Abstract

Abstract. This paper examines the relationship between offshoring activity by U.S. multinational firms and the structure of U.S trade preferences. We combine firm level panel data on U.S. foreign affiliate activity from the U.S. Bureau of Economic Analysis (BEA) with detailed measures of U.S. trade preferences from the U.S. International Trade Commission (USITC) to create a three-way panel that spans 80 industries, 184 countries, and ten years (1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006). Consistent with existing theory, we find that offshoring and preferential market access are positively and consistently correlated, both in the pooled sample and within countries, industries, and years. Using both instrumental variables and simultaneous equations approaches to address the endogeneity of export-oriented foreign investment, we find that a 10% increase in U.S. foreign affiliate exports to the U.S. from a particular industry and country is associated with a 4 percentage point increase in the rate of preferential duty-free access. Restricting attention to the Generalized System of Preferences (GSP) among potentially eligible developing countries, we find that the influence of multinational affiliate sales on preferential market access more than triples. Date: August 26, 2013. The statistical analysis of firm-level data on U.S. multinational companies was conducted at the Bureau of Economic Analysis, U.S. Department of Commerce under arrangements that maintain legal confidentiality requirements. The views expressed are those of the authors and do not reflect official positions of the U.S. Department of Commerce. We thank Bill Zeile and Ray Mataloni of the Bureau of Economic Analysis for sharing their expertise, Shushanik Hakobyan, Martin Bresslein, and Ashley Ulrich for superb research assistance, and Andrew Bernard, Bruce Blonigen, Matt Cole, Jason Fletcher, Wayne Gayle, James Harrigan, John McLaren, Matthew Slaughter, and seminar and conference participants for helpful comments. Blanchard gratefully acknowledges funding from the Bankard Fund for Political Economy at the University of Virginia. Tuck School of Business at Dartmouth College; emily.j.blanchard@tuck.dartmouth.edu. Fachbereich IV-Volkswirtschaftslehre, Universität Trier (Germany); matschke@uni-trier.de. U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 1 Overview Recent theoretical work suggests that the pattern of international investment and multinational enterprise (MNE) activity may play an important role in shaping government preferences over trade policies: when a multinational firm owns export-oriented (i.e. 'offshoring') affiliates abroad, the MNE's 'home' country government has an incentive to improve market access for imports from its MNEs' foreign affiliates, for the simple reason that greater market access means higher rates of return to the government's MNE constituents. To the extent that governments tailor their commercial policies in response to the interests of constituent industries, differences in the pattern of firm operations across the globe may be reflected in trade policy. These ideas are formalized in two theory papers, Our empirical strategy sidesteps these difficulties by focusing on the potential influence of MNE activity not on MFN tariffs, but on the recent proliferation of various preferential trade agreements and the Generalized Systems of Preferences (GSP) by which industrial nations grant developing countries facilitated market access. Preferen- tial treatment of trade flows is exempt from MFN (under Article XXIV and the Enabling Clause of the GATT, for FTAs and GSP respectively), and therefore may be considered a closer reflection of a government's unilateral trade preferences. 1 A second potential complication for empirical testing lies in differentiating exportoriented (vertical) FDI apart from import-substituting (horizontal) FDI. While theory predicts that export-oriented FDI will exert downward pressure on tariffs in the investment-source country, the converse holds for market-seeking investment. To the extent that multinationals operate horizontal 'tariff jumping' operations abroad, those 1 While in principle Article XXIV and the Enabling Clause limit the extent of unilateral discretion, the intended uniformity (across industries and/or countries) of MFN exemptions appears to be weak in practice. (See Blanchard and Hakobyan (2012) for a detailed accounting of the sources and extent of discretion under the GSP.) Indeed, were bilateral free trade deals truly non-discriminatory across industries -or were GSP preferences in practice determined only by country and industry (but without discretion at the level of the country-industry pair or year-to-year variation) -our results would not withstand country-, industry-, and year-fixed effects, as they do. See the discussion in Section 3.1 for more background on the institutional structure governing U.S. trade preferences, and U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 3 activities will have either a negligible or small positive effect on the investment-source country's optimal tariff. 2 Indeed, market-seeking FDI and trade protection are often positively correlated in practice, largely due to the reverse causality as firms circumvent protectionist barriers abroad by building tariff jumping factories in their target market. 3 Fortunately, the richness of the BEA data offers us an empirical solution. In our data, MNE sales are disaggregated by destination; we are thus able to distinguish exportoriented investment (measured as U.S. foreign subsidiaries' sales of goods back to the U.S.) apart from horizontal investment (subsidiaries' sales to the foreign local market). 4 The last and biggest hurdle in identifying a potential effect of MNE activity on trade policy is the obvious potential endogeneity of export platform investment. A candidate FDI host country becomes a more attractive venue for offshoring operations if it has preferential market access to the anticipated market for its exports. Mexico's Maquiladora program, the well known North American Free Trade Agreement (NAFTA) predecessor, is an obvious example: duty-free access to the U.S. market was precisely the carrot offered to entice investors to set up export-oriented manufacturing bases south of the border. Indeed, as the theory laid out formally in Blanchard (2007) makes quite clear, export-oriented FDI will in general increase as tariffs to the export-destination market are lowered. Our response is to use an instrumental variables (IV) approach to control for the endogeneity of export-oriented FDI. The theory makes it clear that horizontal investment should not itself influence or be influenced by U.S. tariffs. In practice, however, we expect that market-seeking foreign investment is likely to be positively correlated with exportoriented investment, since both rely on a favorable climate for investment (comparative 2 In a general equilibrium model, an increase in the investment source country's tariff will cause the world relative price of the foreign import good (and thus the return to horizontal FDI) to rise. 3 Recall the prominent example of 'tariff jumping' Japanese car manufacturing plant construction in the U.S. in response to the voluntary export restraints in the 1980s, studied by Bhagwati, Brecher, Dinopoulos, and Srinivasan (1987) among others. 4 The BEA data include a third category: MNE sales to the rest of the world (ROW). Although these sales are clearly also export platform, they would not benefit directly from improved access to the U.S. market; we use the information in these 'other' sales to the ROW in falsification exercises. EMILY BLANCHARD AND XENIA MATSCHKE advantage via natural resources, policy, or otherwise). Thus, we use (horizontal) MNE sales to the local market as an instrument for (export-oriented) MNE goods sales to the U.S. Crucially, we construct our instrument using local sales by only those multinational affiliates that do not also sell goods back to the U.S., which addresses the potential concern that internal returns of scale at the affiliate level could challenge the exclusion restriction. 5 For this paper, we assemble a three-way panel data set including 80 industries, 184 countries, and ten years (1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006) to answer the question of whether export-oriented operations by U.S. MNEs cause higher rates of trade preferences for imports originating from countries where U.S. firms have set up shop. Our findings are consistent with the presence of such a causal relationship: conservatively, a 10% increase in U.S. foreign affiliate exports to the U.S. is associated with an increase in the duty-free access rate of about 4 percentage points (corresponding to about 20% of the mean value of preferential market access), controlling for the endogeneity of MNE foreign affiliates' exports to the U.S. Among potentially GSP-eligible developing countries, the estimated effect for GSP preferences is more than three times higher: a 10% increase in affiliate goods exports to the U.S. is associated with a 14 percentage point increase in the rate of duty-free access under the Generalized System of Preferences (more than seventy percent of the mean level of GSP access among potentially eligible countries). In a simultaneous equations version of the model, we find evidence consistent with the theoretical prediction that causality does indeed run both ways: expanded market access seems to increases MNE sales to the U.S. The baseline estimates prove to be remarkably robust in a variety of different empirical specifications. Our empirical results offer compelling evidence that offshoring MNE activity spurs preferential trade liberalization to the MNE's home country, which in turn may deepen economic integration between the investment host and investment source countries even further. To the extent that more generous preferential tariff treatment fosters additional 5 Intuitively, we would worry that U.S. trade policy could influence an affiliate's local sales if the decision to enter (or expand in) the market depends on both the level of exports back to the U.S. and local sales, as could arise under increasing returns to scale. U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 5 export-oriented investment, the cycle of improved market access and increased FDI may continue. At the same time, however, it stands to fear that the same mechanism can lead to substantial trade and investment diversion; a particular concern is that just as some trading partners experience ever-greater economic integration through this investmenttrade nexus, other countries may be left out entirely. The remainder of the paper proceeds in the usual sequence. In the next section, we briefly relate this paper to earlier work. Section 3 then reviews the theoretical motivation that guides our empirical approach and discusses the political process through which U.S. trade policy solicits and responds to multinational firms' concerns. Section 4 outlines our empirical strategy; Section 5 describes the data; and Section 6 presents the results. We describe a series of extensions and robustness tests in Sections 7 and 8 before concluding in Section 9. Related Literature Our study complements and considerably extends the empirical literature on the determinants of preferential treatment and the influence of international investment. The literature on trading blocs is sufficiently broad that discussion is restricted here to the small subset of work most closely related to this paper. 6 Most relevant to this paper's objective are Magee 6 See Bhagwati, Krishna, and Panagariya (1999) for a broader review. A second subset of related work consists of a handful of studies that examine the role of (exogenous) PTAs in determining investment flows. The seminal theoretical work by U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 7 Finally, this project provides an important complement to a companion paper, Blanchard and Matschke (2006). There, we use United Nations (UN) data on aggregate bilateral investment positions to predict the future formation of free trade agreements, in the spirit of Magee Theoretical Motivation and Political Mechanisms In this section, we review the theory that motivates and guides our empirical analysis and offer a short overview of how multinational firms can influence trade preferences in practice. Because the focus of this paper is empirical, we review the existing theory briefly and in broad terms; the interested reader is referred directly to the cited papers for additional detail. In a flexible general equilibrium framework, EMILY BLANCHARD AND XENIA MATSCHKE offshoring investments by multinational firms), the terms of trade externality is internalized, causing the optimal tariff to decline in the investment source country. 9 If trade protection is politically motivated and offshore investors (or multinational firms) actively lobby their local government, we would expect tariffs to fall even further. In a many country world, and in the presence of the GATT/WTO most favored nation clause, Blanchard Our first and most important empirical prediction thus follows directly from the theory: ceteris paribus, offshoring activity by firms from a given source country -here the United States -in a particular host country and industry will increase the incentives for U.S. policy makers to improve market access for imports from U.S.-affiliated producers overseas. 10 Because trade policy by law cannot discriminate at the firm level (with the notable exception of anti-dumping cases), we expect a greater rate of trade preferences to be afforded at the (HS-8) tariff line level for the country in question. According to the theory, the degree to which trade preferences will respond to multinational firms' offshoring operations depends on four key factors: the sensitivity of MNE investors' profits to U.S. trade policy; the political influence of the multinational affiliates relative to potential import-competing domestic groups; the opportunity cost of lost tariff revenues; and the extent to which preferences offered to one country and industry dilute the multinational affiliate profits derived from imports from the rest of the world. 11 Translating to empirically testable predictions in the context of U.S. trade preferences and multinational affiliate activity, we draw the following predictions: 9 While offshoring operations are only one form of export-oriented foreign investment, they are also the most easily measurable, which is why we use them as the focus of study in this paper. 10 See Lemma 3.5 in Blanchard 11 The first three predictions can be found in both Blanchard U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 9 Empirical Prediction 1. U.S. trade preferences for a given product j from country c will be: (i) increasing in U.S. multinational affiliates' sales of product j from country c back to the U.S.; (ii) decreasing in U.S. domestic sales of product j; (iii) decreasing in total U.S. imports of product j from country c; (iv) decreasing in total U.S. imports of product j from multinational affiliates located in the ROW. These predictions are largely self-evident, but a few points of explanation are worth noting. First, predictions (i) and (iv) use MNE affiliate sales of goods back to the U.S. as the relevant measure of the sensitivity of U.S. multinational investors' profits to U.S. trade policy; our choice of this measure rests on the assumption that an increase in preferential market access will translate to higher local per-unit revenues, and thus higher profits, earned by those multinational affiliates supplying the U.S. 12 In prediction (ii), we use the size of domestic U.S. sales as a proxy for potential protectionist pressure, with greater domestic sales associated with a more protectionist stance. In our empirical specifications, we also include a quadratic term for U.S. domestic sales, to capture potential concavity in the political response, loosely in the spirit of Bombardini and Trebbi (2011). 13 Prediction (iii) reflects the opportunity cost of tariff preferences, which is greater the larger the volume of imports that would no longer pay MFN duties. Prediction (iv) captures the possibility that enhanced market access for one country could, via trade diversion, cannibalize profits from other MNE affiliate suppliers to the U.S. located elsewhere in the world. 12 By Hotelling's lemma, recall that the derivative of the profit function with respect to selling price p is supply ( ∂π(p, w) ∂p = y(p, w)). Thus, if foreign investors are the residual claimants of affiliate profits and view local wages as fixed, then our sales measure is exactly the derivative theory suggests; otherwise the relationship is positive, but not exact. 13 The mechanism could be quite different, however, as their focus is the relationship between industry contributions (not the political outcome) and industry size. EMILY BLANCHARD AND XENIA MATSCHKE In addition, we expect that both MFN tariffs and total U.S. industry imports could play an important role, but the sign of the predictions is ambiguous. A higher MFN tariff implies a greater opportunity cost of offering tariff preferences (which would reduce equilibrium preferences), but at the same time, greater MFN tariffs may induce more active MNE political lobbying for discretionary policies (which would work in the opposite direction). 14 The net influence of the level of total U.S. imports is similarly unclear: else equal, greater total U.S. imports may imply less domestic political pressure against preferences (countering domestic sales and thus increasing the equilibrium use of preferences). Conversely, the higher the level of total U.S. imports, the greater the potential loss in U.S. tariff revenue from the rest of the world as the result of trade diversion (lowering predicted equilibrium preferences). Finally, the theory issues a clear warning about the endogeneity of export-oriented investment (or multinational affiliate operations), while also offering a potential identification strategy. First, it is immediately obvious that countries with preferential market access to the U.S. will be more attractive to potential MNE investors planning to sell goods back to the U.S., all else equal. Blanchard (2007) models this channel explicitly, demonstrating that trade liberalization and export-platform investment are complementary. The theory is equally clear, however, that market-seeking (horizontal) investment should have no such trade liberalizing effect; indeed, in a general equilibrium framework, market-seeking investment that competes for resources with the host-country's export sector could cause optimal tariffs in the investment source country to rise. To the extent that a many-sector economy more closely resembles a partial equilibrium framework, there would be no relationship between improved market access to the U.S. and horizontal investment. At the same time, however, we might expect vertical and horizontal investment to be positively correlated to the extent that both respond to the local investment climate. Absent better measures of the local investment conditions at 14 Anecdotally, MNEs do seem more politically active in practice when there is more on the line. We offer some evidence below. The bottom line is that the institutional structure governing U.S. trade policy leaves scope for discretionary application of trade preferences on a year-to-year basis, even within countries and industries. We document the extent to which we observe such variation empirically when we turn to the data in Section 5. Our last point about the existing institutional structure is that we do in fact observe multinational firms publicly and actively engaged in the political process that determines the structure of U.S. trade preferences. Below, we offer a handful of examples publicly available from government documents via the U.S. Trade Representative (USTR). All of the information below was collected independently from public petitions to the USTR and bears no relationship to the BEA data. In With a careful understanding of both the theoretical motivation and the institutional mechanisms that drive trade preferences in practice, we now turn to the empirical analysis. Empirical Strategy As in both 25 Under U.S. trade law, preferences must be binary -either eligible for duty-free access or subject to the MFN rate -so that in principle the observed tariff exemption θ cjt is either 0 or 1. Although we cannot observe θ * cjt directly, we can observe many of its determinants. We define the econometric model: where F DI cjt is a measure of offshoring activity by U.S. multinationals, X cjt is a k × 1 vector of other explanatory country-year, product-year, and country-product-year varying characteristics,α 0 andα 1 are scalar parameters, andβ is a 1×k vector of parameters. The parametersγ c ,γ j , andγ t stand for country-, product-, and year-fixed effects, respectively. The remaining error term,˜ cjt , represents unobserved heterogeneity in each country-product-year observation and is assumed to be independent of both X cjt and Assuming that˜ cjt is normally distributed, the model in (4.1) should be estimated using Probit. The theory predicts that among otherwise identical country-product-year pairings, U.S. trading partners with more offshoring activity by U.S. multinationals should have an increased likelihood of receiving preferential market access in those sectors. Thus, the key theoretical prediction is thatα 1 > 0. There are, however, several practical complications that call for modifications to our estimation strategy. First, several of our variables are not available at the product level j, but only at the industry level i, where each industry produces several products. In particular, while preferential market access is determined at the 8-digit HTS (tariffline) level, our information on U.S. multinational affiliates is coded only to the more aggregated 4-digit NAICS level. For this reason, the dependent latent variable is recast as the average preference share for all imports within a given 4-digit NAICS category: For both of these reasons -aggregation and part-year program eligibility -the dependent variable in our data must be treated as continuous. Our econometric model thus becomes: where i stands for a 4-digit NAICS industry and t stands for calendar year. We include in the vector X cit both the variables directly suggested by the theory, as well as a small set of standard political economy and gravity control variables. As discussed earlier, the key explanatory variables in X cit include total U.S. domestic sales in industry i and year t and its square; 27 total U.S. imports from the world in industry i and year t; total industry i exports to the U.S. from country c in year t; U.S. MNE sales from the ROW (other than country c) to the U.S. industry i and year t; and the U.S. MFN ad-valorem tariff rate in industry i and year t. Additionally, we include a set of industry-year controls for U.S. domestic political pressure: U.S. domestic payroll, number of establishments, import penetration, number of employees, and the year-onyear log change in U.S. employment and import penetration. Finally, we add two gravity variables: GDP per capita and population, which vary at the country-year level. 28 (Our results are robust whether or not we include these additional political economy or gravity variables; we find them to be sensible additions as they are both intuitive and prove to be statistically significant.) With a normally distributed error, the correct specification is a double-censored Tobit model. In practice, however, the three dimensions of fixed effects specified in (4.3) introduce a tradeoff between computational feasibility and adherence to this ex-ante preferred non-linear (Tobit) model specification. Thus, we pursue two different strategies for implementing our empirical test. In the first, we estimate a linear probability model in which we can remove industry-and country-fixed effects γ i and γ c by demeaning the data -but of course we must then ignore empirically the censoring process that generates the mass points for θ cit at 0 and 1. 29 In a second version of the empirical 27 Again, including the quadratic term to capture concavity. 28 In a pooled sample that we use in a robustness check, we also include country-specific, timeinvariant characteristics for distance from the U.S. and indicator variables for whether the country was a communist or terrorist state during our ten-year sample period, as well as time invariant industry-level dummies for agriculture and textile sectors. 29 Year-fixed effects are still removed by dummy variables. Because our data set is not balanced, we take care to demean the year dummies. We also correct the standard errors to account for the U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 17 strategy, we adopt the non-linear double censored Tobit model with country-, industry-, and year-fixed effects included as dummy variables, but the need to achieve convergence limits both our choice of estimator and the set of control variables we can include. 30 We address the potential simultaneity between offshoring activity and preferential access by instrumenting for our measure of multinational affiliates sales of goods back to the United States. Finding suitable excluded instruments that predict U.S. imports from offshore MNE affiliates, but are at the same time uncorrelated with the error term, is in general quite challenging. Fortunately, the BEA data separate MNE sales data by both type (goods vs. services) and destination (local, to the U.S., or to the rest of the world (ROW) ('other' in BEA parlance)). 31 Guided by theory, we argue that market-seeking (horizontal) MNE sales to the local market should be independent of U.S. trade preferences. At the same time, it seems likely -and we confirm in the data -that local and offshoring MNE sales are positively correlated, presumably because both capture the attractiveness of the local market for foreign investors. 32 Thus, we instrument for (export-oriented) sales of goods to the U.S. with (market-seeking) MNE affiliate sales (of goods and services) to the local market. We are careful to construct our instrument using only those affiliates that do not also sell goods to the U.S., to address the potential concern that increasing returns to scale at the affiliate level could cause a firm's joint decision to enter and sell to the U.S. and local markets to depend directly demeaning (to account for the difference in degrees of freedom); given the large number of observations, the standard error correction is negligible, however. 30 To achieve convergence in the instrumented Tobit specification, we adopt the Newey (1987) efficient two-step estimation and reduce the set of controls to only those variables explicitly indicated by the theory. 31 The BEA sales data are additionally categorized into sales to affiliated vs. unaffiliated buyers, a distinction we use in robustness tests reported later in the paper. 32 One might reasonably raise the concern that MNE sales to the local market could be a substitute for MNE sales to the U.S., and thus be (negatively) correlated with the second-stage error term. The first-stage results suggest a strong positive correlation between sales to each destination, however, which argues against the substitutes story. Data Preferential Market Access. The dependent variable of interest is preferential market access by industry, country, and year to the U.S. market. There are two ways to construct the preferential access variable using slightly different data sources; we consider both definitions to evaluate the robustness of our findings, and thus to ensure that our results rest on meaningful economic factors rather than specific variable definitions. Our first data source for preferential market access comes from the U.S. Trade Representative harmonized tariff schedules (HTS-US). Imposed tariff rates and the relevant indicators for preferential program eligibility are reported at the 8-digit HTS level by 33 Each 4-digit NAICS industry category allows a host of firms selling a range of products within each industry. In general, the firms and products selling to the local market are distinct from those selling to the U.S. On average, only about a third of firms in our sample sell both locally and to the U.S.; accordingly, 62% of all local sales are made by affiliates that do not also sell goods to the U.S. 34 To preserve zeroes, we add one to (all) variables' values before taking the natural log (thus we replace x with ln(1 + x)); the usual caveats apply. See, e.g., Wooldridge (2009), page 192. 35 Indeed, an eight-month lag is institutionalized in the GSP annual review process -decisions made in (typically) autumn of one year are not implemented until July 1st of the subsequent calendar year. U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 19 country, industry, and year. 36 Thus, one way to define the preferential treatment variable is so that θ cjt = 1 if the country-product pair is eligible for a special rate code in year t, and 0 otherwise. 37 When we aggregate to the 4-digit NAICS level (necessary to concord the preferential tariff data to the investment data), we construct both historic trade weighted 38 and straight (unweighted) averages across the relevant subcategories. We view the trade-weighted version as the more appropriate measure, as it captures the empty promise of preferences for goods that are not produced by a beneficiary country (mangoes from Iceland, semiconductors from Afghanistan, etc.). In the tables, we label the trade weighted eligibility measure of duty-free access under any preference program El Any. We report robustness results for the unweighted eligibility measures in the footnotes of Section 8; the results are qualitatively unchanged by removing the trade weights. While this simple eligibility-based definition of the preferential treatment variable is appealing in its parsimony, one can easily challenge the definition on the grounds that even when preferential eligibility is indicated by the HTS-US, preferential treatment is often afforded to only a subset of the imports in question. Partial-year program eligibility is a key concern, as many program changes (including virtually all changes under GSP) are effective July 1st (or idiosyncratically), rather than January 1st of a calendar year. Moreover, GSP preferences can be (and often are) limited by additional "competitive need limitations" (CNLs), which offer duty-free treatment only until a certain level of exports is reached. Finally, restrictive rules of origin restrictions or other bureaucratic costs under some programs may make de jure preference eligibility useless in practice. A preference measure based on countries' actual usage of the programs can capture such otherwise unobserved limitations to program use in practice. 36 As of January 1, 2006 (the last year in our sample), the HTS-US schedule included 19 preferential treatment codes for GSP (and its subcategories), regional agreements, etc. 37 We code a country-product-year observation as preference eligible if it is eligible for more than one quarter of the given calendar year. 38 Time invariant trade weights are constructed using 1997 trade flows, the year immediately preceding the first year in our sample. EMILY BLANCHARD AND XENIA MATSCHKE With these caveats in mind, we define a second form of the dependent variable using more detailed data from the U.S. International Trade Commission (USITC). Each year, the USITC reports the proportion of bilateral trade that clears U.S. Customs under each preferential program code, by industry and country of origin. 39 We use this information to construct our baseline measure of the dependent variable, Any Pref Share, so that θ cjt is the (exact) share of country c exports of product j in year t that entered U.S. Customs claiming duty-free access under any preferential program code. Note that this version of the dependent variable based on U.S. Customs data offers the additional advantage that it does not require an ad-hoc weighting scheme to aggregate to 4-digit NAICS. 40 In addition to these two measures of U.S. preferential market access under any and all preference programs, we create three alternative dependent variables: two measures of duty-free market access under GSP only -one eligibility based, El GSP, another based on Customs data and actual GSP use, GSP Share, and a third alternative for a robustness check, Non-GSP Share, the rate at which a country's exports claim any preferential market access under a program other than GSP. There are several reasons that we find 39 For instance, when a product enters the U.S. under a GSP eligibility clause, it receives special tariff code A (or A*, A+ depending on the particular sub-classification of GSP eligibility). When a product enters the U.S. duty-free under a free trade agreement, then an agreement-specific code is entered (for instance "MX" for Mexican products entering under the NAFTA or "R" for products entering under the Caribbean Basin Trade Partnership). These codes match those used by the USTR. 40 It should be noted that some of the deviation between the Customs-use and eligibility-based measures is caused by foreign exporters failing to (or choosing not to) claim preferential access when eligible. (See recent work by Hakobyan (2010) on underutilization of GSP preferences.) The relevant concern for our study would be if MNE foreign affiliates are more able to use preferential trade programs than are commensurate locally owned firms (which, to be fair, is not obvious -while U.S. MNE affiliates may find it easier to file the paperwork necessary to get preferential market access, their disproportionate use of sourced inputs may make rules of origin harder to satisfy). The robustness of our results to either preference variable definition -program use, which could potentially be contaminated by differential uptake of preference program usage by MNE affiliates (but is otherwise a more precise measure), and eligibility, which is clearly free of any such concern (but is also a less perfect measure given aggregation and related issues outlined earlier) -suggests that this discrepancy is of minimal consequence in the context of our study. U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 21 GSP preferences of particular interest. The first is simply that the GSP program comprises roughly two-thirds of country-industry-year observations with preferential market access in the U.S. 41 At the same time, the mechanics of the GSP program feature important institutional differences compared to regional or other preferential programs. As mentioned earlier, there is a formal process of annual reviews of GSP eligibility in which domestic and foreign firms, foreign governments, labor unions, and other interest groups can, and regularly do, take part. Year to year, GSP eligibility can be revoked on a discretionary product level basis due to human rights, labor, or intellectual property violations, and is regularly limited by binding competitive need limitations that cap imports from the most productive developing countries. Finally, from an econometric perspective, the GSP-based preference measure is largely immune from concerns over reciprocal trade policy implications or bilateral investment protections that may arise under Article XXIV free trade agreements. It is worth spending a moment to examine the extent of variation in our dependent variable, whichever way it is defined. Per the earlier discussion, a literal reading of Article XXIV rules would lead one to expect that country-fixed effects would explain most of the observed variation in trade preferences under regional or bilateral free trade deals. The Enabling Clause, which authorizes GSP programs, allows more discretion at the industry level -certain industries may be excluded from GSP eligibility entirelybut these industry exclusions must be uniform across GSP beneficiary countries (hence the moniker, Generalized System of Preferences). 42 In principle, we then would expect that together, country-and industry-fixed effects would account for virtually all of the observed variation in the GSP variables. Of course, it is well understood that there is some leeway in incorporating exemptions and exclusions both during the initial negotiation of preferential agreements and through 41 Weighting by trade volume, however, GSP is predictably a smaller share; total U.S. imports under GSP comprise 7% of all preferential imports. 42 The U.S. GSP includes a two-tiered system to allow enhanced market access for "least developed beneficiary countries," but as will be clear from both the following figures and our empirical results, the two-branch system is not responsible for our results. 22 EMILY BLANCHARD AND XENIA MATSCHKE subsequent formal and informal review processes thereafter. To convince the reader that there is, in fact, sufficient variation left to explain after including country-, industry-, and year-fixed effects, we offer the following simple plots. On the vertical axis, we plot the residual of our dependent variable, preferential market access after controlling for country-, industry-, and year-fixed effects. On the horizontal axis, we plot country per capita GDP. Each plot is for a different definition of the dependent variable: Any Pref Share and GSP Share in the top row, and El Any and El GSP in the bottom row. For the two GSP-specific measures, we include data only for the set of countries that is potentially GSP eligible. In all four plots, we see quite clearly that there is considerable variation in every measure of preferential market access and that the degree of variation is higher among the U.S.' less developed (low GDP per capita) trading partners. 46 For our instruments, we use the two additional measures from the BEA data discussed earlier: total MNE affiliate sales of goods and services to the local market by only those firms that do not also sell goods to the U.S., and MNE affiliate sales of services (worldwide), again by only those firms that do not simultaneously sell goods to the U.S. Control Variables. Finally, we include the control variables at the country, industry, country-year, industry-year, and country-industry-year levels as discussed in the previous section. When we include country-and industry-fixed effects, of course, the time-invariant country and industry level controls (such as distance to the U.S. and indicator variables for textile or agricultural industries) are dropped. In principle, our data set would have 204,160 observations: 232 countries by 88 industries for 10 years (1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006). In practice, however, our sample is smaller. We have data for a subset of 184 countries and 80 industries over the ten-year sample. When we lag the independent variables for one year (as seems most appropriate given the time needed for policy to change) and include variables for the log change in U.S. employment or import penetration, our data are further reduced to a 9-year panel beginning in 1998. 47 Of these remaining data, another limitation arises. When we report the preference variable as the share of imports entering the U.S. with preferential treatment, we lose all observations for which U.S. imports (the denominator of the ratio) are zero, leaving 68,130 observations. To make our baseline results consistent across specifications, we use this smallest data set for all specifications of the model, even in cases where we have access to more observations (i.e. for the eligibility-based definitions of the preference variables, which need not preclude observations for which the U.S. import volume was zero). 47 While some preference programs are reviewed only on an ad hoc basis, others, like the GSP, have a formal annual review process for petition, study (by the USITC), and ultimate implementation by USTR on a regular yearly schedule. U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 25 6. Results establishments, and country population. Of these, only the number of U.S. establishments is significant at the 10% level in any of the specifications, with an economically modest coefficient estimate (.017 * * (.009) in specification (1)). 50 From 51 Our coefficient estimates imply that the derivative of Any Pref Share with respect to U.S. domestic sales is negative at the sample minimum of domestic sales, but positive at the sample maximum, suggesting a U-shaped relationship. EMILY BLANCHARD AND XENIA MATSCHKE preferential market access, which may reflect a higher opportunity cost of offering dutyfree access to these sectors and countries. We find that preferences are also less generous for industries with higher MNE affiliate sales from the rest of the world (which may compete with the exports in question), which is once again consistent with our initial empirical predictions. The theory did not offer clear priors for the sign of the remaining coefficients, but we venture that by at least a rough approximation, all of the estimates make sense. Our estimated coefficients are positive and negative for MFN tariffs and total U.S. imports, respectively. Tying back to the discussion in Section 3, the first result suggests that higher MFN tariffs may indeed shift political attention towards the use of preferences, while the second is consistent with the concern that increasing preferences for one country may reduce overall tariff revenue from the rest of the world via trade diversion. Continuing down the table, we find that countries with lower GDP per capita receive ceteris paribus a greater rate of preferential market access, presumably due to the large role of GSP preferences in the full sample. Finally, as one might expect, healthier U.S. industries with larger payrolls and growing employment tend to offer more generous preferential access (or put another way, senescent industries receive more protection); at the same time, bigger industries as measured solely by number of employees (holding payroll fixed) receive more protection in the form of a lower trade preference share. Column 3 presents our baseline specification, in which we reestimate the overidentified version of the model under the more conservative assumption allowing for correlation of the error term within countries. Clustering by country, the standard error of the coefficient on MNE goods sales to the U.S. increases slightly relative to the unclustered version, but the coefficient remains statistically significant at the 5% level with a p-value of .045. The remaining columns offer three robustness checks. Column 4 tests the robustness of our findings to changing the set of control variables. We alter the estimation by reducing the set of controls to include only those variables that are directly indicated by the theory. The estimate of the effect of a 10% increase in multinational sales to the U.S. drops only slightly to 3.6 percentage points and is significant at the 10% level, even when clustering by country. The remaining variables still included in the regression remain virtually unchanged from the baseline version in column 3, while the first-stage Finally, columns 5 and 6 push the data to the limit, adopting increasingly strict assumptions about the structure of the error term. Clustering by country-industry pair in column 6 reduces the statistical significance to the 10% level, but only clustering by country and industry (column 5) finally pushes the p-value above 10% to .19. It is worth noting that the first stage estimation results remain sharp under even the most restrictive assumptions about the error term: the p-values for significance of the instruments in the first stage remain below .001 and .03 (for local and service sales respectively), even with two-way clustering in the presence of three-way fixed effects. We now turn in 54 In the reduced sample, 91% of observations are for currently GSP eligible countries, opposed to 57% in the full sample. The first column presents results for the Any Pref Share variable, for comparison with the benchmark specification in the full sample The effect of removing the de jure GSP ineligible countries is striking: the coefficient estimates for the impact of MNE sales on all U.S. trade preferences U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 29 estimates are consistent with our results in the full sample with two notable exceptions: perhaps unsurprisingly, per capita GDP is no longer a statistically significant predictor of trade preferences when we restrict attention to developing countries. Likewise, the effect of MNE affiliate sales from the ROW is muted among developing countries, as one might expect to the extent that the trade diversion induced by a developing country is less likely to impose a competitive disadvantage on affiliate imports from the ROW. Columns 2-5 report the results when the dependent variable is redefined as the share of U.S. imports that enter duty-free under only the GSP program. We again report the just-identified results in the first specification (column 2). The benchmark specification (over-identified, clustered by country) for GSP preferences is in column 3, and the clustering variants (with two-way clustering and by country-industry pair) are in columns 4 and 5, respectively. All specifications of the model yield a coefficient estimate for MNE goods sales to the U.S. that is statistically different from zero at at least the 5% level. A priori, one might have anticipated that trade preferences for poorer countries would be less subject to domestic U.S. influences such as the change in the number of employees or U.S. domestic sales, but we find no support for such a view. Indeed, in the benchmark specification of the GSP model in column 3, the estimated impact of a 10% increase in the level of MNE affiliate exports back to the U.S. is nearly a 15 percentage point increase in the rate of GSP market access among potentially GSP eligible developing countries -roughly a 75% increase over the mean and just slightly higher than the estimated effect of multinational affiliate sales on preferences overall. The remaining coefficient estimates are consistent with the results in column 1 and expectations, so we see no need to discuss them further. The instruments are somewhat weaker with this restricted sample, presumably because there are fewer industries among the developing countries with both marketseeking (local sales) affiliates and offshoring MNE affiliates selling goods back to the 30 EMILY BLANCHARD AND XENIA MATSCHKE U.S. 55 At the same time, instrument validity appears of little concern given J-statistic pvalues of approximately 100% in these GSP-program specific specifications of the model. We now turn in 56 The first two columns of The results are qualitatively similar to the results of the IV panel estimation, in that we find a strong positive effect of MNE sales to the U.S. on preferences. For the Any Pref Share variable in column 1, the effect of MNE sales to the U.S. is more than triple in size compared to the IV panel specification, 57 which suggests that in the full sample, our baseline linear specification may understate the effect of offshoring by multinationals on U.S. trade preferences. The coefficient estimate is even higher if the dependent variable is the GSP share, which indicates that a 1% increase in MNE sales back to the United States is associated with a 2 percentage point increase in the rate of preferential access 55 Interestingly, among potentially GSP-eligible developing countries, a greater share of total local sales are by firms that do not also sell goods back to the U.S. -70%, compared to 62% in the full sample. It seems the line between vertical, offshoring MNEs and market-seeking, horizontal MNEs is even sharper in developing countries. 56 Unfortunately, Newey two-step estimation precludes clustering and limits the set of available post estimation statistics. 57 The coefficient on MNE Sales is 1.54 * * * in U.S. MULTINATIONALS AND PREFERENTIAL MARKET ACCESS 31 for countries and industries that already have at least some preferential access at the 4-digit NAICS level. Reducing the sample to include only the de jure GSP-eligible countries in columns 3 and 4 leads to roughly the same estimated effect of FDI on preferences as in the full sample, and the results are markedly closer to our linear model findings in Summarizing our results thus far, we draw three broad conclusions. First, the empirical results are qualitatively consistent with the model, and our instruments MNE