Table 1 can be found in Appendi Table international trade. 18 These include distance between countries and whether they share coon borders or languages. 19 In addition, we include the quality of port infrastructure in the two countries, based on survey data fro the Global Copetitiveness Report (GCR) for the years The survey question relating to port quality can be found in Appendi Table A2. Port quality is scaled fro 1 to 7, and we define a easure of port quality for each country pair by ultiplying the iporter and eporter scores. 21 Weighted average tariff rates for each country and year were obtained fro the World Bank Developent Indicators (WBDI). 22 Duy variables for whether two countries are 17 Trade data are available for any ore years than we eaine in this paper, but our analysis is constrained to by the availability of proy variable data on the strength of accounting and auditing standards and corruption. 18 See Head and Mayer (2013) for a survey of the gravity odel literature and variables used for approiating transport costs. 19 Distance, coon border, and coon language are fro the CEPII gravity database, which can be found at 20 Country-level characteristics fro survey questions fro the GCR have been used etensively in epirical studies on international trade and investent. See Carr et al. (2001), Yeaple (2003), Javorcik and Wei (2004), Eckhol et al. (2007), and Kellenberg (2012) for a few of any eaples. 21 Better or ore efficient port infrastructure decreases CIF costs (Clark et al. 2004). 22 Siple unweighted average tariff rates are also available fro the WBDI. However, the siple correlation between unweighted and weighted average tariff rates is All the regressions in the following discussion were also run with the unweighted average tariff rates, with virtually identical results. Thus, we present only the results using the weighted average tariff rates theoretically a ore accurate easure of countries overall average tariff rates. 12

15 ebers of an RTA, which ay be a free trade area, custos union, or econoic integration agreeent, were generated using a data-generating progra described in de Sousa (2012). 23 Like port quality data, country data on average accounting and auditing standards coe fro the GCR surveys. The specific questions relating to the strength of auditing and accounting standards can be found in the second row of Appendi Table A2. This variable is also on a 1-to-7 scale, with higher scores indicating stronger standards. Corruption data coe fro Transparency International s Corruption Perceptions Inde, 24 which uses annual survey data to easure the perceived level of public corruption. The inde ranges fro 0 (very corrupt) to 10 (not corrupt). Table 2 presents several estiates of equation (15). Colun (1) estiates the siplest version of the basic odel where we have included a single proy variable for each of Z, and Z, : port quality, tariffs, and eporter auditing and accounting standards. All three t t coefficients are statistically significant and negative, as epected. Higher tariffs increase the incentive to underreport iports, decreasing the gap between reported iports and eports. Better ports ean lower trade costs, shrinking the gap between true CIF iports and FOB eports. And eporting countries auditing and accounting standards reduce the ability of firs to understate eports, shrinking the log trade gap. In colun (2) we add ore proies for trade costs: distance and indicators for coon borders and coon languages. All three proies are statistically significant and of the epected signs. Distance increases trade costs, increasing the log trade gap, while coon borders and language reduce trade costs with the opposite consequence. In coluns (3) through (5) we add iporter accounting standards an indicator for when both countries are ebers of an RTA as well as an interaction ter between tariffs and RTAs. Better eporter accounting standards reduce the ability of firs to understate eports, shrinking the trade gap, while better iporter accounting standards results in less underreporting of iports, increasing the trade gap. The effect of RTAs on the reported trade gap in coluns (4) and (5) is negative. The estiates across all five specifications in coluns (1) through (5) are rearkably stable, even as we add additional proies to the odel. This stability provides a degree of confidence in the reliability of our proy variable estiation strategy. σ t, 23 The Stata do-files for the RTA progra can be found at These files have been used in several prior studies on RTAs, such as Head et al. (2010), Baghdadi et al. (2013), and Head and Ries (2010)

16 In colun (5) of Table 2 we add the interaction between tariffs and RTAs. The coefficient is positive and significant, indicating that the effect of tariff evasion on the reported trade gap is saller for country pairs that are ebers of an RTA. At the botto of colun (5) we calculate the cobined effects of tariffs for RTA ebers ( 1.69). 25 A 1 percent increase in tariffs decreases the reported trade gap by 1.7 percent. Though statistically significant, this effect is uch saller than the 3.4 coefficient for country pairs that are not ebers of an RTA. That is, the tariff evasion effect is saller between countries that are ebers of an RTA and presuably face lower average tariff rates than for country pairs that are not RTA ebers and face higher average tariff rates. Once we control for the different effects that RTAs have directly and through tariffs, the net effect of RTAs is both negative and statistically significant. Even with a variety of proies included in the estiation in coluns (1) through (5), unobserved heterogeneity across tie or specific country pairs could still bias the estiates. If there are unobserved country characteristics that influence reporting gaps and happen to be correlated with one of our proies, then our estiates ay be biased. For eaple, if stronger auditing and accounting standards across tie happen to be correlated with unobserved technological advanceents in coputing and software capabilities in all countries, and this leads to a decrease in the rando error coponent of isreporting, then our proy estiates on auditing and accounting standards ay be biased. Likewise, if soe trading partners have cooperated to ake sure that the data reported by eporters and iporters atch, that ight shrink the reported trade gap. 26 To the etent that this cooperation is correlated with other country characteristics, our estiates ay be biased. In colun (6) of Table 2 we control for unobserved tie-specific heterogeneity by including year duies. Although soe of the coefficient estiates change slightly fro colun (5), the results are quite siilar, indicating that the effect of any bias fro unobserved heterogeneity over tie is inial. In colun (7) we add country-pair fied effects. Two things are iportant to keep in ind here. First, all estiates are now identified fro changes in variables within country pairs, as opposed to the cross-country pairs estiates in coluns (1) through (6). Second, since we are controlling for tie-invariant unobserved heterogeneity, we 25 The cobined effect of tariffs for RTA ebers is the su of coefficients for [Iporter Tariff ] + [Iporter Tariff RTA]. For colun (5) of Table 2 this is = An F-test was conducted to deterine the joint significance of the su. 26 Since 1990 the United States and Canada have echanged iport data and substituted each other s reported iport data for their own reported eport data, which are typically less reliable (Stoyanov 2012). 14

17 can no longer identify other tie-invariant characteristics, such as distance or coon borders and languages, since they are absorbed by the country-pair fied effects. Only proies that vary across tie within country pairs can be identified. Moreover, soe of the proies that do vary over tie don t vary by uch, leaving little variance for identification. The coefficients on port quality, RTAs, and iporter accounting standards are no longer statistically significant. The coefficients in colun (7) on tariffs and eporter accounting standards reain statistically significant, though their agnitudes change soewhat. The effect of tariffs for countries that are not ebers of the sae RTA is 1.21, but the effect for RTA ebers is not statistically different fro zero ( 0.18). The coefficient on eporters auditing and accounting standards reains negative and significant: a one standard deviation increase in eporters auditing and accounting standards decreases the trade gap by 5.2 percent. The regressions in coluns (3) through (6) of Table 2 are repeated in coluns (1) through (4) of Table 3, using the corruption inde as a proy for isreporting rather than auditing and accounting standards. The agnitudes and significance of the coefficients on tariffs, ports, distance, coon language, coon border, and RTAs reain nearly identical to the results in Table 2. Corruption also has statistically significant effects on the trade gap. A one standard deviation decrease in corruption (i.e., a higher corruption inde score) in the eporting country translates to a 7.7 to 13 percent decrease in the trade gap, whereas a coparable decrease in corruption in the iporting country iplies approiately a 2.8 percent increase in the trade gap. This is consistent with the hypothesis that countries that have less corruption, or do a better job of controlling corruption, ake it ore difficult for eporters or iporters to underreport true trade values. In colun (5) of Table 3 we estiate the odel with both auditing and accounting standards and corruption as proies for isreporting. The results are siilar for all variables ecept that the signs on the coefficients for the iporter and eporter auditing and accounting variables have reversed. This is a bit strange, but given the high siple correlation of the auditing and accounting standards variables and the corruption variables (ρ=0.83), and the fact that the coefficients are identified across country pairs, we suspect a ulticollinearity bias. In colun (6) we again incorporate country-pair fied effects. Here, the coefficients are now identified within country pairs over tie, reducing the collinearity observed across country pairs in colun (5). Both better auditing and accounting standards and less corruption in the eporting 15

18 country now have the epected negative effect on the the trade gap; the results are positive but no longer statistically significant for iporters. It is iportant to note that the country-pair fied effects regressions in colun (7) of Table 2 and coluns (4) and (6) of Table 3 have far greater eplanatory power than the regressions without country-pair fied effects, indicating substantial unobserved country-pair heterogeneity in the trade gap. In the regressions that follow, we constrain our analysis to regressions that contain country-pair fied effects. In Table 4 we check the robustness of the results with respect to outlier observations. Recall fro Figure 2 that the ean of the reported trade gap is clustered slightly above zero, but the gap increases again for values ±0.75. To ensure that our results are not driven by the tails of the distribution, we repeat our ost restrictive specification in colun (7) of Table 3 using different cutoff values for the tails of the distribution: we drop observations where the trade gap is ±0.99, ±0.75, and ±0.5. Although the agnitude of the coefficients on tariffs and eporter auditing and accounting standards and corruption falls as the saple size is further restricted, the coefficients reain econoically and statistically significant, indicating that their effects are not being driven solely by etree country-pair observations. V. WTO Mebership, Product Taes, and Capital Controls In Table 5 we eplore additional versions of equation (15). Colun (1) adds a duy for World Trade Organization (WTO) ebership as well as an interaction with tariff rates to see whether WTO ebership has had siilar effects on the trade gap as RTA ebership. 27 Although WTO ebership does not reduce average tariff rates to the etent a free trade agreeent or custos union ight, ost-favored-nation status ay lower average tariff rates for ebers and provide benefits siilar to RTA ebership. However, 90 percent of the annual trade flows in our data were between pairs of contries that were WTO ebers over the entire saple period, aking any WTO effect difficult to identify. This ay eplain the insignificance of both the direct WTO coefficient and the WTO effect through tariffs. Nonetheless, the insignificant net effect of WTO ebership on the overall trade gap is consistent with Javorcik and Narciso (2013), who also found that WTO ebership has led to no net difference in the reported trade gap for 15 recent eber countries. Iportantly, paraeter estiates on tariffs, 27 Data on World Trade Organization ebership were obtained fro accession dates on the WTO website, at 16

19 auditing and accounting standards, and corruption in the eporting country reain consistent and significant. In coluns (2) and (3) of Table 5 we eaine product ta rates. Higher product taes in the eporting country give firs the incentive to understate eports, increasing the reported gap easured as V V. But higher product taes in the iporting country lead to understated iports, shrinking the reported trade gap. We calculate net product ta rates as the ta revenues (inus subsidies) that are related to the sale, production, or use of goods and services in a country, divided by GDP. This easure provides a proy of the net ta rates on goods and services faced by iporters and eporters across countries. 28 Data coverage on ta rates, however, is not as etensive as for the other variables in Table 5, with only 64,429 observations available. Moreover, ost of the lost observations coe fro less developed countries with less reliable ta data. Given that we are not only adding ta rates as additional proies but also substantially altering our saple, in colun (2) we first run the sae regression as in colun (7) of Table 3 to understand how changing the saple size affects the results. Then in colun (3) we add net taes on products for both eporters and iporters. Despite the decrease in observations, the results in colun (2) are rearkably siilar to those using the full saple. In colun (3) of Table 5 the coefficient on eporter net product taes (1.34) is large, of the epected sign, and statistically significant. A 1 percent increase in eporting country ta rates increases eporters underreporting and leads to a 1.34 percent increase in the reported trade gap. Finally, in coluns (4) through (5) of Table 5 we include proies for capital controls. In theory, capital eport controls give firs incentives to understate eports, increasing the trade gap ( V V ), whereas capital iport controls give firs the incentive to understate iports, decreasing the trade gap. Although we don t have data on the strength of capital controls, one potential proy is inflation rates. Inflation can be correlated with large capital inflows that ake it difficult for countries to control onetary policy without capital control restrictions (Cordero 28 Data on both net taes on products and GDP are fro the WBDI database. We also eplored but do not present here other ta rate proies found in the WBDI, including easures of value added ta revenues and overall ta revenues as a percentage of GDP. However, these other easures were always insignificant, did not qualitatively change the results of any of the other relevant variables in the odel, and suffered fro ore severe issing data probles. 17

20 and Montecino 2010; Lartey 2012). And Leen and Eijffinger (1996) find a positive correlation between inflation rates and easures of capital controls. Again, incoplete inflation data across countries reduce our available saple size to 83, To ensure that any changes in coefficient estiates are not the result of the changing saple, in colun (4) of Table 5 we again estiate the base odel of colun (7) in Table 2 on the new saller saple. The effects of tariffs and eporter accounting standards and corruption reain stable and statistically significant. In colun (5) the signs on eporter and iporter inflation rates are of opposing signs, consistent with firs having incentives to use trade isreporting to ove capital. These effects are not statistically significant, however. Iportantly, the estiates on tariff rates and eporter accounting standards and corruption continue to reain consistently significant and robust to changing saples and the inclusion of alternative proies. VI. Different Effects for Rich and Poor Countries As an additional check on our results, we eplore whether tariffs, auditing and accounting standards, and corruption have different effects on countries of different developent levels. Defining high-incoe countries using the World Bank s classification, we interact an indicator for these countries with our tariff, auditing and accounting, and corruption easures. The interaction ters tell us how the arginal effects of high-incoe countries differ fro the arginal effects of lower-incoe countries, which we define as all countries not classified as high incoe. The results of the interactions are presented in Table 6, where colun (1) includes interactions with tariffs, colun (2) has interactions with accounting standards, and colun (3) has interactions with corruption. In general, there appears to be no difference in the effects of auditing and accounting standards for high- and lower-incoe countries. There are, however, intriguing differences in tariff evasion and iporter corruption effects. To facilitate discussion of these differences, Table 7 reports the cobined arginal effects of high- and lower-incoe countries fro the regression results in Table 6. For both country categories, tariffs have a negative and statistically significant effect on the reported trade gap when a country iports fro a non-rta trading partner, but the effect in high-incoe countries is three ties larger than the effect in lower-incoe countries. This difference is both 29 Inflation rate data also coe fro the WBDI database. 18

21 statistically and econoically significant. However, when trading partners are ebers of the sae RTA, for lower-incoe iporters the tariff effect is not statistically different fro zero, but for high-incoe iporters it reains significant. The results indicate that tariff evasion is ore prevelant in high-incoe iporting countries both within and outside RTAs. We cannot say why there is such a discrepancy in the trade gap between high- and lower-incoe iporting countries with respect to the tariff elasticity. However, given that high-incoe countries have ore ultinational firs 30 and consue a greater proportion of differentiated products with larger arkups, 31 and given that those arkups create ore opportunity for transfer pricing within firs, 32 we speculate that ultinational fir transfer pricing in high-incoe iporting countries ay play a role in the larger trade elasticity gap. However, testing this is outside the scope of this paper, given ulticountry panel data liitations for ultinational enterprises. The arginal effect of stronger auditing and accounting standards is statistically significant in both high- and lower-incoe countries, with the effect in lower-incoe countries slightly stronger. The difference between the two, however, is not statistically significant. Finally, the arginal effects of corruption in the eporting countries is siilar and statistically significant for both high- and lower-incoe countries. Additionally, corruption has a positive and significant 33 sign for lower-incoe iporting countries, as would be epected (recall that the corruption inde score is higher for less perceived corruption). Less corruption is likely correlated with less underrreporting of iports, which would increase the reported trade gap. In this case, a one standard deviation increase in the corruption inde in a low-incoe iporting country leads to a 3.7 percentage increase in the reported trade gap. 34 Our results suggest that although iproveents in auditing and accounting standards across countries can substantially reduce the etent of isreported trade, the eliination of corruption can have even larger effects, particularly for lower-incoe countries. VII. Conclusion This paper finds evidence of intentionally isreported trade by eaining the difference between trade reported by eporters and iporters. Our approach differs in several ways fro 30 Markusen (2002). 31 Markusen (2013). 32 Clausing (2003). 33 Significant at the 10 percent level. 34 The standard deviation in the log of the corruption perception inde for lower-incoe countries is

22 prior studies. First, by using a panel of data on aggregate trade aong hundreds of countries over 11 years, rather than detailed industry data for specific pairs of countries or regions, we can test for general patterns of tariff evasion. In this sense, our results apply to a broad range of countries at different levels of developent and across all sectors of the econoy. Second, we can break down those general results and copare tariff evasion in rich and poor countries as well as aong trade agreeent ebers and nonebers. And third, we can eaine country characteristics aside fro tariffs that cause firs to isreport trade, including taes, auditing standards, corruption, and capital controls. For countries that are not ebers of regional trade agreeents, higher tariffs are associated with significant underreporting of iports. And for high-incoe non-rta ebers, the tariff evasion elasticity is three ties larger than for low-incoe non-rta ebers. For country pairs that are ebers of the sae RTA, the tariff evasion effect disappears, although not copletely for high-incoe iporting countries. Tariff evasion is siply not a relevant otivation for isreporting when average tariff rates are near zero for ost countries. The results indicate that tariff evasion effects are dependent on both the level of developent across countries as well as their participation in RTAs. In addition, we find robust evidence that underreporting of eports increases when auditing and accounting standards are lower and eporters ta rates are higher. Likewise, eliinating corruption decreases isreporting for both iporters and eporters, particularly in low-incoe countries. These last results confir that when we look beyond disaggregated tariff rates, policyrelevant country characteristics such as doestic ta rates, the strength of auditing and accounting standards, and corruption are significant in eplaining firs otivations for underreporting of eports. These factors ust be taken into account by governents and custos authorities seeking to protect revenue sources associated with tariff and ta evasion on trade across international borders. 20

23 References Baghdadi, Leila, Iaculada Martinez-Zarzoso, Celestino Suarez-Burguet, and Habib Zitouna Are RTA agreeents with environental provisions reducing eissions? Journal of International Econoics 90(2): Bouët, Antoine, and Devesh Roy Trade protection and ta evasion: Evidence fro Kenya, Mauritius, and Nigeria. Journal of International Trade & Econoic Developent: An International and Coparative Review 21(2): Carr, David L., Jaes R. Markusen, and Keith E. Maskus Estiating the knowledgecapital odel of the ultinational enterprise. Aerican Econoic Review 91(3): Carrére, Celine Revisiting the effects of regional trade agreeents on trade flows with proper specification of the gravity odel. European Econoic Review 50(2): Clark, Xiena, David Dollar, and Alejandro Micco Port efficiency, aritie transport costs, and bilateral trade. Journal of Developent Econoics 75(2): Clausing, Kiberly A Ta-otivated transfer pricing and US intrafir trade prices. Journal of Public Econoics 87(9): Collier, Paul, and Anthony J. Venables Illusory revenue: Iport tariffs in resource-rich and aid-rich econoies. Journal of Developent Econoics 94: Cordero, José Antonio, and Juan Antonio Montecino Capital controls and onetary policy in developing countries. CEPR working paper. de Sousa, Jose The currency union effect on trade is decreasing over tie. Econoics Letters 117(3): Dutt, Pushan, and Daniel Traca Corruption and bilateral trade flows: Etortion or evasion. Review of Econoics and Statistics 92(4): Ekhol, Karolina, Rikard Forslid, and Jaes R. Markusen Eport platfor foreign direct investent. Journal of the European Econoic Association 5(4): Epaphra, Manaba Ta rates and ta evasion: Evidence fro issing iports in Tanzania. International Journal of Econoics and Finance. 7(2): Ferrantino, Michael J., Xuepeng Liu, and Zhi Wang Evasion behaviors of eporters and iporters: Evidence fro the U.S. China trade data discrepancy. Journal of International Econoics 86:

24 Fisan, Rayond, and Shang-Jin Wei Ta rates and ta evasion: Evidence fro issing iports in China. Journal of Political Econoy 112(2): Head, Keith, and Thierry Mayer Gravity equations: Workhorse, toolkit, and cookbook. CEPII working paper Head, Keith, and John Ries Do trade issions increase trade? Canadian Journal of Econoics 43(3): Head, Keith, Thierry Mayer, and John Ries The erosion of colonial trade linkages after independence. Journal of International Econoics 81(1): Javorcik, Beata S., and Gaia Narciso Accession to the World Trade Organization and tariff evasion. CEPR discussion paper Javorcik, Beata S., and Shang-Jin Wei Pollution havens and foreign direct investent: Dirty secret or popular yth? Contributions to Econoic Analysis & Policy 3(2): Article 8. Kar, Dev, and Joseph Spanjers Illicit financial flows fro developing countries: Global Financial Integrity report. Available at Kellenberg Trading Wastes. Journal of Environental Econoics and Manageent 64(1): Ki, Sunghyun H., and M. Ayhan Kose Welfare iplications of trade liberalization and fiscal refor: A quantitative eperient. Journal of International Econoics 92: Lartey, Eanuel K.K Financial openness, nontradable inflation and optial onetary policy. Econoics Letters 117: Leen, Jan J.G., and Sylvester C.W. Eijffinger The fundaental deterinants of financial integration in the European Union. Weltwirtschaftliches Archiv 123(3): Magee, Chris New easures of trade creation and trade diversion. Journal of International Econoics 75: Markusen, Jaes R Multinational firs and the theory of international trade. Cabridge, MA: MIT Press Putting per-capita incoe back into trade theory. Journal of International Econoics 90(2):

25 Mishra, Prachi, Arvind Subraanian, and Petia Topalova Tariffs, enforceent, and custos evasion: Evidence fro India. Journal of Public Econoics 92: Patnaik, Ila, Abhijit Sen Gupta, and Ajay Shah Deterinants of trade isinvoicing. Open Econoies Review 23: Prasad, Eswar S., and Raghura G. Rajan A pragatic approach to capital account liberalization. Journal of Econoic Perspectives 22(3): Stoyanov, Andrey Tariff evasion and rules of origin violations under the Canada-U.S. Free Trade Agreeent. Canadian Journal of Econoics 45(3): Wooldridge, Jeffrey Econoetric analysis of cross section and panel data. Cabridge, MA: MIT Press. Yeaple, Stephen R The role of skill endowents in the structure of U.S. outward foreign direct investent. Review of Econoics and Statistics 85(3):

26 Figure 1: Total Global Trade as Reported by Iporters and Eporters Figure 2: Trade Reporting Gap by Country-Pair Year 24

27 Figure 3: Trade Reporting Gap by Tariff Figure 4: Trade Reporting Gap by Eporter Characteristics 25