NBER WORKING PAPER SERIES PUTTING THE LID ON LOBBYING: TARIFF STRUCTURE AND LONG-TERM GROWTH WHEN PROTECTION IS FOR SALE. Nathan Nunn Daniel Trefler

Transcription

1 NBER WORKING PAPER SERIES PUTTING THE LID ON LOBBYING: TARIFF STRUCE AND LONG-TERM GROWTH WHEN PROTECTION IS FOR SALE Nathan Nunn Daniel Trefler Working Paper NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massahusetts Avenue Cambridge, MA April 2006 We thank members of the CIAR for valuable omments, as well as seminar partiipants at the CEA Annual Meetings, the University of British Columbia, the University of Toronto and the World Bank. The Soial Sienes and Humanities Researh Counil of Canada (SSHRC) and the CIAR kindly provided funding for this projet. The views expressed herein are those of the author(s) and do not neessarily reflet the views of the National Bureau of Eonomi Researh by Nathan Nunn and Daniel Trefler. All rights reserved. Short setions of text, not to exeed two paragraphs, may be quoted without expliit permission provided that full redit, inluding notie, is given to the soure.

2 Putting the Lid on Lobbying: Tariff Struture and Long-Term Growth when Protetion is for Sale Nathan Nunn and Daniel Trefler NBER Working Paper No April 2006 JEL No. F10, F14, O24, O40 ABSTRACT It has long been reognized that a ountry's tariffs are the endogenous outome of a rent-seeking game whose equilibrium reflets national institutions. Thus, the struture of tariffs aross industries provides insights into how institutions, as refleted in tariff poliies, affet long-term growth. We start with the ommonplae pereption among politiians that protetion of skill-intensive industries generates a growth-enhaning externality. Modifying the Grossman-Helpman protetion for sale model to allow for this, we make two preditions. First, a ountry with good institutions will tolerate high average tariffs provided tariffs are biased towards skill-intensive industries. Seond, there need not be any relationship between average tariffs and good institutions. Using data for 17 manufaturing industries in 59 ountries over approximately 25 years, we find that average tariffs are unorrelated with output growth and that the skill-bias of tariff struture is positively orrelated with output growth. We interpret this to mean that ountries grow faster if they are able and willing to put a lid on the rent-seeking behaviour of speial interest lobby groups. We show that our results are not ompatible with explanations that appeal to (1) externalities per se, (2) initial industrial struture that is skewed towards skill-intensive industries, or (3) the effets of broader institutions suh as rule of law and ontrol of orruption. Nathan Nunn Department of Eonomis University of British Columbia # East Mall Vanouver, BC V6T 1Z1 Canada nnunn@interhange.ub.a Daniel Trefler Rotman Shool of Management University of Toronto 105 St. George Street Toronto, Ontario M5S 3E6 ADA and NBER dtrefler@rotman.utoronto.a

3 1. Introdution It has long been reognized that a ountry s tariffs are the endogenous outome of a rent-seeking game whose equilibrium reflets national institutions e.g., Krueger (1974), Grossman and Helpman (1994) and the review by Gawande and Krishna (2003). Thus, the struture of tariffs aross industries an provide insights into how institutions, as refleted in tariff poliies, affet long-term growth. We investigate the relationship between a ountry s domesti institutions and its tariff struture using Grossman and Helpman s (1994; 1995; 2001) protetion for sale model. We then take the theoretial preditions to the data by estimating the industry level relationship between tariffs and output growth for a sample of 17 industries in 59 ountries over the last 25 years. Our empirial point of departure is a rih literature on the relationship between average tariffs and per apita GDP growth e.g., Edwards (1992, 1998), O Rourke (2000), Vamvakidis (2002), Yanikkaya (2003), Clemens and Williamson (2004), Jaks (2005) and DeJong and Ripoll (2006). 1 These papers make a number of ontributions, highlighting, for example, how the tariff-growth relationship has hanged over time and how it varies with the level of development. However, there are two very speifi senses in whih these papers are limited in their analysis of the impat of tariffs on growth. First, many of these papers argue that a positive oeffiient on average tariffs should be understood to mean that there is a positive externality. Yet a key result of Grossman and Helpman s (1990; 1991) analysis of tariffs, externalities, and endogenous growth is that what matters is not just the average tariff, but the struture of tariffs. For example, a ountry with high average tariffs will grow rapidly if tariffs are highest in researh-intensive industries and will grow slowly if tariffs are highest in unskilled-intensive industries. It is not the average tariff alone that matters, but the struture of tariffs. Seond, these papers ignore the endogeneity of tariffs e.g., Trefler (1993), Goldberg and Maggi (1999) and Gawande and Krishna (2005). If institutions determine tariff poliies then the tariff oeffiient in a 1 These are papers that deal expliitly with average tariffs. There are of ourse many other papers that deal with openness to international trade more generally. 1

4 growth regression may simply be piking up the familiar orrelation between institutions and growth e.g., La Porta, Lopez-de-Silanes, Shleifer, and Vishny (1997), Hall and Jones (1999) and Aemoglu, Johnson, and Robinson (2001, 2002). To inorporate tariff struture and tariff endogeneity into the existing empirial literature, one needs a theoretial model that allows for multiple industries, externalities that vary aross industries, and tariff endogeneity. This turns out to be possible with just a minor modifiation of the Grossman and Helpman (1994) protetion for sale model. In partiular, we introdue externalities into their model. The model yields two preditions about institutions and tariffs. In the model, institutions are defined narrowly in terms of the game played out between industry lobbyists and the government. (See Greif (2006) for a tighter definition of institutions.) In ountries with good institutions, poliy makers plae a heavy weight on both onsumer surplus and future growth while plaing little weight on the politial ontributions of industry lobby groups. In our model, tariffs affet future growth via externalities and these externalities vary aross industries. Our first predition is obvious: there need not be any relationship between good institutions and the average level of tariffs. While onsumer surplus is redued by high average tariffs, future growth is enhaned by high tariffs in industries that generate positive externalities. Our seond theoretial result is more subtle and based on two empirial observations. (1) In almost all ountries, unskilled-intensive industries have higher tariffs than skillintensive industries. (We show this below for our sample as well.) Countries that heavily weight onsumer surplus will want to redue prie distortions within the tradeables setor and this requires a redution in the tariffs of unskilled-intensive industries. (2) Skill-intensive industries generate positive externalities. 2 Thus, ountries whose poliy makers heavily weight future growth will want high tariffs in skill-intensive industries. Putting these two observations together, the ratio of tariffs in skill-intensive industries to 2 It is not important here whether this is true or whether it is pereived to be true by poliy makers. It is also not important whether skill-intensive industries generate a positive externality or whether unskilledintensive industries generate a negative externality. See Antweiler and Trefler (2002) for supporting general equilibrium evidene of the existene of positive externalities in skill-intense industries. 2

5 tariffs in unskilled-intensive industries will be highest in ountries with good institutions. More suintly, ountries with good institutions will have a skill-biased tariff struture. Summarizing the preditions of our amended protetion for sale model, average tariffs are not informative about whether a ountry has good institutions. In ontrast, the skillbias of the tariff struture is indiative of good institutions. We use these insights to struture an empirial investigation of the impat of tariffs on long-term industry output growth. We provide a new panel database on tariffs and output for 17 manufaturing industries in 59 ountries. The panel ends in 2000 and ountries enter as early as We regress long-term industry output growth on own-industry tariffs as well as on two ountry level tariff measures, the average tariff and the skill-bias of the tariff struture. We find that average tariffs are unorrelated with long-term industry output growth: ountries with lower average tariffs do not grow faster. In ontrast, the skill-bias of the tariff struture is highly signifiant both eonomially and statistially: ountries with relatively higher tariffs on skill-intensive industries grow faster. These results also hold in a ountry level analysis of the impat of tariffs on long-term per apita GDP growth. There are several possible interpretations of this result beyond what our model provides. A large part of this paper is devoted to sorting out these possibilities whih explains why a formal model is useful and to disriminating empirially between alternative interpretations. There are three groups of alternative interpretations. The first is that we are apturing the broader effets of institutions on growth that have already been doumented by La Porta et al. (1997), Hall and Jones (1999), Aemoglu et al. (2001, 2002) and others. This is not the ase. For one, by inluding nine regional dummies we are explaining sample variation within narrowly defined regions. In ontrast, other studies inlude few or no regional dummies and are largely explaining global sample variation. For another, we inlude in a single regression all six World Bank measures of institutions (Kaufmann, Kraay, and Mastruzzi, 2003). The skill-bias of tariff struture is more powerful than any one of these institutional measures and is almost as powerful as 3

6 the joint effet of all six. We interpret this to mean that we are apturing the effet of the narrow lobbying institution desribed by Grossman and Helpman s protetion for sale model. The seond alternative interpretation of our results is that we are apturing externalities alone. Tariffs in skill-intensive industries promote externalities and this in turn promotes growth. However, unless externalities are muh larger than typially estimated, the pure externalities interpretation requires there to have been a massive expansion of skill-intensive industries. We find no evidene of this. Further, the externalities would have had to operate at the eonomy-wide level rather than the industry level and we find no eonomy-wide effets suh as indued human apital aumulation. The third alternative interpretation is that ountries with a skill-biased tariff struture initially had a large skill-intensive setor. Under this interpretation, all we are finding is that initial prodution struture matters for future growth. However, we ontrol for initial prodution struture in all of our speifiations. In summary, average tariffs are unorrelated with both industry level output growth and ountry level per apita GDP growth. In ontrast, the skill-bias of the tariff struture is an important orrelate of long-term growth. We explain this fat using the protetion for sale model modified to allow for externalities that affet future growth. Rapid growth ours in ountries whose governments are both willing and able to put a lid on the rent-seeking behaviour of speial interest lobby groups. Setions 2 to 4 lay out the model. Setions 5 to 7 present the ore empirial work. 3 The remaining setions examine alternative explanations of our results and establish robustness. 2. Protetion for Sale We develop a minor modifiation of the Grossman and Helpman (1994) protetion for sale model that generates a rih set of preditions relating the quality of a ountry s institutions to its ross-industry distribution of tariffs. The minor modifiation is an 3 Our ore empirial results appear in tables 6 and 7. The impatient reader an jump straight to these. 4

7 inter-setoral externality. Sine almost all of the set-up draws diretly from Grossman and Helpman (1994), we proeed with a terse exposition. 4 A representative onsumer has separable utility over n + 1 goods. Goods i = 1,..., n are tradeable while good i = 0 is nontradeable. We take good 0 to be the numeraire. Let p i be the domesti prie of good i and let p = (p 1,...,p n ) be the vetor of domesti pries. One unit of good 0 output is produed with 1 unit of labor so that the wage rate is unity. Goods i > 0 are produed under onstant returns to sale using labor and a speifi fator. As a result, output q i depends only on p i. The onsumer resides in a small open eonomy faing exogenous world pries. Government poliies put a perentage gap τ i = (p i pi )/p i between domesti pries and exogenous world pries pi. We take these poliies to be tariff poliies though a more general interpretation is possible. To redue notation we set pi = 1 in what follows. The model with pi unrestrited is relegated to the appendix. The existene of a speifi fator provides a motive for protetion. Owners of the speifi fator oalese into a lobby whose welfare is desribed by W i (p). Exatly as in Grossman and Helpman (1994), this aptures four soures of benefit: (1) rents that arue to the owners of the speifi fator used in industry i, (2) labor inome of lobby members, (3) a share α i of the eonomy-wide onsumer surplus, and (4) a share α i of the eonomy-wide tariff revenue. The lobby s share α i is just the fration of the voting population that owns some of the speifi fator of industry i. In return for protetion, the lobby gives the government a politial ontribution C i (p). W i and C i depend on pries in all setors beause there is a peuniary externality: pries in other setors affet items 3 and 4. We now introdue our only departure from the Grossman-Helpman set-up. We allow for the possibility that a tariff in one setor generates an eonomy-wide externality. 4 See also Grossman and Helpman (1995, 2001), Helpman (1997), Goldberg and Maggi (1999), Gawande and Bandyopadhyay (2000), Gawande and Krishna (2005) and Mitra, Thomakos, and Ulubaşoğlu (2006). A number of papers onsider extensions of the original protetion for sale model. Mitra (1999) and Magee (2002) allow for endogenous lobby formation, Bombardini (2005) allows for firm heterogeneity and Chang (2005) allows for monopolisti ompetition. 5

8 Consider just two examples. First, protetion of skill-intensive industries raises the demand for skills, thus leading to human apital aumulation and possibly to faster future growth. Conversely, protetion of unskilled-intensive industries lowers the returns to human apital aumulation and hene retards growth. Seond, skill-intensive industries tend to be industries with omplex prodution (Levhenko, 2004) and industries requiring relationship-speifi investments (Nunn, 2005). Beause of these features of the prodution proess, these industries are subjet to opportunisti behaviour. Poliies that promote the growth of these industries raise the demand for a good legal environment, thus leading to growth-enhaning institutional reforms. Along these lines, Aemoglu, Johnson, and Robinson (2005) argue that the rise of international trade in the Atlanti eonomies during the early modern period promoted a demand for institutional reforms that were growthenhaning. Puga and Trefler (2006) make a related argument in the ontext of medieval Venie and then show how rapid growth pushed the politial equilibrium towards a more protetionist regime that ultimately hoked off the Venetian eonomy. For our empirial work, we do not want to prejudge the externality. Thus, we allow the externality to be either positive or negative. To model the externality, let G(p) be the welfare (possibly negative) assoiated with the externalities generated by prie (or tariff) shedule p. We do not need to plae any restritions on G for the results related to our empirial work. However, it will simplify the mathematial expressions if we assume that a tariff in industry i affets the externality only via its affet on net exports x i (p i ). 5 More onretely, we replae G(p) with G(x(p)) where x(p) (x 1 (p 1 ),..., x n (p n )). Results without this assumption appear in Appendix A. Further, we an make an independent ontribution to the theoretial protetion for sale literature if in addition we assume separability: G(x(p)) = Σi=1 n G i(x i (p i )) for arbitrary differentiable funtions G i. Notie that the externality operates via G, but not via urrent output, whih is denoted by q i. This is not important for our theoretial model. We have hosen this route beause in the empirial work we separate out fators that affet urrent output from those that affet 5 The Grossman-Helpman assumptions ensure that demand and supply in setor i depend only on p i. Hene x i depends only on p i. 6

9 future output. Therefore, empirially we tend to think of G as the future welfare stream that arues from a tariff-indued, growth-promoting or growth-retarding externality. Of ourse, the theoretial interpretation of G is far more general. Let γ i G(p) be the externality-indued welfare stream that arues to lobby i where Σi=0 n γ i = 1. Then the objetive funtion for lobby i is W i (p) + γ i G(x(p)) C i (p). (1) A key modeling insight in Grossman and Helpman s original paper is to write the government s objetive funtion as aw(p) + C(p) for some onstant a > 0 where W(p) Σ n i=0 W i(p) and C(p) Σ n i=1 C i(p). a aptures the weight the government plaes on gross welfare relative to politial ontributions. We extend their approah by defining government welfare as aw(p) + bg(x(p)) + C(p) a, b > 0 (2) where b aptures the weight the government plaes on welfare from externalities relative to politial ontributions. 6 Equations (1) and (2) satisfy the Bernheim and Whinston (1986) assumptions. Assuming differentiability of all funtions, our equations thus also satisfy the Grossman- Helpman assumptions. We an therefore use the wonderfully simple Grossman-Helpman solution methods. We refer the reader to their paper for additional details, inluding the onept of subgame-perfet Nash equilibrium that we will be using. 3. Charaterization of the Equilibrium Tariff Struture We turn now to haraterizing the equilibrium tariff struture (τ 1,..., τ n ) = (p 1 1,..., p n 1). Let I i be a binary indiator taking on a value of 1 if the industry has an 6 All of our results hold when a = b. However, if one thinks of G as the future welfare stream that arues from the poliy-indued externality then b may be determined by a different set of fators than a. For example, there is a well known tendeny for governments to overly disount the impat of urrent deisions on future outomes. This phenomenon has been repeatedly doumented in the literature on politial yles. For example, Besley and Case (1995) find that lak of eletoral aountability due to term limits on U.S. state governors leads to fisal yles and slower state growth. This suggests that politiians disount the future beause of unertainty about re-eletion. Suh onsiderations do not affet a, but do affet b. 7

10 organized lobby and a value of 0 if the industry is not organized. Let L be the set of industries that are organized. Let α L Σ i L α i be the share of voters that own a speifi fator in an organized industry i L. Let γ L Σ i L γ i be the share of the externality that arues to these voters. Theorem 1 (Protetion for Sale with an Externality) The subgame-perfet Nash equilibrium of the trade-poliy game satisfies for i > 0 where x i x i/ p i > 0 and G i G/ x i. τ i = I i α L a + α L qi(p i ) x i (p i) + b + γ L a + α L G i (x i(p)). (3) The proof appears in Appendix A. The first term on the right-hand side of equation (3) is idential to Grossman and Helpman s ore proposition 2 (as expressed in their footnote 10). The seond term on the right-hand side of (3) is new. To understand its signifiane, without loss of generality hoose i so that industries are ordered by the G i : G 1 < < G n. (4) This ordering of industries means that a tariff in industry i + 1 is more growth-enhaning or less growth-retarding than a tariff in industry i. Growth here refers to a one-time inrease in welfare. With this hoie of industry ranking, the seond term in equation (3) is inreasing in i. That is, all else equal, equilibrium tariffs are highest in industries where the impat of tariffs is most growth-enhaning or least growth-retarding. Further, and this is entral to our argument, the more weight the government plaes on the externality (the larger is b), the more skewed is the tariff struture towards high-i setors. 4. Institutions and the Struture of Protetion The Grossman-Helpman game played by the government and n lobbies is part of what Greif (2006) defines as an institution. We fous on the parameters a and b, whih are the exogenous pay-off relevant part of Greif s definition. For brevity, we refer to a and b as 8

11 a ountry s institutions. It is natural to think of a and b as parameterizing the degree to whih national institutions put a lid on rent-seeking behaviour. There is abundant evidene that measures of the prevalene of rent-seeking, and poor institutions generally, are negatively orrelated with long-term growth. It is therefore appropriate to say that the larger are a and b, the better are institutions. 7 We start with a general statement about the effet of institutions on tariffs. Changes in a and b have two effets. There is a diret effet that operates through equation (3), holding q i /x i and G i fixed. This diret effet is trivial to alulate simply differentiate equation (3) with respet to a and b. We will sometimes say that the diret effet holds onstant the struture of prodution. There is also an indiret effet that aptures how hanges in a and b lead to hanges in τ i, whih lead to hanges q i /x i and G i, whih feed bak into additional hanges in τ i. This is a far more diffiult derivative to alulate. The next theorem deals with both the diret and indiret effets. The proof appears in Appendix B. Theorem 2 (Institutions and Tariff Struture) For i = 1,...,n: 1. The total (diret plus indiret) effet of institutions on tariffs is given by for some funtion φ i > 0. τ i dτ i = φ i (p i ) da + G i φ i (p i ) db 2. The diret effet (holding onstant the struture of prodution) of institutions on tariffs is given by τ i dτ i = da + G i db. a + α L a + α L 7 Returning to our interpretation of G as the future welfare stream that arues from a tariff-indued, growth-promoting or growth-retarding externality, it is useful to think about why b might vary aross ountries. Institutions suh as the onstitution may promote transpareny and aountability, thus foring the government to be more forward looking. For example, Persson and Tabellini (2003) find that ountries with onstitutionally mandated strong aountability via the ballot box have redued levels of both orruption and government ineffetiveness, have better protetion of property rights and, ultimately, have faster growth. Thus, one interpretation of b is that it is a measure of aountability via the ballot box. Persson and Tabellini (2003) also find that relative to parliamentary regimes, presidential regimes are assoiated with signifiantly worse eonomi performane due to worse strutural poliies and a legal system that is less respetful of property rights. Differenes in b aross ountries an thus be due to more subtle ross-ountry differenes in eletoral systems. Overall, we expet b to be highest in ountries with politial systems that display transpareny and aountability, and are thus able to find a better balane between the urrent interests of rent-seeking lobbyists and the future welfare of the soiety. 9

12 This means that better institutions as measured by a larger a, lead to a smaller value of τ i i.e., to less of a prie distortion. 8 9 Better institutions as measured by b lead to more omplex effets on tariffs that we will explain shortly. Theorem 2 will now be used to draw impliations for standard measures of the tariff struture. In our empirial setting we always ontrol for the struture of prodution. Therefore, we work with the diret effet only in what follows. A. Distortions Between Tradeables and Nontradeables: The Average Tariff We turn first to the textbook ineffiieny reated when a tariff-indued prie distortion leads to a stati mis-alloation of resoures. We are interested in how lobbying affets the size of the distortion. Consider first the textbook ineffiieny aused by tariffs that distort the prie of tradeables relative to nontradeables. By assumption, the nontradeable i = 0 setor has a zero tariff. Let us assume for the purposes of empirial work that all tariffs τ i are non-negative. Then the distortion between the prie of tradeables and nontradeables an be modeled as the weighted average tariff Eτ Σ n i=1 θ iτ i 0 where the arbitrary positive weights θ i sum to unity. From theorem 2, Eτ/ a < 0. That is, good institutions as measured by a lead to lower tariffs. The same is not true for institutions as measured by b: Eτ/ b annot be signed without additional information about the G i. A ountry with good institutions as measured by large values of b will tolerate high tariffs in industries that generate positive externalities. 8 To understand φ i, let Ω be the Goldberg and Maggi (1999) objetive funtion adopted to our setting with externalities. Let Ω jj be its seond derivative with respet to p j. Then φ j (p j ) Ω jj (p j )/x j (p j). The seond order onditions for a maximum (Ω jj < 0) ensure φ j > 0. See Appendix B for details. 9 An obvious question is about the onditions under whih the diret effet is larger than the total effet: when is φ i less than a + α L? With quadrati utility, quadrati profits and linear externalities, φ i = (a + α L ) (I i α L )q i /x i. Hene, the total effet of institutions is larger than the diret effet if and only if industry i is organized (I i = 1). Further, the bigger is the supply response (q i ), the smaller is φ i. That is, ountries with good institutions put more weight on Ramsey rule onsiderations. 10

13 This is of interest beause researhers who regress per apita GDP growth on average tariffs often appeal to externalities to explain how higher average tariffs an ause higher growth (Edwards, 1992, 1998; O Rourke, 2000; Vamvakidis, 2002; Yanikkaya, 2003; Clemens and Williamson, 2004; Jaks, 2005; DeJong and Ripoll, 2006). Our results show that beause of the endogeneity of tariffs, there is no lear theoretial relationship between growth and average tariffs. B. Distortions within the Tradeables Setor: The Variane of Tariffs We have explored prie distortions between the tradeable and nontradeable setors. The rihness of the Grossman-Helpman framework allows us to examine prie distortions within the tradeables setor. One measure of these prie distortions is the variane of tariffs aross setors, Part 2 of theorem 2 implies Vτ Σ n i=1 θ i(τ i Eτ) 2. dvτ = 2 Vτ da + 2 Cov(τ a + α L a + α i, G i ) db (5) L where the Cov operator means the ovariane. See Appendix C for a proof. This implies that Vτ/ a < 0 i.e., better institutions as measured by a shrink all tariffs to zero, thus reduing the prie distortion within the tradeables setor. Consider the sign of Vτ/ b. Casual empiriism suggests that industries whih generate positive externalities are often organized and suessful at getting what they want from law-makers. Thus, one expets G i and (I i α L ) to be positively orrelated. In this ase, we show in Appendix C that Cov(τ i, G i ) > 0. Thus equation (5) implies Vτ/ b > 0 i.e., better institutions as measured by b lead to greater prie distortions within the tradeable setor. Thus, Vτ/ a and Vτ/ b are of opposite sign: better institutions an be assoiated with either larger or smaller prie distortions within the tradeables setor. 11

14 C. Externalities and the Skill-Bias of Tariff Struture We next onsider the externality bias of the tariff struture. Reall that in equation (4) we ranked industries by G i > G i 1. That is, a tariff in industry i is more growth-enhaning or less growth-retarding than a tariff in industry i 1. An immediate impliation of theorem 2 is the following result about the impat of institutions on tariffs holding onstant the struture of prodution: [ d(τ i τ i 1 ) = τ i τ i 1 a + α L ] da + [ G i G i 1 a + α L ] db. (6) It is lear that better institutions as measured by b will lead the government to favour industries that generate positive externalities or at least that do not generate overly large negative externalities: (τ i τ i 1 )/ b > 0. Interestingly, there are solid empirial reasons for thinking that (τ i τ i 1 )/ a > 0. As is well known, almost every ountry in the world has a tariff struture that is biased towards unskilled-intensive industries e.g., in both rih and poor ountries lothing is proteted more heavily than omputers. More generally, in the data we will be presenting on 3-digit ISIC industries in 63 ountries, the ross-industry orrelation between tariffs and skill intensity is negative in 52 of 63 ountries and signifiantly positive in only 1 ountry. The interesting insight omes from equating the ordering of industries by G i with the ordering of industries by skill intensity. The onnetion between skill intensity, human apital and externalities is a familiar one (e.g., Moretti, 2004). If orderings by externalities and skill intensities overlap, then we know empirially that τ i < τ i 1. Hene equation (6) implies (τ i τ i 1 )/ a > 0. As a rises, tariffs fall most in unskilled intensive industries. Why? Sine tariffs are highest in unskilled intensive industries the easiest way to redue average tariffs is to redue tariffs disproportionately on unskilled intensive industries. We have now established that better institutions as defined by both a and b are onsistent with a skill-biased tariff struture. This is the first time in our paper that inreases in a and b move tariff struture in the same diretion. It suggests that measures of the skill-bias of protetion play a unique role for determining the orrelation between growth and the struture of tariffs. This novel finding omes out of the protetion for sale logi. 12

15 Table 1. Summary of the Relationships Between Domesti Institutions and the Tariff Struture. Average Variane of skill-bias of Tariffs: Eτ Tariffs: Vτ Tariffs: SBτ Prie of tradeables Pries within vs. nontradeables tradeables Externality Eonomi Effet: (Growth)/ (tariff variable) + Politial Eonomy Effet: (tariff variable)/ a + (tariff variable)/ b? Notes: 1 Positive if G i > 0 for all i and negative G i < 0 for all i. Indeterminate if the sign of G i varies aross i. 2 Requires an additional assumption that the ovariane between (I i α L ) and G i is either positive or not too negative. Table 1 summarizes the results of our theory. The Eonomi Effet row of the table shows, for eah feature of the tariff struture, the predited eonomi effet on growth. Beause average tariffs distort the prie of tradeables relative to nontradeables, higher average tariffs redue growth. (This is a one-time hit to output.) As well, the variane of tariffs distort pries within tradeables, reduing growth. If there is an externality, the eonomi effet of a skill-biased tariff struture is positive beause it distorts the alloation of resoures towards industries in a manner that is either more growth enhaning or less growth retarding. The Politial Eonomy Effet rows show the orrelation between tariff struture and institutions a and b. Only for the skill-bias of the tariff struture does the orrelation with a have the same sign as the orrelation with b. In a regression of growth on tariff measures, these rows give the diretion of the endogeneity bias aused by the orrelation between tariff poliy and institutions a and b. The table will thus guide the implementation of the regressions and the interpretation of the oeffiients. 13

16 5. Estimating Equations and Data We now introdue our basi regressions. Let q it be the output of industry i in ountry in year t. We onsider only long-term growth so that we take t = 0 to be the initial year (usually 1972) and t = 1 to be the final year (usually 2000). Our dependent variable is the average annual hange in log output, denoted ln q i1 /q i0. 10 Let τ i0 be the log tariff in industry i in ountry in the initial year t = 0. Let Eτ 0 and Vτ 0 be the output-weighted average tariff and variane of tariffs, respetively, in ountry in the initial year 0. Let SBτ 0 be the skill-bias of tariffs in ountry in the initial year 0. We will define it arefully below. We are interested in the impat of tariffs τ i0 on output growth. However, if we are interested in the full effet of tariffs, then we must reognize that in general equilibrium tariffs in one industry affet output in other industries. Therefore, our industry regressions inlude the national measures of tariff struture Eτ 0, Vτ 0 and SBτ 0. This leads to our estimating equation: ln q i1 /q i0 = β τ τ i0 + β E Eτ 0 + β V Vτ 0 + β SB SBτ 0 + β q ln q i0 + X 0 β X + α i + ε i. (7) Following the tradition of the tariffs-and-growth literature, we regress output hanges on tariff levels. See Edwards (1992, 1998), O Rourke (2000), Vamvakidis (2002), Yanikkaya (2003), Clemens and Williamson (2004), Jaks (2005) and DeJong and Ripoll (2006). This strategy is partiularly relevant in our ontext where tariffs are orrelated with institutions a and b that are best viewed as time-invariant ountry harateristis. Our results about tariffs and institutions examined diret effets. In terms of theorem 2, this means holding onstant industrial struture q i and x i. Assuming that the slope of the export funtion is the same for all ountries, x i is perfetly orrelated with our industry fixed effets α i. That is, we do not need to separately inlude x i in our equation. This leaves only q i, whih appears in our regression as ln q i0. We also ontrol for a large number of other ountry-speifi variables X 0 that will be introdued as used. 10 Note that this is an abuse of notation. More preise notation for our dependent variable would be 1 t 1 t 0 ln q i,,t1 /q i,,t0 where t 0 is the initial year and t 1 is the final year. 14

17 We are also interested in the relationship between tariff struture and the average annual growth rate of per apita GDP, ln y 1 /y 0, where y t is per apita GDP in ountry in year t. 11 This leads to our ountry level estimating equation: ln y 1 /y 0 = β E Eτ 0 + β V Vτ 0 + β SB SBτ 0 + X 0 β X + ε. (8) The sample is determined primarily by the availability of industry level tariff data. The industry level sample onsists of 59 ountries and the ountry level sample inludes an additional 4 ountries. The list of ountries appears in table 3. To math the SITC-based tariff data with ISIC-based output data we have aggregated the data up to a ommon set of 17 manufaturing industries. This is desribed in Appendix D. There are potentially 1,003 (= 59 17) industry level observations. However, some of the output data are missing, leaving us with only 942 industry level observations. Output data are from UNIDO s STAT prodution database as desribed in UNIDO (2003). There are 63 (= ) ountry level observations. Per apita GDP data are from the Penn World Tables (PWT), Mark 6.1. Tariff data are from Lai and Trefler (2002) and UNCTAD (1994). As noted above, in the regressions we use the log of ad valorem tariffs rather than the ad valorem tariffs themselves. This is a minor point that is explained in Appendix D. 6. OLS Results for the Mean and Variane of Tariffs In this setion we present results for the growth-tariff nexus exluding the skill-bias of tariffs SBτ 0. We make two points. First and foremost, we wish to familiarize the reader with the type of regressions we are estimating. Seond, we show that there is no relationship between growth and either average tariffs Eτ 0 or the variane of tariffs Vτ 0. We will, of ourse, have to persuade the reader that our regressors are orthogonal to the residuals. That is, we must deal with endogeneity bias and omitted-variable bias. Consider table 4. The first set of olumns deal with the industry-ountry level regressions of equation (7). The seond set of olumns deal with the ountry level regressions of 11 Again, more preise notation would be 1 t 1 t 0 ln y,t1 /y,t0. 15

18 equation (8). We begin by explaining the many anillary regressors that omprise X 0. Consider first the Cohort Fixed Effets. The availability of disaggregated tariff data is different for eah ountry so that we are unable to use the same time period for all ountries. For 21 ountries tariffs are available beginning in 1972, for 30 ountries tariffs are available beginning in the period and for 12 ountries tariffs are available beginning in the period. See table 3 for details. To ontrol for the three different entry periods of these three ohorts, we inlude ohort fixed effets i.e., three dummy variables with dummy D j equal to 1 if ountry entered in ohort j = 1, 2, 3. (The 1972 ohort dummy is the omitted dummy.) From table 4, these ohort dummies are jointly insignifiant. 12 Consider next the region fixed effets. Ideally, we would like to inlude ountry fixed effets. However, our tariff variables Eτ 0 and Vτ 0 are ountry harateristis that would be annihilated by ountry fixed effets. Instead, we take the next best approah whih is to inlude detailed region fixed effets. We define 10 regions distilled from the PWT regional lassifiation and inlude 9 region dummies in our regressions. See table 3 for a list of whih ountries are in whih regions. We are using far more regional dummies than is typial. This is an important point. It means the sample variation that we are explaining is all within narrowly defined regions. Thus, for example, we are using tariff struture to explain growth differenes within North Afria/Middle East (Algeria, Egypt, Iran, Moroo, Syria and Tunisia). We are not using tariff struture to explain why Algeria has grown slower than the United States. In addition to these region fixed effets we also inlude the familiar ountry harateristis that appear in growth regressions. These are the log of initial per apita GDP, the log of initial human apital, and the initial investment to GDP ratio. See Appendix D for data soures. Without the 9 region fixed effets, these three ountry harateristis are statistially signifiant. However, one the region fixed effets are inluded the three 12 The data end in However, in the ountry level regressions, there are three ountries for whih per apita GDP data end in 1996: Singapore, Cyprus and Sierra Leone. A dummy for this 1996 ohort is also statistially insignifiant. 16

19 ountry harateristis beome statistially insignifiant. Our main point here is that it will be very hard to find a statistially signifiant ountry harateristi when so many region fixed effets are inluded. This, we believe, makes our results for the skill-bias of tariffs more persuasive. Finally, we inlude industry fixed effets. With 17 industries, we inlude 16 industry dummies. We also inlude the log of initial industry output in ountry, ln q i0. We now turn to the tariff results. Consider first the own tariff effet aptured by the tariff in industry i in ountry, τ i0. This an only appear in the industry-ountry level regressions. The oeffiient is positive and statistially signifiant (β τ = 0.005, t = 2.20). As expeted, higher tariffs in industry i in ountry lead to higher output in industry i in ountry. The estimated magnitude is large. For example, a one standard deviation inrease in the own tariff leads to a 1.0% standard deviation inrease in average annual output growth. 13 Turning to the tariff struture variables of our theory, average tariffs Eτ 0 do not matter for output growth at the industry-ountry level ( β E = 0.007, t = 1.82). Neither does the variane of tariffs Vτ 0 ( β V = 0.019, t = 0.53). When regressing industry-ountry level variables on ountry level regressors suh as Eτ 0, we report t-statistis alulated from lustered standard errors. The third olumn in the table shows that the lustered standard errors are always smaller than the OLS standard errors. For the remainder of this paper, for ountry level regressors that appear in the industry-ountry output regressions we report t-statistis alulated from lustered standard errors. One way of measuring whether the effets of Eτ 0 and Vτ 0 are large is to onsider their overall effet on average annual per apita GDP growth. We do this in the seond set of olumns of table 4. Consistent with the industry-ountry level results, at the ountry level the mean and variane of tariffs are statistially insignifiant. They are also eonomially small. For example, a one standard deviation fall in the variane of tariffs leads to only a.0006 standard deviation rise in the growth of per apita GDP. 13 See table 2 for the sample statistis used to derive this. 17

20 It is possible that the insignifiane of Eτ 0 and Vτ 0 results from a high degree of orrelation of these two variables with the many ountry harateristis that are also inluded in the regression. This is not the ase. Figure 1 shows that there is no relationship between average tariffs and growth. Figure 1 is a graph of the bivariate plot (i.e., without ontrols) of Eτ 0 against the average annual growth of per apita GDP. 14 Figure 2 shows the partial regression plot (i.e., with ontrols) from the regression reported in table 4. No relationship is apparent in this plot either. Although we do not report them here, the plots for Vτ 0 are similar. These results may seem surprising. However, they are ompletely onsistent with the reent empirial evidene showing that there is no lear relationship between average tariffs and subsequent eonomi growth. Whether or not there is a signifiant relationship between average tariffs and growth and whether this relationship is positive or negative depends ritially on the ountries in the sample (DeJong and Ripoll, 2006) and on the time period under onsideration (O Rourke, 2000; Vamvakidis, 2002; Yanikkaya, 2003; Clemens and Williamson, 2004; Jaks, 2005). Having established that there is no relationship between average tariffs and growth or tariff dispersion and growth, we turn our fous to the skill-bias of tariffs. 7. The Skill-Bias of Tariff Struture We onstrut our measure of the skill-bias of the tariff struture SBτ 0 as follows. Define unskilled workers as those with less than 12 years of shooling. All other workers are defined as skilled. Next, rank industries based on the ratio of skilled workers (S i ) to unskilled workers (L i ). Table 5 displays S i /L i for the United States in Data are from Antweiler and Trefler (2002). Our first measure of the skill-bias of tariff struture is the ross-industry orrelation between skill intensity and initial tariffs: ρ = Corr{τ i0, S i /L i } The Correlation Measure of the Skill-Bias of Tariff Struture. 14 To prevent the observations from being too ompressed in figure 1 we omit Hong Kong. 18

21 As noted earlier, most of the ountries in our sample have a negative value of ρ. This means that tariffs tend to be high in unskilled-intensive industries and low in skillintensive industries. Figure 3 plots ρ against initial per apita GDP in order to give the reader a sense of the ρ. The two are positively orrelated. This is an indiation of the endogeneity around whih our theory was built. Our seond measure is onstruted as follows. Choose an arbitrary ut-off industry i and define all industries i with S i /L i less than S i /L i as unskilled-intensive and all remaining industries as skill-intensive. Let τ Unskill average tariff for unskilled-intensive industries and let τ Skill be the initial year, output-weighted be the initial year, outputweighted average tariff for skill-intensive industries. Our seond measure of skill-bias is then DIFF τ Skill τ Unskill The Differene Measure of the Skill-Bias of Tariff Struture. An important question is whether our results are sensitive to the hoie of i. As we show in setion 11 below, it does not matter what we hoose for i provided it is not extremely lose to the very top or very bottom of table 5. For now, we proeed by reporting all results for two hoies of i. From table 5, there are two values of i for whih the ratio of skilled to unskilled workers jumps up. In the main text we use these jump points to define two different i, whih we refer to as the low ut-off and the high ut-off. Let DIFF Low and DIFF High be the orresponding differene measures. Another question is whether our results depend on our definition of S i /L i or our use of 1972 U.S. data when onstruting our skill intensity rankings. We will also report results using rankings based on alternative measures of skill intensity. These alternative measures use different definitions of S i /L i and use data from South Afria in 1997 and from Brazil in Virtually idential results are obtained with the alternative measures so we relegate these results to setion 11. The explanation for this surprising robustness is simple. While skill intensities vary aross time and ountries, the relative ranking of industries based on skill intensity barely varies: leather goods are always very unskilled-intensive and 19

22 professional equipment is always very skill-intensive. 15 In total, we have three measures of the skill-bias of tariffs SBτ 0. These are ρ, DIFF Low and DIFF High. To give the reader a sense of the data and how the SBτ 0 measures are related to the quality of domesti institutions we plot DIFF Low against the two most ommonly used measures of a ountry s institutions: (1) demoray, measured using the average demoray sore from 1970 to 1994 from the Polity III database and (2) absene of orruption in government index, measured as the average from 1982 to 1995 from the International Country Risk Guide (ICRG) published by the Politial Risk Servies Group (PRS). The absene of orruption measure aptures the extent to whih high government offiials are likely to demand and aept bribes onneted with speial privileges suh as import and export lienses, tax breaks, loans, et. We show the bivariate relationship between the two variables and DIFF Low in figures 4 and 5. Similar pitures are obtained if one uses ρ and DIFF High. The orrelation oeffiient between DIFF Low and demoray is 0.60 and the orrelation between DIFF Low and orruption is In ontrast, the orrelation between DIFF Low and per apita GDP growth is only Thus, onsistent with our protetion for sale framework, SBτ 0 is highly orrelated with measures of the quality of domesti institutions (reall table 1). A. Estimates of the Impat of Skill-Biased Tariff Struture Table 6 reports our OLS estimates of the industry-ountry level output growth regressions (equation 7) with measures of the skill-bias of tariffs inluded. Table 7 reports the orresponding estimates of the ountry level per apita GDP growth regressions (equation 8). These are the ore empirial tables of the paper. 15 Note that politiians in rih and poor ountries will in general disagree as to whih industries should be proteted i.e., as to whih industry has the largest G i. For example, proteting lothing may make sense in Mauritius whereas proteting airraft may make sense in the United States. See Harberger s (1998) mushrooms model. This is an interesting point whih shows up in some of our results. In partiular, for poor ountries the model does best for relatively small i whereas for rih ountries the model tends to do relatively best for high i. In addition, the best-fit i tends to be lower using the South Afrian and Brazilian skill-intensity rankings than using the U.S. skill-intensity rankings. We do not pursue this interesting point further. 20

23 Column 1 of tables 6 and 7 report the results using ρ as our measure of the skill-bias of the tariff struture. Reall that with so many ontrols we have not yet found a single ountry level variable that is statistially signifiant. In ontrast, the oeffiient for ρ is signifiant. The t-statisti is 4.23 in the industry output growth regression and 3.66 in the per apita GDP growth regression. The magnitude of the oeffiient is very large, a fat that we will detail at the end of this setion. Further, the ontribution to the R 2 is enormous. Adding just ρ to the ountry level per apita GDP growth regression raises the R 2 from 0.60 to The idea that a single regressor ould have suh a huge effet in a speifiation that already has 16 regressors is remarkable. Column 3 of tables 6 and 7 replae ρ with DIFF Low as our measure of the skill-bias of the tariff struture. The results are again statistially signifiant with a t-statisti of 3.46 in the industry output growth regression and 3.51 in the per apita GDP growth regression. In the latter, DIFF Low raises the R 2 from 0.60 to Column 7 uses DIFF High as our measure. The results are even a little stronger than for DIFF Low, with the R 2 rising to To peel bak the onion on what is driving the DIFF introdue τ Skill and τ Unskill expeted, the oeffiient on τ Skill τ Skill τ Unskill results, we separately. We do this for the low ut-off in olumn 5. As is positive and signifiant: t = 4.16 in the industry output growth regression and t = 3.66 in the per apita GDP growth regression. Also as expeted, the oeffiient on τ Unskill is negative and signifiant: t = 2.99 in the industry output growth regression and t = 3.40 in the per apita GDP growth regression. To hek that the results are not being driven by outliers, figures 6 and 7 display partial regression plots for the per apita GDP growth regressions. Figure 6 shows the partial regression plot using ρ (olumn 1 of table 7). Figure 7 shows the partial regression plot using DIFF Low (olumn 3 of table 7). From the plots it is lear that the relationships are not being driven by a small number of influential observations. In setion 11 below we provide a battery of regression diagnostis aimed at showing that the results are robust to a variety of speifiation searhes. We have shown that the oeffiient on the skill-bias of the tariff struture is statistially 21