94665 EXPORT DIVERSIFICATION: WHAT’S BEHIND THE HUMP? ` re, and Vanessa Strauss-Kahn* ´ line Carre Olivier Cadot, Ce Abstract—The paper explores the evolution of export diversification patterns The negative correlation between natural resources and growth along the economic development path. Using a large database with 156 countries over 19 years at the HS6 level of disaggregation (4,991 product was, however, questioned by, among others, Brunnschweiler lines), we look for action at the intensive and extensive margins. We find a (2008) and Brunnschweiler and Bulte (2008), who argued that hump-shaped pattern of export diversification similar to what Imbs and Wac- regressing growth on the share of primary products in exports ziarg (2003) found for production. Diversification and subsequent reconcen- tration take place mostly along the extensive margin. This hump-shaped pat- or GDP suffered from fatal endogeneity problems. tern is consistent with the conjecture that countries travel across While the relationship of endowments, trade, and growth diversification cones, as discussed in Schott (2003, 2004) and Xiang (2007). has remained a controversial issue, how export patterns vary across time and countries has become a subject of I. Introduction intense descriptive analysis in recent years. Several papers (for example, Evenett & Venables, 2002; Hummels & Kle- W HY should export diversification be taken as a policy objective per se? There are two reasons that it should not. First, according to Ricardo, countries should specialize, now, 2005; Kehoe & Ruhl, 2006; Brenton & Newfarmer, 2007) decompose cross-country export variations into intensive and extensive (new-products or new-markets) not diversify. Second, the Heckscher-Ohlin model implies that margins and study the contribution of these margins in export patterns are largely determined by endowments, so, if export growth.2 Digging deeper into the extensive margin, anything, we should worry about factor accumulation, not Hausmann and Klinger (2006) proposed a measure of diversification. Yet export diversification is a constant preoc- ‘‘product proximity’’ based on the conditional probability cupation of policymakers in developing countries. As de Fer- that one product is exported given that the other is also ranti et al. (2002) note, ‘‘A recurrent preoccupation of [Latin exported. American] policymakers is that their natural riches produce a In parallel with this literature, a widely cited paper by highly concentrated structure of export revenues, which then Imbs and Wacziarg (2003) uncovered a nonmonotone path leads to economic volatility and lower growth’’ (p. 38). of production and employment diversification as functions The notion that export patterns are fully determined by of per capita incomes, with diversification followed by endowments is of course naive. The relationship of endow- reconcentration. Imbs and Wacziarg’s work naturally raised ments, trade, and growth is a complex and imperfectly under- the question of whether a similar pattern would hold for stood one. Intra industry trade models showed long ago that exports as well. Klinger and Lederman (2004, 2006) indeed many factors other than endowments, including market failures found that exports diversify and then reconcentrate with and policies, can affect trade patterns. More recently, Haus- income. While Imbs and Wacziarg’s exercise was essen- mann, Hwang, and Rodrik (2007) argued that export patterns tially an empirical one, Klinger and Lederman built on can display path dependence in the presence of externalities. Hausmann and Rodrik (2003) to explore a causal link from Policy concerns about a linkage between the concentration market failures to insufficient diversification. The argument of exports on primary products and deteriorating terms of trade, is that opening up new export lines is an entrepreneurial income volatility and, ultimately, low growth go back to the gamble; if it is successful, it is quickly imitated. The inabil- work of Prebisch (1950) and Singer (1950). Subsequent work ity of ‘‘export entrepreneurs’’ to keep private the benefits of (for example, Neary & van Wijnbergen, 1986; Gelb, 1988; their activity thus leads to a classic public-good problem. Auty, 1990; Sachs & Warner, 1999) showed a robustly nega- We revisit the issue using a different perspective, in tive correlation between dependence on primary products and which we derive and analyze a decomposition of Theil’s future growth, a finding called the ‘‘natural-resource curse.’’1 concentration index that maps directly into the extensive and intensive margins of export diversification. In order to Received for publication December 7, 2007. Revision accepted for pub- analyze how the two margins evolve as functions of GDP lication January 6, 2010. * Cadot: World Bank, HEC Lausanne, CERDI, CEPR and CEPRE- per capita, we construct a very large database covering 156 MAP; Carre ` re: CERDI-CNRS, Universite ´ d’Auvergne; Strauss-Kahn: countries (including 141 developing ones) over all years ESCP-Europe and CEPII. available from the COMTRADE database at the highest Research on this paper was supported by a grant from the Interamerican Development Bank and by Switzerland’s Fonds National pour la disaggregation level (HS6). Using this database, we calcu- Recherche Scientifique. We thank the FERDI (Fondation pour les Etudes late for all countries and years three classes of variables of et Recherches sur le De´ veloppement International) for publication support. interest: export concentration indices (focusing on Theil’s Special thanks go to Julien Gourdon for calling our attention to key data and estimation issues. We also thank Marius Bru ¨ lhart, Antoni Estevadeor- index and its decomposition), the number of active lines dal, Christopher Grigoriou, Jaime de Melo, Marcelo Olarreaga, Christian (lines with nonzero exports), and a measure of ‘‘new export Volpe, two anonymous referees, and the editor for useful comments. 1 The Prebisch-Singer hypothesis implies that low growth is caused by 2 dependence on primary products, not necessarily by concentration per se. The intensive margin reflects variation in export values among exist- However, preliminary findings by Dutt, Mihov and van Zandt (2008) sug- ing exports, whereas the extensive margin reflects variation in the number gest that diversification does accelerate future growth, especially when it of new products exported or in the number of new markets for existing is accompanied by convergence toward the U.S. pattern of exports. exports. The Review of Economics and Statistics, May 2011, 93(2): 590–605 Ó 2011 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology EXPORT DIVERSIFICATION 591 products.’’ We use these three variables to explore action database at the HS6 level (4,991 lines).3 The baseline sam- along the intensive and extensive margins. In essence, we ple covers 156 countries representing all regions and all propose a decomposition of the Theil index in between- levels of development between 1988 and 2006 (19 years), groups and within-groups components that can be easily including 141 developing countries—non-high-income mapped into the extensive and intensive margins, respec- countries, defined by the World Bank as countries with tively. 2006 per capita GDP under $16,000 in constant 2005 PPP We find a hump-shaped relationship between economic international dollars. After we take out missing-year data, development and export diversification, like Imbs-Wacziarg the usable sample has 2,797 observations (country-years). and Klinger-Lederman, with a turning point around In this section, we compute several measures of export $25,000 per capita at purchasing power parity (PPP). The concentration/diversification for each country and year: observed reconcentration might be spurious in a number of Herfindahl concentration indices, Theil and Gini indices of ways. For instance, it could be driven by small, rich, and inequality in export shares, and the number of active export concentrated oil producers. It could also be an artifact of lines. The Herfindahl index, normalized to range between 0 the Harmonized system (see Appendix A). This would be and 1, is the case if low- and middle-income countries were mainly P exporting products from sectors with large numbers of à ðsk Þ2 À 1=n H ¼ k ; export lines such as the textile sector. Alternatively, 1 À 1=n observed concentration pattern could be driven by unex- plained heterogeneity between countries. We find that none P n where sk ¼ xk = xk is the share of export line k (with of the obvious culprits stands scrutiny. In particular, the k¼1 reconcentration holds strongly within country: all countries amount exported xk) in total exports and n is the number of to the right of the turning point reconcentrate over time. export lines (omitting country and time subscripts). We use At income levels below the turning point, we find diver- the following formula for the Gini index: sification at both the extensive and intensive margins, but X n mostly along the extensive margin until around PPP G¼1À ðXk À XkÀ1 Þ=n; $22,000. The intensive margin briefly dominates around the k¼1 turning point; thereafter, the extensive margin retakes the lead and explains the reconcentration, suggesting that rich P k where Xk ¼ sl represents the cumulative export shares. countries close export lines. What are those products disap- l¼1 pearing from rich-country export portfolios? We find that Theil’s entropy index (Theil, 1972) is given by the factor intensities of those products are typically far   away from the countries’ endowments, as if they were left- 1X n xk xk 1X n T¼ ln where l ¼ xk : ð1Þ overs from old export patterns kept alive only by hysteresis. n k¼1 l l n k ¼1 That is, our evidence suggests that as countries travel across diversification cones, they fail to close a tail of export lines that no longer belong to their comparative advantage but Table 1 shows descriptive statistics for these indices. artificially inflate their diversification, until finally compara- Observe that Gini indices are very high. The reason has tive advantage catches up. to do with the level of disaggregation: we use a very disag- The paper is organized as follows. Section II reports gregated trade nomenclature. At that level, we have a large econometric evidence on the stages of export diversification number of product lines with small trade values, while a in the process of economic development. In order to better relatively limited number of them account for the bulk of understand what is behind the hump-shaped diversification all countries’ trade (especially for developing countries, but curve, section III analyzes action along the intensive and even for industrial ones). As for the average number of extensive margins by examining the evolution of the within positive export lines—active lines with non zero trade and between component of the Theil concentration index. It values—it is relatively low at 2,062 per country per year— also explores the specificities of the new export products a little less than half the total, with a minimum of 8 for Kiri- that generate diversification. Section IV explores potential bati in 1993 and a maximum of 4,988 for Germany in 1994 explanations behind the diversification curve. Section V and the United States in 1995. This implies that there is concludes. room for a substantial extensive margin for developing countries, especially the poorest and least diversified ones. II. Stages of Diversification: Estimation Per capita GDPs are taken from the World Bank’s World Development Indicators (WDI) and are expressed in 2005 A. Measures of Export Concentration/Diversification purchasing power parity (PPP) dollars for comparability. Our dataset comprises data on trade and income per 3 The appendix provides further information on the COMTRADE HS6 capita. Export data are from UNCTAD’s COMTRADE level database. 592 THE REVIEW OF ECONOMICS AND STATISTICS TABLE 1.—DESCRIPTIVE STATISTICS: 156 COUNTRIES OVER 1988–2006 Variable Observation Mean S. D. Minimum Maximum Gini 2,797 0.967 0.045 0.773 1.000 Herfindahl 2,797 0.189 0.235 0.002 0.989 Theil 2,797 4.865 1.797 1.478 8.465 Nber of active lines 2,797 2,061.8 1,669.6 8 4,988 GDPpc PPP, constant international 2005 dollars 2,695 9,442.1 11,130.9 136.5 73,276.9 Share of oil in exports 2,797 0.190 0.287 0 0.996 Author calculations using COMTRADE. FIGURE 1.—PREDICTED THEIL’S CONCENTRATION INDEX & NUMBER OF ACTIVE tion between the export concentration and per capita GDP EXPORT LINES (figure 1). One issue is whether the turning point is driven by micro- states and island economies, which are very heterogeneous in GDP per capita and at the same time very concen- trated—say, in bananas or fish products. Because micro- states are potential outliers, we omit them in the rest of the analysis (that is, we exclude fifteen countries with popula- tions below 1 million). A second issue is that of omitted variables. First, spurious correlation could be introduced by fluctuations in the world price of oil and other commodities, as higher commodity prices would raise both per capita incomes and export con- centration for primary-product exporters. Columns 1 to 4 of table 2, which report pooled estimates with time effects, show a turning point around 25,000 PPP international (2005 Note: Quadratic corresponds to the OLS estimation of Yit ¼ a0 þ a1 GDPpcit þ a2 GDPpc2 constant) dollars. This turning point is quite similar to the it þ mit , with Yit being alternatively the Theil index and the number of active export lines. Nonparametric corresponds to one found by Imbs and Wacziarg for production and by smoother nonparametric regressions of Yit on the GDPpc. Author calculations using COMTRADE. Klinger and Lederman (2004) for exports on a panel of 130 countries from 1992 to 2003 ($22,500 in constant 2000 dol- lars).6 B. Parametric Evidence Second, given the panel structure of our data set, a nat- Figure 1 depicts curves representing predicted values of ural question is the type of estimator—within, between, or Theil’s index, as well as curves representing the predicted pooled—we should use. Imbs and Wacziarg’s estimation on number of active export lines.4 The latter, which are con- production data relies on fixed effects (that is, within). Col- cave and increasing at the origin, are easy to distinguish umns 5 to 12 of table 2 show our results using the within from the former, which are convex and decreasing at the and between estimators. The turning point stays significant origin. and at a similar level of GDP per capita. Apart from its The Theil curve is fitted using quadratic polynomial level, what matters is which countries are on either side of regressions of the Theil concentration index on per capita the turning point. When Theil regressions are used the GDP using pooled OLS with White-corrected standard between and pooled estimators return the same list of 21 errors. We find a turning point around $30,000 in PPP countries to the right of the turning point. The within esti- (2005 constant).5 We also estimated smoother nonpara- mator adds only 2 (Israel and New Zealand).7 metric regressions (dashed curves). This consists of reesti- Table 3 reports a number of robustness checks. First, we mating the regression for overlapping samples centered on consider censoring, as Gini coefficients are bounded left each observation. Smoother regressions impose no func- and right, at 0 and 1, respectively, although neither is bind- tional form and are therefore suited to the exploration of ing the strictest sense. We thus perform a logistics transfor- highly nonlinear relationships. The nonparametric estimates validate the use of a quadratic form to approximate the rela- 6 The value of our turning point is not directly comparable to that of Imbs and Wacziarg, as they used Summers-Heston per capita incomes in constant 1985 dollars. They note, however, that their turning point occurs roughly at the level of income reached by Ireland in 1992. Our turning point corresponds roughly to Ireland’s income level in 1996. 4 7 Fitted curves for Herfindahl and Gini indices have similar shapes. Measurement errors in explanatory variables, if they are correlated 5 We also explore the turning point’s stability across different defini- with the error term, create a downward bias in estimated coefficients that tions of GDP per capita (i.e., per capita GDP at PPP from the Penn World is especially severe with fixed effects (see Griliches and Hausman, 1986). Tables and per capita GDP in constant US dollars from the WDI). If present, this would push the turning point to the left compared to Results, which are similar across definitions, are available on request. pooled and between estimates. TABLE 2.—POOLED, WITHIN, AND BETWEEN ESTIMATES Pooled Within Between (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Dependent HHI Theil Gini Nber HHI Theil Gini Nber HHI Theil Gini Nber GDPpc À1.89E-05 À0.0002516 À5.98E-06 2.65E-01 À6.50E-06 À0.0000779 À2.63E-06 3.90E-01 À1.89E-05 À0.0002573 À5.84E-06 2.68E-01 12.48*** 23.40*** 21.34*** 36.53*** 2.22*** 4.91*** 9.46*** 23.96*** 2.57** 4.85*** 4,52*** 7.68*** GDPpc2 4.09E-10 4.99E-09 1.12E-10 À4.67E-06 1.38E-10 1.83E-09 5.87E-11 À6.98E-06 4.21E-10 5.20E-09 9.95E-11 À4.79E-06 9.49*** 15.40*** 10.27*** 21.00*** 2.52*** 6.18*** 11.27*** 18.67*** 1.90* 3.11*** 2.00** 4.26*** Turning point ($) 23,105 25,210 26,744 28,396 23,551 21,284 22,402 27,928 22,447 24,740 29,347 28,012 R2 0.12 0.37 0.50 0.64 0.10 0.32 0.43 0.58 0.10 0.36 0.51 0.63 Observations 2,497 2,497 2,497 2,497 2,497 2,497 2,497 2,497 141 141 141 141 Number of countries 141 141 141 141 141 141 141 141 141 141 141 141 Period 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 Countries on the Right of the Turning Point, 2006 Australia Australia Australia Australia Australia Australia Australia Australia Australia Australia Australia Australia Austria Austria Austria Austria Austria Austria Austria Austria Austria Austria Austria Austria Belgium Belgium Belgium Belgium Belgium Belgium Belgium Belgium Belgium Belgium Belgium Belgium Canada Canada Canada Canada Canada Canada Canada Canada Canada Canada Canada Canada Denmark Denmark Denmark Denmark Denmark Denmark Denmark Denmark Denmark Denmark Denmark Denmark Finland Finland Finland Finland Finland Finland Finland Finland Finland Finland Finland Finland France France France France France France France France France France France Greece Greece Greece Greece Greece Greece Greece Greece Greece Greece Germany Germany Germany Germany Germany Germany Germany Germany Germany Germany Germany Germany Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Hong Kong Ireland Ireland Ireland Ireland Ireland Ireland Ireland Ireland Ireland Ireland Ireland Ireland Israel Israel Israel Israel Italy Italy Italy Italy Italy Italy Italy Italy Italy Japan Japan Japan Japan Japan Japan Japan Japan Japan Japan Japan Netherlands Netherlands Netherlands Netherlands Netherlands Netherlands Netherlands Netherlands Netherlands Netherlands Netherlands Netherlands New Zealand New Zealand New Zealand New Zealand Norway Norway Norway Norway Norway Norway Norway Norway Norway Norway Norway Norway Singapore Singapore Singapore Singapore Singapore Singapore Singapore Singapore Singapore Singapore Singapore Singapore Spain Spain Spain Spain Spain Spain Spain Spain Sweden Sweden Sweden Sweden Sweden Sweden Sweden Sweden Sweden Sweden Sweden Sweden Switzerland Switzerland Switzerland Switzerland Switzerland Switzerland Switzerland Switzerland Switzerland Switzerland Switzerland Switzerland United United United United United United United United United United United United Kingdom Kingdom Kingdom Kingdom Kingdom Kingdom Kingdom Kingdom Kingdom Kingdom Kingdom Kingdom United States United States United States United States United States United States United States United States United States United States United States United States Absolute value of robust t-statistics under coefficients. ***, **, *: Significant at, respectively, 1%, 5%, and 10% level. The full sample except microstates, GDP per capita PPP in constant 2005 international dollars from WDI. Author calculations using COMTRADE. 594 THE REVIEW OF ECONOMICS AND STATISTICS TABLE 3.—ROBUSTNESS Logistic Transformation Negative Binomial System GMM (1) (2) (3) (4) (5) (6) (7) Dependent Gini Nber Nber HHI Theil Gini Nber GDPpc À2.55E-04 3.18E-04 1.45E-04 À2.14E-05 À2.80E-04 À6.95E-06 2.97E-01 25.57*** 31.20*** 29.32** 4.23*** 7.10*** 7.18*** 11.42*** 2 GDPpc 4.85E-09 À4.98E-09 À2.83E-09 4.46E-10 5.61E-09 1.20E-10 À5.41E-06 16.21*** 16.30*** 20.42*** 3.49*** 5.35*** 4.61*** 7.46*** Turning point ($) 26,320 31,908 25,583 23,991 24,955 28,958 27,412 Observations 2497 2497 2497 2497 2497 2497 2497 Number of countries 141 141 141 141 141 141 141 Period 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 Absolute value of robust t-statistics under coefficients. ***, **, *: Significant at respectively 1%, 5%, and 10% level. Full sample except microstates. GDP per capita PPP in constant 2005 international dollars from WDI. Author calculations using COMTRADE. mation whose results are reported in columns 1 and 2. The and export diversification. Our next task is to understand turning point is at the usual level of about $26,000. Second, what is behind the hump. we correct for the potential endogeneity of GDP per capita to export concentration. As we have no valid outside instru- III. Stages of Diversification: Extensive versus ment for GDP per capita for our large panel, we carry out a Intensive Margins system GMM estimation. Results, presented in the columns 4 to 7, show a turning point varying between $24,000 (Her- That export diversification would proceed in parallel with findahl) and $29,000 (Gini), with the same countries to the economic development is to be expected. Pretty much like right of the turning point.8 Thus, by and large, both the exis- human beings colonized new land to alleviate competitive tence of a turning point in export concentration and its loca- pressure on existing pastures, entrepreneurs can be expected tion around a GDP per capita of about $22,000 to $27,000 to look for new pastures and open up production and export at PPP in constant 2005 international dollars—a very late lines at the extensive margin. As capital accumulates, this point in the development process—are fairly robust. becomes easier. But the later reconcentration, although con- A glance at the columns entitled ‘‘Nber’’ in tables 2 and sistent with Imbs and Wacziarg’s findings for production 3 shows a clear hump-shaped relation between the number and employment, is somewhat of a puzzle. In order to better of active export lines and GDP per capita. The turning point understand what is behind the hump in the curve, we turn to for the number of active export lines is always roughly at a systematic analysis of the intensive and extensive margins the same level of GDP per capita as that of the concentra- using the decomposability property of Theil’s index. tion indices (see also figure 1). As the number of lines is a The nonmonotone pattern of diversification revealed in count variable, we also run a negative binomial estimation. section II (decreasing concentration up to $25,000 and Results, reported in column 3 of table 3 are consistent with increasing concentration thereafter) could be explained by previous findings. The rising part of the curve corresponds change at the extensive margin, the intensive margin, or to the introduction of new products as countries develop. Its both. Diversification at the extensive margin occurs when decreasing part illustrates one of the striking findings of this the number of active lines rises. Diversification at the inten- paper: that high-income countries tend to close down export sive margin occurs when the distribution of trade values lines faster than they open up new ones, resulting in recon- across existing export lines becomes more even. That is, centration at the extensive margin. We return to this point diversification at the intensive margin during a period t0 to later. t1 means convergence in export shares among goods that Thus, our analysis, regressing concentration indices and were exported at t0. The evolution in the number of active the number of active lines on GDP per capita, shows a lines identified in section II is suggestive of action at the hump-shaped relationship between economic development extensive margin. In order to shed more light on the issue, we turn to a decomposition of Theil’s index, which can be usefully mapped into the intensive and extensive margins 8 thus defined. A crucial issue with system GMM (Blundell & Bond, 1998) is the number of instruments to use. This number should not exceed the number of individuals in the panel (see Roodman, 2006). We make the standard A. Mapping the Theil Decomposition with the Extensive choice of using two lags for the instruments of the differenced equation and one lag for the instruments of the level equation. Following Arellano and Intensive Margins and Bond (1991), we use the Sargan/Hansen test of overidentifying restrictions and a direct test for the absence of second-order serial correla- In this section, we combine the classic decomposition of tion. Both fail to reject the null of no serial correlation. Theil’s index into between- and within-groups components EXPORT DIVERSIFICATION 595 with a partition of export lines into active and inactive ones. for their mean to also tend to 0), it follows that n1 l1 ! nl, The result is a perfect mapping of changes in the between- so groups component of Theil’s index into changes in the     extensive margin of exports and of changes in its within- B l1 n lim T ¼ ln ¼ ln : ð6Þ groups components into changes in the intensive margin of l0 !0 l n1 exports. Letting D denote a period-to-period change and obser- Theil’s index has the property that it can be calculated ving that n is time invariant, we have finally that for groups of individuals (export lines) and decomposed additively into within-groups and between-groups compo- lim DT B ¼ ÀD ln n1 : ð7Þ nents (that is, the within- and between-groups components l0 !0 add up to the overall index). Specifically, let n be the total number of potential export lines (the 4,991 lines of the HS6 That is, given our partition, changes in the between- system) and l their average dollar value. Consider some groups component of Theil’s index measure changes at the partition of that total number of potential exports (of a extensive margin (proportional changes in the number of given country in a given year) into J þ 1 groups denoted active lines). Gj, j ¼ 0,. . .,J. Let nj be the number of export lines in group As for the within-groups component, it is a weighted j and lj their average dollar value. Also let Tj stand for average of terms combining group-specific means (lj/l) Theil’s index for group j, calculated using equation (1) on and group-specific Theil indices Tj (the terms in square the nj lines making up group j. Finally, let xk be the dollar brackets), the weights being nj/n. In our case, TW reduces to value of export line k, regardless of which group it belongs T1, the group Theil index for active lines. To see this, write to. The between-groups component of Theil’s index is equation (3) in full as "  # defined as n l 1 X x xk 0 0 k   TW ¼ ln XJ nj lj lj n l n0 k2G l0 l0 B T ¼ ln ; ð2Þ " 0  # ð8Þ j¼0 n l l n1 l1 1 X xk xk þ ln : n l n1 k2G l1 l1 and its within-groups component is defined as 1 2 !3 X J l XJ l X In group G0, suppose that all lines have the same arbi- n j j nj j4 1 x k xk 5 TW ¼ Tj ¼ ln : ð3Þ trary, strictly positive value x0, so l0 ¼ x0. Then the first j¼0 n l j¼0 n l n j l k 2G j l j term in equation (8) is well defined and boils down to j n0 l0 lnð1Þ ¼ 0: n l It is easily verified that T W þ T B ¼ T . Suppose that for a given country and year, we partition Moreover, this remains true as x0 is made arbitrarily the 4,991 lines making up the HS6 nomenclature into two close to 0. Thus, groups: G1 made of active export lines for that country and "  # year, and G0 made of inactive export lines. We want to use W n1 l1 1 X xk xk this partition to construct group Theil subindices, one for lim T ¼ ln : x0 !0 n l n1 k2G l1 l1 each group j ¼ 0,1, and the within and between components 1 of the Theil. The between-groups subindex is not defined ð9Þ since xk ¼ 0 for all k in G0, so that l0 ¼ 0 and consequently the logarithm in expression (2) is not defined for j ¼ 0. Now, as x0 tends to 0, we noted already that n1 l1 ! nl. However, applying L’Ho ˆ pital’s rule gives It follows that    ! W 1 X xk xk l l lim T ¼ ln ¼ T1 : ð10Þ lim 0 ln 0 ¼ 0; ð4Þ x0 !0 n1 k2G l1 l1 l0 !0 l l 1 so given our partition, Thus, given our partition, changes in the within-groups   Theil index (DTW) measure changes at the intensive margin n1 l1 l (DT1, that is, changes in concentration among active lines lim T B ¼ ln 1 : ð5Þ l0 !0 n l l only). In sum, the decomposition of Theil’s index with our parti- P n P tion of export lines into active and inactive ones allows dis- As l ¼ xk =n, l1 ¼ xk =n1 , and liml0 !0 tinguishing changes in overall concentration into extensive- P P n k ¼1 k2G1 and intensive-margin changes. The evolution of the between xk ¼ xk (since lines outside G1 must all tend to 0 k2G1 k¼1 component of the Theil corresponds to changes at the exten- 596 THE REVIEW OF ECONOMICS AND STATISTICS FIGURE 2.—WITHIN AND BETWEEN COMPONENTS OF THEIL’S INDEX (figure 4b confirms that high-income countries specialize in chemicals), chapters 29 and 28 rank high in their number of closed lines. The simultaneous occurrence of rising specia- lization and line closures in the chemical sector is, however, consistent with Schott’s (2004) finding that specialization occurs within sectors, as high-tech exports replace low-tech ones when countries become more prosperous. The closure of export lines in the leather sector, by contrast, suggests between-product specialization, as leather or cotton works are labor-intensive activities in which countries lose com- parative advantage when they grow. We explore this last point more intensively in section IV. B. What Are the New Export Products That Generate Trade Diversification? Author calculations using COMTRADE (quadratic estimates). Although the most intriguing feature of the U-shape pat- tern is the exports’ reconcentration of the richest countries, sive margin, whereas the evolution of the within component patterns of diversification at lower income levels are also of of the Theil reflects changes at the intensive margin. interest. As most of the diversification occurs at the exten- We now put this decomposition to work. Figure 2 depicts sive margin, one may indeed wonder what the characteris- the contribution of the between and within components to tics of those new export products (new lines at the HS6 the overall Theil. We observe that in levels, the within com- level) are. ponent dominates the index, but in terms of evolution, most The number of new export products should be interpreted of the action is in the between component.9 somewhat cautiously, as these products are not necessarily Until about PPP$22,000, the between component shrinks true entrepreneurial discoveries. In most cases, they corre- faster than the within, so diversification occurs mostly at spond to the opening of new export lines that are already the extensive margin. Past that point and until the turn- active in other countries. This is particularly true for devel- around (at around PPP$25,000), it is the within component oping countries that are copying existing products invented that decreases faster, so diversification occurs mostly at the elsewhere and exporting those products as new export lines. intensive margin. That is, individual export values (and In contrast, genuine innovations are incorporated within the shares) converge among active lines. HS6 classification in the course of periodic revisions and Beyond the turning point, the index starts rising again, may not show up as new export lines.10 Our new export and its rise is driven almost exclusively by the between products thus correspond to what Klinger and Lederman component. That is, reconcentration occurs at the extensive (2006) called ‘‘inside-the-frontier innovations.’’ The focus margin as countries close down active export lines. What of our paper is not innovation but export diversification are those lines? within an existing (although arbitrarily limited) product Table B1 in appendix B shows the sectors and chapters nomenclature. Exporting a product for the first time (that is, mostly concerned with closure. The majority of chapters opening a new export line), even if it were already produced listed in table B1 are declining industries in high-income or exported to other destinations, is an entrepreneurial risk countries. Among the fifteen chapters that experienced the worth investigating. highest number of closed lines, three belong to the textiles There is no conventional definition of new export pro- sector, a fourth concerns raw hides and skins and leather, ducts. In order to stay as close as possible to the definition two belong to the vegetable products sector, two others to of active lines and in the tradition of Besedes and Prusa the live animal and animal products, two are from the (2006b), we first define new export products for a year and mineral products sector, and one concerns iron and steel. country as those lines that were not active in the country’s Textiles (chapter 53) and leather (chapter 41) are among the export trade in the preceding year but were exported in the most active ‘‘closers’’ (8.6 percent of the chapter’s active following year (one-year cutoff). This definition, based on a lines for the former, 9.4 percent for the latter). The case of moving three-year window, reduces the sample period to chemicals (chapters 29 and 28) is worth investigating. 1989 to 2005, one year being taken out at both ends. As Although the chemicals sector does not necessarily come alternatives, we use (a) Klinger and Lederman’s (2006) across as a declining sector for most developed countries definition and (b) lines that were inactive in the country’s 10 At the HS6 level, reclassifications are limited, but we follow Besedes 9 When the slope of the overall Theil is at least twice that of its within and Prusa (2006a) in treating them as censored, that is, a spell of, say, five component, the between component contributes for more than 50 percent years ending with a reclassification is treated as a spell of at least five to the overall index’s decrease. years, like one at the end of the sample. EXPORT DIVERSIFICATION 597 FIGURE 3.—PREDICTED NEW EXPORT LINES: NONPARAMETRIC ESTIMATES (a) Using the available data for each definition of new products (1989–2005 for the one year cut-off, 1990–2004 for the two year cut-off, and 1997–2006 for Klinger and Lederman). (b) All new products are computed over the 1997–2005 period. Author calculations using COMTRADE. export trade in the preceding two years but were exported ent in the figure could conceivably be due to equally rapid in the following two years (two-years cutoff). This latter convergence toward the absolute barrier to diversification definition strikes a balance between the very conservative (the 5,000 lines of the HS system), but it is not, as few definition used by Klinger and Lederman (2006) and the countries approach this barrier and certainly not those at very liberal one used by Besedes and Prusa (2006b). GDP per capita levels around $5,000 to $10,000.11 Klinger and Lederman (2006) define ‘‘discoveries’’ as The relationship between income and new export pro- products not exported in the early part of their sample ducts is robust to the choice of definition of new products. (1994–1996) but with over $10,000 of exports in the latter The lower number of Klinger and Lederman’s new export part (2002–2003). What is the difference between this defi- products in figure 3a could be expected from the more con- nition and definitions that account for years of inactivity servative aspect of their definition. It could also result from and activity around the first appearance of a product (one- the shorter time frame on which new products are mea- year or two-year cutoff)? Conceptually these notions of sured. As ten years are required to compute a new product new export products are essentially the same, being based according to Klinger and Lederman, we measured these on the idea that imperfectly informed entrepreneurs search new products for the 1997–2006 period against 1989–2005 for profitable export opportunities. Uncertainty can be about for the Besedes and Prusa definition (one-year cutoff) and production costs, as in Hausmann and Rodrik (2003), or 1990–2004 for the two-year cutoff. Figure 3b depicts the about foreign demand, as in Vettas (2000), but the point is nonparametric estimates of the predicted number of new that starting to export a product is an entrepreneurial gam- export products against GDP per capita for the 1997–2005 ble that may fail. Whereas Klinger and Lederman’s defini- period, which is common to all definitions. Once corrected tion singles out successful export line development (new for the number of years available, new export products per lines that reach a threshold value), the other definitions Klinger and Lederman are similar to new export products include small-volume, short-spell lines in order to pick up defined by the two-year cutoff. The one-year cutoff unsur- the trial-and-error process at the extensive margin. The prisingly counts more new products because it includes sev- shorter the spell, the more discoveries or new products there eral of these new exports with extremely short spells, which should be, as new entrepreneurs try again a few months or can be assimilated to trial-and-error export products. years later. Besedes and Prusa (2006a) found that over half We finally ask whether new export products are any dif- of all trade relationships were observed for a single year ferent from other traditional exports. Table 4 gives a char- and 80 percent lasted less than five years. Our more aggre- acterization of export goods using Rauch’s index of product gated HS6 data are likely to smooth some of those entries differentiation. Rauch (1999) distinguished products traded and exits, but Besedes and Prusa showed the high churning on organized exchanges, products with reference prices, rate to be robust to aggregation. and differentiated ones. Table 4 shows the proportion of Figure 3 shows the predicted number of new export pro- each of Rauch’s categories in traditional and new export ducts (per country-year, with several alternative definitions lines as measured according to Besedes and Prusa’s of new export products) against GDP per capita using the nonparametric (smoother) estimator. In all cases, the turn- ing point comes very early: in the PPP $5,000 to 10,000 11 Recall that on average, only half the HS6 lines are active for any range. The rapid decrease in export entrepreneurship appar- country and year. 598 THE REVIEW OF ECONOMICS AND STATISTICS TABLE 4.—CHARACTERIZATION OF PRODUCTS BY DEGREE OF DIFFERENTIATION New Products New World Trade, (count number) Products All Products 1990 (Rauch)b Conservative classificationa Homogeneous 7.6% 15.0% 31.9% 12.6% Reference priced 28.1% 32.5% 27.4% 20.3% Differentiated 64.4% 52.5% 41.0% 67.1% Liberal classificationa Homogeneous 12.0% 22.6% 39.1% 16.0% Reference priced 26.8% 28.2% 19.7% 19.5% Differentiated 61.2% 49.2% 41.3% 64.2% In value of total trade unless otherwise indicated. Author calculations using COMTRADE. a Because the classification of some products cannot be asserted unambiguously, Rauch’s conservative classification assigns fewer products to the homogeneous and reference-priced categories than his liberal ones. b From table 2 of Rauch (1999). (2006b) definition. Using other definitions for new export dle-income countries or very high-income ones in our data- products provides similar shares. base and (b) the structure of the HS6 COMTRADE We find a lower share (in terms of export value) of classification, as textiles and clothing, essentially exported homogeneous product exports among new than among tra- by low- to middle-income countries, have a large number ditional ones (15.0% versus 31.9% using Rauch’s conserva- of lines per dollar of export. We show in the next section tive classification and 22.6% versus 39.1% according to his that the hump-shaped relationship is robust to controls for liberal classification). The reverse is true for reference- these alternative explanations and then explore characteris- priced and differentiated goods, suggesting that the bulk of tics of closed lines that may help us understand what drives diversification is made on these types of products. This fea- the hump shape. ture is emphasized by the proportion of each of Rauch’s categories in terms of the number of new lines. Differen- tiated goods account for 61.2% to 64.4% of new export A. Primary Products lines in average over the 1989–2005 period. Finally, like Besedes and Prusa (2006b) and Rauch and We consider here the prevalence of primary resources in Watson (2003), we observe that initial trade in homoge- exports as an explanation for the U-shaped pattern of export neous products requires higher values than initial trade in concentration evidenced in section II. Where do we find differentiated products. The proportion of homogeneous large primary-resource exporters along the income axis? goods in the total number of new export lines is smaller Figures 4 shows selected sectoral shares against GDP per than its proportion in the total value of these new exports capita. (7.6% versus 15% using Rauch’s conservative classification Figure 4a for minerals (HS section 5) shows a fairly dis- and 12.0% versus 22.6% according to his liberal classifica- tinct pattern whereby large exporters of mineral products tion). The contrary is true for differentiated products (those for which mineral products represent over 50% of (64.4% versus 52.5% using Rauch’s conservative classifica- exports) are either low/middle income countries or very tion and 61.2% versus 49.2% according to his liberal classi- high-income ones. This pattern, which is confirmed by the fication). nonparametric regression curve, is likely to contribute to Thus, new export products are essentially low-value-dif- the U-shaped pattern of export concentration. ferentiated goods traded by low-income countries. These As the large primary-product exporter status is a largely findings are consistent with the existing literature. Interest- time-invariant country characteristic, the country fixed- ingly, they are independent of the definition chosen. effects estimator used in section II already suggests that the U-shaped pattern of export concentration is not a spurious IV. Stages of Diversification: Alternative Explanations one due to primary product exports. However, given the importance of primary product exports in the debate linking Our decomposition of the Theil index highlights the export concentration and development, we choose to go importance of distinguishing the extensive from the inten- beyond the country-fixed-effects approach in two ways. sive margins in the evolution of export diversification. It First, we exploit the time variation in the share of pri- also suggested slow adjustment across diversification cones. mary products in exports over the 1988–2006 period by We must, however, consider alternative explanations that including this variable (in an additive way) in our usual could artificially create or reinforce a hump-shaped pattern. quadratic. We thus introduce in the model the share of HS The diversification curve may result from spurious statisti- chapters 26 (ores, slag, and ashes) and 27 (mineral fuels, cal effects, for example. Alternative explanations include mineral oils, and products of their distillation).12 Results (a) the potential role of primary resource exports as large are shown in table 5. exporters of mineral products (those for which mineral pro- 12 ducts represent over 50% of exports) are either low- or mid- Chapters 26 and 27 belong to section 5. EXPORT DIVERSIFICATION 599 FIGURES 4.—SELECTED SECTORAL SHARES AGAINST GDP PER CAPITA Author calculations using COMTRADE. Unsurprisingly, the share of raw materials comes out as a Second, we want to know if the share of raw materials positive and significant contributor to export concentration changes only the level of export concentration or if it also (this is to be expected, as a large share of one narrow class has an impact on the magnitude of the U-shape and the of products is likely to be associated with high concentra- level of the turning point. We thus interact the share of raw tion) and as a negative one to the number of active lines materials in exports with GDP per capita (columns 5 to 8 of (columns 1 to 4 of table 5). But the striking result is that table 5). Figure 5 plots predicted Theil indices against GDP coefficients on GDP per capita and its square are not per capita for various levels of raw material export shares. affected by much, nor is the turning point. Except for very high values of the share of raw materials 600 THE REVIEW OF ECONOMICS AND STATISTICS TABLE 5.—ESTIMATES WITH RAW MATERIAL EXPORT SHARES (1) (2) (3) (4) (5) (6) (7) (8) Dependent HHI Theil Gini Nber HHI Theil Gini Nber GDPpc À1.91E-05 À2.53E-04 À6.00E-06 2.66E-01 À2.80E-05 À3.23E-04 À6.73E-06 3.11E-01 16.49*** 33.46*** 27.22*** 39.98*** 19.98*** 33.96*** 23.39*** 44.87*** GDPpc2 4.15E-10 5.03E-09 1.12E-10 À4.68E-06 6.76E-10 6.76E-09 9.15E-11 À5.76E-06 12.6*** 22.60*** 13.78*** 23.02*** 16.88*** 22.87*** 9.70*** 27.31*** Raw materials 0.5142 3.4746 0.0533 À1245.7 0.3425 1.5409 À0.0013 74.01 36.39*** 45.72*** 24.96*** 16.03*** 12.57*** 15.44*** 0.88 0.78 GDPpc  Raw materials 4.01E-05 3.80E-04 7.82E-06 À2.53E-01 10.84*** 22.25*** 18.4*** 12.35*** GDPpc2  Raw materials À1.05E-09 À8.49E-09 À1.04E-10 5.50E-06 11.24*** 16.5*** 7.48*** 9.17*** Turning point ($) 23,012 25,139 26,690 28,385 – – – – Year effects Yes Yes Yes Yes Yes Yes Yes Yes Observations 2,497 2,497 2,497 2,497 2,497 2,497 2,497 2,497 Number of countries 141 141 141 141 141 141 141 141 Period 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 1988–2006 Absolute value of robust t-statistics under coefficients. ***, **, *: Significant at, respectively, 1%, 5%, and 10% level. Full sample except microstates. GDP per capita PPP in constant 2005 international dollars, from WDI. Author calculations using COMTRADE. (over 70%), the U-shaped relationship is maintained with FIGURE 5.—THEIL INDICES AGAINST GDP AND THE SHARE OF RAW MATERIALS an almost unchanged turning point. B. The Harmonized System’s Classification The harmonized system’s classification used by COM- TRADE could also potentially explain the hump-shaped relationship between economic development and export diversification. This classification is derived from nomen- clatures originally designed for tariff collection purposes rather than to generate meaningful economics. Conse- quently, some sections have a large number of economic- ally irrelevant categories (for example, the textile-clothing sector, section 11), whereas in other sections (for example, machinery, section 16), economically important categories are lumped together in a few lines. Now assume that pro- ducts in section 11 are essentially exported by middle- Author calculations using COMTRADE. income countries, whereas products in section 16 are essen- tially exported by high-income countries (assumptions con- firmed by figures 4d and 4f, respectively). Then the observed shrinks drastically, reducing the average value per line to a diversification and reconcentration pattern could be an illu- level comparable to that of other sections, as reported in fig- sion caused by the structure of the HS6 classification. ure 6b. Figure 6a, which plots, for each section of the HS6 classi- Our new classification (HS4 for sections 15, 6, or 11 and fication, total export value versus number of lines provides HS6 otherwise) contains 3,336 product lines instead of evidence of this feature. Sections 6, 11, 15, and 16 have a 4,991 for the benchmark classification. Results obtained much higher number of lines than others sectors of the HS6 with Theil indices calculated on the modified database are classification. Section 16, however, differs from sections 6, not significantly different from the ones obtained above: the 11, and 15 as it is well above the 45 degree line, reflecting a turning point is consistent with previous findings under disproportionate high value per export line, whereas sec- pooled or within estimation.13 tions 6, 11, and 15 include a large number of small lines. Figure 6 reveals that section 16 has both a large number In order to control for the conjecture that the U-shape of lines and a disproportionately high value per export line pattern of diversification may be a consequence of the struc- (the section represents around 25 percent of the total value ture of the HS6 classification, we went back to our raw of exports). The high value per export lines suggests that database and reaggregated the lines in sections 6, 11, and the number of existing lines is not extended enough to 15 from HS6 (subheading) to HS4 (heading) level (because represent production in this section in a similar way as other of its specificity, we treated section 16 separately, as we 13 explain below). The number of lines in these sectors thus Results available on request. EXPORT DIVERSIFICATION 601 FIGURE 6.—SECTION SHARES IN NUMBER OF LINES AND TRADE VALUE Author calculations using COMTRADE. sections of the HS6 classification. Mammoth lines may tially a transitory phenomenon between two steady states in indeed include many more products than lines in other sec- terms of industrial specialization. tions. This could artificially lead to the high concentration Besedes and Prusa’s (2006b) finding that the hazard rate of high-income countries. decreases rapidly in the first years of an export spell is We thus need to control for the particular design of sec- indeed suggestive of a dual regime with high infant mortal- tion 16. As we cannot further disaggregate section 16, we ity, consistent with Hausmann and Rodrik’s (2003) view of dropped this sector from the database. Our final classifica- an entrepreneurial trial-and-error process, and persistence tion thus contains 2,575 product lines. Results (not reported among ‘‘old’’ spells, consistent with the conjecture above. It here but available on request) are similar to the one is also consistent with Schott’s (2003) finding that ‘‘esti- obtained with the benchmark classification: the turning mated development paths deviate substantially from the point is robust to the aggregation of section 6, 11, or 15 and theoretical archetypes of figures 4 [a systematic pattern of the elimination of section 16 in the pooled as well as in the births for new-cone industries and deaths for old-cone within estimation. ones]. Many sectors, including apparel and footwear, exhi- The hump-shaped relationship between economic devel- bit positive value-added per worker in more than two opment and export diversification is thus not a consequence cones’’ (pp. 693–696). Apparel and footwear could indeed of spurious composition effects.14 be slowly dying industries in many countries, not only on the import-competing side but also on the export side (the EU, for instance, is still a major exporter of textile and C. Traveling across Diversification Cones apparel products). If that were the case, the high diversifica- tion characterizing the middle part of the economic devel- As Schott (2003, 2004) and Xiang (2007) discussed, opment process would not be a desirable outcome per se countries travel across diversification cones when they but simply an out-of-equilibrium one characterizing the accumulate capital. As they do, ‘‘old-cone’’ lines should transition from one steady state to another, each character- become inactive while ‘‘new-cone’’ ones should become ized by specialization according to comparative advantage. active. Suppose that old-cone lines are slow to die because A comparison of figures 4d and 4f, which show, respec- of incumbency advantages, established ties with customers, tively, the shares of textile and apparel products (section or any other kind of support they may get. During the tran- 11) and machinery (section 16) in exports as a function of sition phase, new-cone lines become active, while old-cone GDP per capita, partly bears out this story, as the former ones do not want to die. As a result, exports diversify, and follows a decreasing and only mildly convex trajectory (see the total number of active lines rises. As time passes, how- the smoother fitted curves), while the latter follows a rising ever, comparative advantage catches up on old lines, and and concave one. The combination of the two generates a they slowly die, reducing diversification. Viewed in this decrease in export concentration up to the $10,000 thresh- way, high diversification at middle-income levels is essen- old, after which there is not much more action as both tex- tiles and machinery stabilize at low (5 percent) and high 14 We also ran our baseline concentration regression with the share of (30 percent) shares, respectively. service exports in GDP on the right-hand side. Results (available on request) were unchanged: the turning point was nearly the same. We Suppose that when a country reconcentrates, export lines thank Carsten Fink for giving us the service data. that it closes are old-cone lines that were still in that coun- 602 THE REVIEW OF ECONOMICS AND STATISTICS FIGURE 7.—KERNEL DENSITY OF FACTOR-INTENSITY DISTANCES FOR CLOSED LINES The horizontal axis measures the Euclidean distance between the factor intensity of closed lines (see the text for details on the calculation) and the factor endowment of the country closing it, all for the year in which the closure takes place. Author calculations using COMTRADE. try’s export portfolio essentially by inertia. In that case, distances between their factor intensities and the endow- lines closed by a country to the right of the diversification ments of countries closing them (‘‘accidental’’ closures).15 e turning point would lie further from its comparative advan- By contrast, the density of dik for lines closed by countries tage than lines closed, in the process of normal churning, by to the right of the turning point (broken line) peaks far from countries to the left of the turning point. the vertical axis, suggesting large distances (products far This is a conjecture we can verify, albeit indirectly. To from the closing country’s current diversification cone). To do this, we use a database compiled by Cadot, Shihotori, make the argument clear, the average intensity of lines and Tumurchudur (2008). The databases contain national closed by countries to the right of the turning point is factor endowments (capital per worker and educational between the factor endowments of Chile and Malaysia, achievement) as well as revealed factor intensities calcu- whose income is about half the turning point. lated at the HS6 level as weighted averages of the factor The right panel of figure 7 provides a counterfactual. endowments of countries exporting each good. The con- Densities estimated in a similar way for new export lines struction of these revealed factor intensities follows the peak near 0, suggesting that the factor intensity of new logic of Hausmann et al.’s (2007) PRODY. That is, the export lines coincides roughly with the endowment of the revealed capital intensity of product k is countries introducing them. Moreover, there is no clear dif- X ference between the lines introduced by countries to the ^k ¼ j x j; i ik i ð11Þ right of the turning point and those introduced by countries to the left. where ji is country i’s capital/labor endowment calculated In order to go beyond descriptive statistics, we regressed according to Easterly-Levine (2001) and xik is its (Balassa) endowment-intensity distances (dik e ) on the status of coun- index of revealed comparative advantage in good k. Human tries (a dummy variable equal to 1 for countries to the right capital intensities (hi for country i) are from Barro and of the turning point and 0 otherwise) first on the subsample Lee’s (2000) national educational achievements database, of closed lines, and then on the subsample of new lines for and the revealed human capital intensity of product k is cal- the counterfactual. Table 6 presents the results (see columns culated in a way similar to equation (11). We compare the 1 and 2, respectively) and confirms the findings of figure 7. revealed factor intensity of closed line k, computed in this The coefficient on the status dummy is positive and signifi- way, with the endowment of the country closing it, using a cant for closed lines but insignificant for new lines. Euclidean distance formula: Columns 3 and 4 of table 6 show that the factor intensi- hÀ Á i1=2 ties of lines closed to the right of the turning point are not e dik ¼ hi À h ^k 2 þ ðji À j ^k Þ2 : ð12Þ 15 In order to limit the number of one year trial-and-error cases in our e If our conjecture were correct, dik should be larger for estimation, we define closed lines (in a similar way as ‘‘new export lines closed by countries to the right of the turning point lines’’) as lines that were open for two years and remained subsequently closed for two years. The kernel estimation is thus performed on lines (declining industries) than for lines closed by countries to closed between 1990 and 2003 (endowment-intensity distances are not the left of it (normal churning). The left panel of figure 7 available in the Cadot et al., 2008, database for 2004). Note that we also e run the exercise defining closed lines as lines that had been open for one shows just that pattern. The density of dik for lines closed year and remained subsequently closed for one year. Although there are by countries to the left of the diversification turning point around five times more closed lines with this definition, we observe the (solid line) peaks near the vertical axis, suggesting small same patterns as the ones described in this section. EXPORT DIVERSIFICATION 603 TABLE 6.—REGRESSION RESULTS: INTENSITY/ENDOWMENT DISTANCES ON CLOSED LINE STATUS Intensity/Endowment Difference Dependent Variable Intensity/Endowment Distance Human Capital Capital Sample: Closed Lines New Lines Closed Lines Closed Lines (1) (2) (3) (4) Status (ATP¼1) 25,171.40 465.22 4.96 108,333.00 (39.5)*** (0.8) (84.8)*** (129.8)*** Observations 31,372 98,390 31,372 31,372 R3 0.06 0.01 0.19 0.36 Estimation is by OLS; year dummies are not reported in order to save space. Absolute value of robust t-statistics under coefficients. ***, **, *: Significant at respectively 1%, 5%, and 10% level. The dependent variable in columns 1 and 2 is the Euclidean distance between the factor intensity of closed lines (see text for details on the calculation) and the factor endowment of the country closing it, all for the year in which the closure takes place. Columns 3 and 4, report the algebraic difference between the factor endowment of the closing country and the factor intensity of the closed line (for human and physical capital respectively). The status regressor is a dummy variable equal to 1 when the line is closed by a country to the right of the turning point in year t. Thus, ignoring the year dummies, column 4 says that DK ¼ À22; 909 þ 108; 333IR ; where capital is measured in 2000 PPP dollars and the status dummy IR is & 1 if country is to the right of the turning point in t IR ¼ 0 otherwise: Thus, the negative intercept means that a closed line is on average $22,909 more capital intensive than the endowment of the country closing it when it is left of the turning point, and $108,333À27,909 ¼ $80,424 less intensive to the right of the turning point. By way of comparison, France’s capital endowment (capital per worker at 2000 PPP dollars) was $139,000 in 2003. Author calculations using COMTRADE. just far from the endowments of the countries closing tors being faster to appear than old ones are to die. We find them, but also less intensive in human capital and physical evidence that countries to the right of the turning point capital. That is, in column 3, the dependent variable is close lines that are typically, in terms of factor intensities, h j dik ¼ hi À hk , and in column 4, it is dik ¼ ji À jk . The sta- far from their endowments—outliers in their export portfo- tus dummy is again positive and highly significant. lios. The hump-shaped relationship between diversification The evidence brought together in this section is only sug- and development may be explained by this slow adjustment gestive of a pattern whereby the closure of export lines in as countries travel across diversification cones. declining industries is delayed, but it certainly goes in that direction. It also confirms the prima facie evidence in Appendix B, where declining industries figure prominently among closed lines. REFERENCES Arellano, M., and S. Bond, ‘‘Some Tests of Specification for Panel Data: V. Conclusion Monte Carlo Evidence and an Application to Employment Equa- tions,’’ Review of Economic Studies 58 (1991), 277–297. The results presented so far suggest two observations. Auty, R., Resource-Based Industrialization: Sowing the Oil in Eight Developing Countries (New York: Oxford University Press, 1990). First, there seems to be, across countries and time, a robust Barro, R. J., and J.-W. Lee, ‘‘International Data on Educational Attain- hump-shaped relationship between export diversification ment: Updates and Implications,’’ CID working paper no. 42 and the level of income (the mirror image of our U-shaped (2000). Besedes, T., and T. Prusa, ‘‘Surviving the U.S. Import Market: The Role concentration indices). This nonmonotonicity holds both of Product Differentiation,’’ Journal of International Economics between and within countries. The reconcentration of 70 (2006a), 339–358. exports above a threshold around PPP $25,000 is especially ——— ‘‘Ins, Outs, and the Duration of Trade,’’ Canadian Journal of Eco- nomics 39 (2006b), 266–295. striking. Diversification occurs mostly at the extensive mar- Blundell, R., and S. Bond, ‘‘Initial Conditions and Moment Restrictions gin, especially early in the development process, as new in Dynamic Panel Data Models,’’ Journal of Econometrics 87 export items multiply and are marketed at increasingly (1998), 115–143. Brenton, P., and R. Newfarmer, ‘‘Watching More Than the Discovery large initial scales. This relationship does not appear to be Channel: Export Cycles and Diversification in Development,’’ spurious or driven only by variations in the share of primary World Bank policy research working paper, no. 4302 (2007). products. From a policy perspective, it thus appears as a Brunnschweiler, C., ‘‘Cursing the Blessings? Natural Resource Abun- dance, Institutions, and Economic Growth,’’ World Development key element of the economic development process and is, if 36 (2008), 399–419. not necessarily an objective per se, at least an important Brunnschweiler, C., and E. Bulte, ‘‘Linking Natural Resources to Slow policy indicator. From an econometric perspective, our Growth and More Conflict,’’ Science 320 (2008), 616–617. Cadot, O., M. Shirotori, and B. Tumurchudur, ‘‘Revealed Factor Intensi- findings justify treating export diversification as endogen- ties at the Product Level,’’ mimeograph, UNCTAD (2008). ous in growth regressions, as de Ferranti et al. (2002) do. de Ferranti, D., G. Perry, D. Lederman, and W. Maloney, From Natural The second observation is that diversification at middle Resources to the Knowledge Economy: Trade and Job Quality (Washington, DC: World Bank, 2002). to high levels of income may simply reflect a slow adjust- Dutt, P., I. Mihov, and T. van Zandt, ‘‘Trade Diversification and Eco- ment process between two equilibria, with new export sec- nomic Development,’’ mimeograph, INSEAD (2008). 604 THE REVIEW OF ECONOMICS AND STATISTICS Easterly, W., and R. Levine, ‘‘It’s Not Factor Accumulation: Stylized ——— ‘‘Across-Product Versus Within-Product Specialization in Inter- Facts and Growth Models,’’ World Bank Economic Review 15 national Trade,’’ Quarterly Journal of Economics 119 (2004), (2001), 177–219. 647–678. Evenett, S., and A. Venables, ‘‘Export Growth in Developing Countries: Singer, H., ‘‘US Foreign Investment in Underdeveloped Areas: The Dis- Market Entry and Bilateral Trade Flows,’’ mimeograph (2002). tribution of Gains Between Investing and Borrowing Countries,’’ Feenstra, Robert C., ‘‘U.S. Exports, 1972–1994: With State Exports and American Economic Review 40 (1950), 473–485. Other U.S. Data,’’ NBER working paper no. 5990 (1997). Theil, H., Statistical Decomposition Analysis (Amsterdam: North-Hol- Feenstra, Robert C., John Romalis, and Peter K. Schott, ‘‘U.S. Imports, land, 1972). Exports, and Tariff Data, 1989–2001,’’ NBER working paper no. Vettas, N., ‘‘Investment Dynamics in Markets with Endogenous 9387 (2002). Demand,’’ Journal of Industrial Economics 48 (2000), 189–203. Gelb, A., Windfall Gains: Blessing or Curse? (New York: Oxford Univer- Xiang, C., ‘‘Diversification Cones, Trade Costs and Factor Market sity Press, 1988). Linkages,’’ Journal of International Economics 71 (2007), 448– Griliches, Z., and J. A. Hausman, ‘‘Errors in Variables in Panel Data,’’ 466. Journal of Econometrics 31 (1986), 93–118. Hausmann, R., J. Hwang, and D. Rodrik, ‘‘What You Export Matters,’’ Journal of Economic Growth 12 (2007), 1–25. APPENDIX A Hausmann, R., B. Klinger, ‘‘Structural Transformation and Patterns of Comparative Advantage in the Product Space,’’ mimeograph, Har- vard University (2006). Data Description Hausmann, R., and D. Rodrik, ‘‘Economic Development as Self-Discov- ery,’’ Journal of Development Economics 72 (2003), 603–633. The Harmonized System’s classification of goods is defined by the Hummels, D., and P. Klenow, ‘‘The Variety and Quality of a Nation’s number of digits used, which goes from 1 (sections, numbering 21) to 2 Exports,’’ American Economic Review 95 (2005), 704–723. (chapters, numbering 99), 4 (headings, numbering 1,243), and 6 (subhead- Imbs, J., and R. Wacziarg, ‘‘Stages of Diversification,’’ American Eco- ings, numbering around 5,000). Between 1988 and 2006, there were three nomic Review 1993 (2003), 63–86. classifications: HS0-1988/1992 (5,015 products), HS1-1996 (5,111 pro- Kehoe, T. J., and K. J. Ruhl, ‘‘How Important Is the New Goods Margin ducts), and HS2-2002 (5,222 products). We convert the HS1 and HS2 in International Trade?’’ Society for Economic Dynamics meeting classifications into HS0 (using WITS conversion tables) and drop 24 HS0 papers no. 733 (2006). lines that were no longer present in the HS1 and HS2 classification. This Klinger, B., and D. Lederman, ‘‘Discovery and Development: An Empiri- yields 4,991 lines. cal Exploration of ‘New’ Products,’’ mimeograph (2004). Further degrees of disaggregation (HS 8, 10, and beyond) are not har- ——— ‘‘Diversification, Innovation, and Imitation inside the Global monized across members of the World Customs Organization and require Technology Frontier,’’ World Bank Policy Research working caution in using. For instance, Eurostat, the European Union’s statistical paper no. 3872 (2006). division, frequently reclassifies goods, shifting them back and forth Neary, R., and S. van Wijnbergen, Natural Resources and the Macroec- between different HS8 codes from one year to another. This problem also onomy (Oxford: Basil Blackwell, and Cambridge, MA: MIT Press, affects U.S. trade data compiled by Feenstra in the NBERTD (see Feen- 1986). stra, 1997, and Feenstra, Romalis, & Schott, 2002). Prebisch, R., The Economic Development of Latin America and Its Princi- COMTRADE does not always report inactive export lines as zero pal Problems (Lake Success, NY: United Nations Department of lines, as national customs often omit those lines. In a first step, we have Economic Affairs, 1950). Reprinted in Economic Bulletin for Latin thus harmonized sample size for all countries and years by adding the America 7 (1962), 11–22. missing lines and assigning them zero trade values. We thus work with Rauch, J., ‘‘Networks versus Markets in International Trade,’’ Journal of 4,991 observation per country-years. However, we do not have a perfectly International Economics 48 (1999), 7–35. balanced country-year database. Actually, for our baseline country-year Rauch, J., and J. Watson, ‘‘Starting Small in an Unfamiliar Environment,’’ regressions, we use 2,497 observations, corresponding to 141 countries International Journal of Industrial Organization 21 (2003), 1021– over 1988 to 2006 (with an average number of observations per country 1042. of 18, with a minimum of 7 and a maximum of 19). Roodman, D., ‘‘How to Do xtabond2: An Introduction to ‘Difference’ and Finally, in order to limit potential errors in reported trade flows, we ‘System’ GMM in Stata,’’ Center for Global Development working use mirrored data. Such data are more accurate than direct export data, in paper no. 103 (2006). particular for developing countries. Actually, it is well known that Sachs, J., and A. Warner, ‘‘The Big Rush, Natural Resource Booms and imports are better reported than exports. Moreover, remaining errors in Growth,’’ Journal of Development Economics 59 (1999), 43–76. reported trade flows, when using mirror data, are no more related to Schott, P., ‘‘One Size Fits All? Heckscher-Ohlin Specialization in Global exporting countries’ income levels, limiting measurement error issues in Production,’’ American Economic Review 93 (2003), 686–708. the estimation. EXPORT DIVERSIFICATION 605 APPENDIX B Closed Lines by Chapter TABLE B1.—CUMULATED CLOSED LINES, 2003–2005, BY MAIN CHAPTERS—COUNTRIES WITH GDP PER CAPITA OVER PPP$25,000 Cumulated Closed Lines 2003-2005, Country Average Number in % Number in % Value in % of of Chapter of Total Total Export Active Lines Closed Lines Value in 2002 Chapter Corresponding Section in 2000 72 Iron and Steel 15 Base Metals and Articles of 2.4% 10.8% 0.0005% Base Metal 28 Inorganic Chemicals; 6 Products of the Chemical or 3.3% 7.4% 0.0002% Organic or Inorganic Allied Industries Compounds of Precious Metals, Of Rare-earth Metals, of Radioactive Elements or of Isotopes 29 Organic Chemicals 6 Products of the Chemical or 2.8% 7.3% 0.0005% Allied Industries 41 Raw Hides and Skins (Other 8 Raw Hides and 9.4% 6.1% 0.0006% Than Furskins) and Leather Skins,Leather, Furskins and Articles Thereof; Saddlery and Harness; Travel Goods, Handbags, and Similar Containers 52 Cotton 11 Textiles and Textile Articles 3.0% 4.8% 0.0001% 25 Salt, Sulphur, Earths and 5 Mineral Products 3.5% 4.0% 0.0002% Stone; Plastering Materials, Lime and Cement 68 Articles of Stone, Plaster, 13 Articles of Stone, Plaster, 2.6% 3.3% 0.0000% Cement, Asbestos, Mica or Cement, Asbestos, Mica or Similar Materials Similar Materials 48 Paper and Paperboard; 10 Pulp of Wood or of other 1.4% 2.9% 0.0003% Articles of Paper Pulp, of Fibrous Cellulosic Material; Paper Or of Paperboard Waste and Scrap of Paper or Paperboard; Paper and Paperboard and Articles Thereof 53 Other Vegetable Textile 11 Textiles and Textile Articles 8.6% 2.8% 0.0001% Fibres; Paper Yarn and Woven Fabrics of Paper Yarn 26 Ores, Slag and Ash 5 Mineral Products 7.8% 2.7% 0.0004% 11 Products of the Milling 2 Vegetable Products 3.9% 2.6% 0.0001% Industry; Malt; Starches; Inulin; Wheat Gluten 3 Fish & Crustaceans, Molluscs 1 Live Animals; Animal 3.4% 2.5% 0.0000% & Other Aquatic Products Invertebrates 12 Oil Seeds and Oleaginous 2 Vegetable Products 4.6% 2.1% 0.0000% Fruits; Misc, Grains, Seeds & Fruit; Industrial or Medicinal Plants; Straw and Fodder 2 Meat and Edible Meat Offal 1 Live Animals; Animal 4.4% 2.1% 0.0001% Products 55 Man-made Staple Fibres 11 Textiles and Textile Articles 2.4% 2.0% 0.0001% Closed lines at date t are defined as lines with positive exports at t À 2 and t À 1 and 0 exports at t, t þ 1 and t þ 2. The sample is restricted here to countries with populations above 1 million (no microstates) and GDP per capita above PPP$25,000 (at the right of the turning point). Data are cumulated over 2003–2005 for robustness. Author calculations using COMTRADE.