The World Economy (2009) doi: 10.1111/j.1467-9701.2009.01164.x An Account of Global Intra-industry

This paper provides a comprehensive description of intra-industry trade patterns and trends, using data on more than 39 million bilateral trade flows. In 2006, 27 percent of global trade was intra-industry if measured at the finest (5-digit) level of statistical aggregation, and 44 percent if measured at a coarser (3-digit) level of statistical aggregation. The observed steady growth in global intra-industry trade since the early 1960s suggests a process of world-wide structural convergence: economies are becoming more similar over time in terms of their sectoral compositions. In particular since the 1990s, this trend appears to be driven to a significant extent by the international fragmentation of vertical production chains. Intra-industry trade is a high-income and middle-income country phenomenon: African trade remains overwhelmingly of the inter-industry type. Moreover, the observed increase in intra-industry trade was not accompanied by a comparable increase in marginal intra-industry trade, suggesting that trade-induced adjustment pressures remain potentially important.


Measurement and Data
3. Global IIT in 2006 4. The Evolution of Global IIT, 1962 5. Some Simple Regressions: IIT, Income and Distance over Four Decades 6. Marginal IIT

Non-Technical Summary
I describe global merchandise trade flows through the lens of intra-industry trade (IIT) indices, using data on more than 39 million bilateral trade flows. IIT indices quantify the extent to which bilateral imports and exports are matched within sectors. A simple description of IIT patterns is of interest for two main purposes: as a gauge of the sectoral similarity of different national economies, and as a proxy for the intensity of factor-market adjustment pressures associated with trade expansion.
It is easy to see how IIT can serve as an indicator of economic similarity: for two countries to be able to export goods of a particular sector to each other, they both need to produce this good. Given the relative paucity of internationally comparable and sectorally disaggregated production and employment data, trade-based measures can provide uniquely comprehensive (though indirect) evidence on international specialization patterns.
The link between IIT and adjustment is similarly intuitive. If tighter international trade integration leaves the sectoral composition of national economies broadly intact by fostering the two-way exchange of different "varieties" of the same type of good, then labor and capital does not have to be reallocated from declining import-competing sectors to expanding export sectors, but simply between different product lines within a given sector. It is primarily due to this "smooth-adjustment hypothesis" that the original discovery of high IIT levels among liberalizing European countries in the late 1960s generated enormous interest among policy-oriented economists and that IIT continues to be used as a diagnostic tool in impact assessments of trade reforms.

Introduction
Merchandise trade is by far the best documented aspect of international economic relations.
Trade data therefore offer a rich source of information on patterns and shifts in the allocation of economic activity around the globe.
In this paper I describe global merchandise trade flows through the lens of intra-industry trade (IIT) indices, which quantify the extent to which bilateral imports and exports are matched within sectors. A simple description of IIT patterns is of interest for two main purposes: as a gauge of the sectoral similarity of different national economies, and as a proxy for the intensity of factor-market adjustment pressures associated with trade expansion.
It is easy to see how IIT can serve as an indicator of economic similarity: for two countries to be able to export goods of a particular sector to each other, they both need to produce this good. 1 Given the relative paucity of internationally comparable and sectorally disaggregated production and employment data, trade-based measures can provide uniquely comprehensive (though indirect) evidence on international specialization patterns.
The link between IIT and adjustment is similarly intuitive. If tighter international trade integration leaves the sectoral composition of national economies broadly intact by fostering the two-way exchange of different "varieties" of the same type of good, then labor and capital does not have to be reallocated from declining import-competing sectors to expanding export sectors, but simply between different product lines within a given sector. It is primarily due to this "smooth-adjustment hypothesis" that the original discovery of high IIT levels among liberalizing European countries in the late 1960s 1 The link between export values and production values is provided by export propensities, which can vary considerably across sectors and destinations. Hence, trade values are a noisy measure of underlying production values. Trade and production specialization may even diverge. Epifani (2005), for example, develops a trade model within which increasing inter-industry specialization in production coincides with rising IIT. The present study relies on the premise that such configurations are the exception, not the rule. Moreover, actual trade data occasionally (and erroneously) report goods that merely transit a country (typically one that hosts an important port) as exports. In this case, trade flows also do not reflect production patterns. Work by Amiti and Venables (2002) and by Venables, Rice and Steward (2003) supports the interpretation of IIT that motivates this study. Venables et al. (2003), for example, conclude that their results "provide strong support for the view that the spatial pattern of IIT is merely reflecting the spatial distribution of country characteristics" (p. 2) and that "close countries do a lot of IIT because they have similar economic structures" (Abstract). generated enormous interest among policy-oriented economists and that IIT continues to be used as a diagnostic tool in impact assessments of trade reforms. 2 The paper is organized as follows. Section 2 presents the IIT measures employed and the data on which they are computed. In Section 3, I provide a snapshot of global IIT patterns in 2006, the last year for which I have data; and in Section 4 I take a longer view by describing the evolution of IIT over the full sample period . The evolution of the main cross-country determinants of IIT, based on annual regression estimates, is described in Section 5. Section 6 reports measures of marginal IIT, which are more closely related to structural adjustment than the standard IIT indices. Section 7 concludes.

Measurement and Data
The Grubel-Lloyd Index IIT is commonly understood as coterminous with the index proposed by Grubel and Lloyd (1975), which expresses IIT as a share of total bilateral trade in a particular industry i: where X cd,i and M cd,i refer to country c's exports and imports respectively, to/from where C delineates the group of countries considered. 3 Three variants of the index in (4) will be distinguished. First, for IIT within a particular country group C (say, among all low-income countries), D c ⊆ C ∀c. Conversely, for IIT between country groups (say, between low-income and high-income countries), D c ⊄ C ∀c.
Finally, country group C's total IIT (say, IIT of low-income countries with all their trading partners) obtains when D c ⊆ {C, C′} ∀c, where C′ denotes the complement to C (i.e. all trading nations that are not part of the group C).
Note that all these indices are computed for pairs of countries. It would be simple to aggregate a country's trade flows across all (or a subset) of that country's trade partners to obtain a measure of "multilateral IIT". However, most of the interest in IIT measures stems from the observation of simultaneous imports and exports between a given pair of countries, and this definition of IIT also serves best to identify similarity of trade 3 I let C symbolize both the number of countries in a particular group and the particular group (set) itself.
3 compositions among country pairs. I therefore use bilateral IIT measures as the basis for all the results reported in this paper. 4 The GL index is highly intuitive and has found near-universal acceptance. Two additional measurement issues nonetheless merit discussion.
Categorical aggregation. The definition of an "industry" is probably the most contentious issue in applied IIT research. Grubel and Lloyd (1975, p. 86) defined IIT as "trade in differentiated products which are close substitutes". Over time, it has become generally accepted that the relevant criterion is substitutability in production (rather than in consumption), since this is the aspect of industries that (a) maps trade flows to production patterns and (b) lies at the heart of the link between IIT and factor-market adjustment. 5 Whilst statistical product classifications are inevitably imperfect in this respect, they are nevertheless largely guided by the correct criterion, i.e. an effort to group together goods with similar input requirements. 6 This still leaves open the question about the most appropriate level of statistical aggregation for the calculation of IIT indices. Whilst many empirical studies use data at the 3-digit level, this choice is mostly motivated by expediency rather than any a priori reason for favoring that level of aggregation. I opt for a narrower definition in this paper, by working mainly with 5-digit sectors and thus distinguishing up to 1,161 different "industries". This minimizes the likelihood of grouping substantially different activities under the same industry heading.
Adjustment for overall trade imbalance. The upper bound of a country's mean GL index is negatively related to the size of that country's overall trade surplus or deficit relative to total trade. Hence, a larger imbalance in the trade account implies lower GL indices on average. Aquino (1978) has suggested a corresponding adjustment method for the GL 4 Through this bilateral definition, our IIT indices are conservative measures of the international fragmentation of production (also referred to as outward processing), as they will not capture sequential production chains that encompass more than two countries (see e.g. Hummels, Ishii and Yi, 2001). 5 Furthermore, it is this definition of IIT that distinguishes it from comparative-advantage based trade and that provided the impetus for economic theorists to develop the "new trade theory" (see Helpman and Krugman, 1985, for a comprehensive statement). 6 In the list of five similarity criteria used by the experts in charge of the third revision of the Standard International Trade Classification (SITC), an earlier version of which my calculations are based on, the first principle was "the nature of the merchandise and the materials used in its production", while "the uses of the product" only ranks third (United Nations, 1986, p. viii). Evidence in favor of reasonable homogeneity of statistical sectors in terms of factor requirements has been found by Elliott, Greenaway and Hine (2000).

4
index. The rationale for such an adjustment has, however, been questioned on the grounds that visible trade imbalances, both bilateral and multilateral, may well be compatible with balance of payments equilibrium (Greenaway and Milner, 1986). 7 Given the difficulty in estimating equilibrium trade imbalances, the professional consensus has been to work with unadjusted GL indices. Furthermore, if IIT measures are to be interpreted as gauges of international specialization patterns, no modification of the basic GL index is warranted. I therefore report unadjusted indices throughout.

Marginal IIT
The GL index refers to the pattern of trade in one year, and in that sense it is a static measure. This is appropriate if one seeks to quantify international specialization patterns at a particular point in time. In the context of structural adjustment, however, it is the structure of changes in trade patterns which is important. This insight has motivated the development of "dynamic" measures referred to as marginal IIT (MIIT). 8 Hamilton and Kniest (1991) first made this distinction by pointing out that the observation of a high proportion of IIT in one particular time period does not justify a priori any prediction of the likely pattern of change in trade flows. Even an observed increase in static IIT between two periods (GL t -GL t-1 > 0) could "hide" a very uneven change in trade flows, concomitant with interrather than intra-industry adjustment.
MIIT denotes parallel increases or decreases of imports and exports in an industry. Such matched changes of sectoral trade volumes can plausibly be associated with a broadly neutral effect on employment. For example, if industry i imports expand, domestic jobs may be threatened in that industry, but if industry i exports expand by a comparable 7 Egger, Egger and Greenaway (2007) propose a similar adjustment motivated by the fact that profit repatriation of multinational firms can imply inherently unbalanced bilateral trade. This is an interesting extension of IIT measurement. However, the bulk of global merchandise trade continues to be arms-length (OECD, 2002). Moreover, while multinational activity may cause bilateral imbalances at the sector level, this is not a necessary implication. 8 The GL index is calculated on the basis of cross-border flows of goods and is thus not a static measure in the strictest sense. Yet, "static" IIT in the sense of the GL index contrasts with "dynamic" measures of MIIT since the latter relate to the change in these flows between two different periods. amount, this may offset lost market share in the domestic market and yield a zero net change in the industry's domestic employment. 9 An illustration of the difference between IIT and MIIT is given in Figure 1. Figure 1A graphs a hypothetical country's bilateral imports and exports in a particular industry. All points along any ray from the origin share the same GL index, since they represent equal sectoral import-export proportions. Assume that P represents the sectoral trade balance in the base year (t-n): home-country imports exceed exports by a ratio of 3:1. The industry thus exhibits a GL index of 0.5. Assume further that the GL index is higher in the end year (t). A move from P to both Q1 and to Q2 would show up as an increase in the GL index from 0.5 to 0.8. However, the pattern of trade change is quite different between the two scenarios. With a shift from P to Q1, exports and imports of increase at the same absolute rate, and both countries (assuming there are only two) have captured an equal share of the increased volume of trade in this sector. If this pattern appears for other industries as well, then the adjustment process is intra-industry, since all countries share equally in the growth (or decline) of all these sectors. A move from P to Q2, however, implies that exports have declined while imports have increased. If this pattern appears also in other industries -with the home country not necessarily always on the 'losing' side -the adjustment process is inter-industry. A rise in the GL index can thus hide both a process of intraand interindustry trade change.
Several MIIT measures have been developed to quantify the "matchedness" of trade changes. The most straightforward of these measures is a transposition of the Grubel-Lloyd index to first differences of sectoral trade flows (country subscripts implied): where Δ stands for the difference between years t and t-T. 10 This index, like the GL index, varies between 0 and 1, where 0 indicates marginal trade in the particular industry to be completely of the inter-industry type, and 1 represents marginal trade to be entirely of the intra-industry type. 9 This conjecture evidently only holds if other relevant variables are held constant. Lovely and Nelson (2000) have shown that, in general equilibrium, MIIT can be associated with inter-industry reallocation of factors if productivity is also allowed to change. 10 See Brülhart (1994).

6
The MIIT index is related strictly to the structure of the change in trading patternsinformation on levels of exports or imports is not required. Hence, MIIT can be mapped onto a plane that is defined by ΔX and ΔM ( Figure 1B). The possibility of such a mapping is what essentially distinguishes MIIT measures from IIT.
The MIIT index shares most of the statistical properties of the GL index. 11 In particular, it can also be summed across industries, by applying the following formula for a weighted average: and where MIIT t is the weighted average of MIIT it over all sectors of the economy or over all the sub-sectors of a sector. Revision 1 has the advantage of offering maximum comparability over the sample period, as trade statistics have been recorded according to this classification since 1960. The 11 For a detailed exploration of the parallels and differences between the IIT and MIIT indices, see Oliveras and Terra (1997). 12 See Brülhart (2002) and Azhar and Elliott (2004) for discussions of the properties of this and alterative MIIT measures. Brülhart, Elliott and Lindley (2006) and Cabral and Silva (2006) are two recent empirical tests of the "smooth adjustment hypothesis" associated with MIIT. 7 disadvantage is that some sectors which are larger and more differentiated now than they were in 1960 are still recorded as a unique "industry". This will imply a tendency toward higher measured IIT in sectors that have experienced product innovation relative to sectors whose traded goods have remained unchanged. Since my main focus is on the geographic pattern of IIT rather than on sector variations, however, my priority is to obtain consistent time series by country.
Most of my calculations are performed at the 5-digit level of the SITC classification, which corresponds to the finest possible definition of an "industry" in the available data. A the 5- One approach is to narrow down the list of countries to those for which coverage is broad enough such that I can be confident that intertemporal comparisons are not driven by variations in country coverage. I have therefore established a list of 56 countries which report trade data in at least 40 of the 45 sample years, to produce an (almost) balanced panel of consistent data. 14 I refer to this as the "long coverage" data set. For this data set, I retain only data reported by the importing countries, as these can be considered to be more reliable on average (customs services having a stronger incentive to monitor imports than to monitor exports). 13 Four examples to illustrate the narrowness of the basic industry definition: in 2006, the smallest 5-digit sector was SITC 3324 ("residual fuel oils"), accounting for 0.000002 percent of the value of recorded world trade; the biggest 5-digit sector was SITC 33101 ("crude petroleum"), accounting for 9.54 percent of world trade; the median 5-digit sector was SITC 71965 ("automatic vending machines"), accounting for 0.00014 percent of world trade; and the mean 5-digit sector was SITC 03201 ("fish, prepared or preserved"), which accounted for 0.087 percent of world trade. 14 In the construction of the balanced panel, I also drop four of the 1,161 5-digit sectors for which COMTRADE does not provide consistent coverage over the sample period. Appendix Table 1 lists the 56 countries included in the "long coverage" data set.
8 As a second approach, I exploit the fact that country coverage is broader if one takes account of reported export data as well as of reported imports. One can take exporting country statistics to infer imports of countries that have not submitted their statistics to the UN. I therefore use exporter data to fill as many gaps as possible for four sample years: 1962, 1975, 1990 and 2006. Since the non-reporting countries are mainly from the developing world, this "wide coverage" data set allows me to incorporate many lowincome countries into the analysis that are not part of the "long coverage" sample. 15 At the 5-digit level, the "long coverage" data set identifies between 565,000 (1962) and It becomes immediately apparent that IIT at the 3-digit level is higher than IIT at the 5-digit level. The unweighted IIT averages are 0.14 at the 3-digit level and 0.07 at the 5-digit level (see final row of Table 1). Table 1 also clearly shows that large trading nations tend to exhibit higher IIT, which explains why these unweighted averages are significantly smaller than the aggregate IIT shares reported above. It suffices to look at the third and fourth data columns to realize that GL indices tend to increase with the size of countries' trade. The simple correlation coefficients between trade shares and GL indices are 0.58 (3-digit) and 0.52 (5-digit).
Furthermore, the second data column of Table 1 shows that larger trading countries also tend to trade in a broader set of industries. France is the country with the highest level of While average IIT shares differ significantly, variations across countries are very similar for the two levels of sectoral aggregation: the correlation coefficient across the 216 countries between the 3-digit and the 5-digit GL indices is 0.97.
In Table 2, I slice the global trade matrix by sector rather than country, and I present trade shares as well 5-digit and 3-digit GL indices separately for the 177 3-digit sectors. Again one can easily observe that 3-digit GL indices are higher than 5-digit GL indices (aggregated to the 3-digit level), the unweighted averages corresponding to 0.28 and 0.21 respectively. And at 0.92, the correlation between the two sets of GL indices is again very high. Sectoral disaggregation thus strongly affects observed average levels of IIT, but it is of secondary importance in a description of broad cross-sectional patterns.
The 3-digit sector with the highest level of observed 5-digit IIT (GL= 0.527) is "Electric Power Machinery and Switchgear", whereas the only 3-digit sector for which I find a 5digit GL index of 0.000 is "Concentrated Uranium and Thorium Ore". categorization and applying the "within" version of the group-level GL index defined in expression (4). Trade among high-income countries is characterized by the highest IIT shares on average. IIT among the low-income countries, in contrast, is virtually nonexistent. Strikingly, however, the highest 5-digit IIT level is observed for trade among lower-middle-income countries -higher even than for trade among high-income economies. There are good reasons to believe that the high IIT among lower-middleincome countries is due to processing trade in vertically fragmented industries (the four main trading nations in this category are China, Thailand, the Philippines and Indonesia, see Table 1).
Finally, Figure 3 reports summary IIT according to a classification of 5-digit sectors by the three main stages of the production chain: primary, intermediate and final goods. 18 Not surprisingly, primary goods are found to exhibit by far the lowest average IIT. It is interesting, however, to observe that average IIT in intermediate goods is considerably higher than IIT in final goods. This again suggests that vertical fragmentation of production processes across country borders might be as important (or even more important) in explaining global IIT patterns as international product differentiation and consumer tastes for variety.

The Evolution of Global IIT, 1962-2006
Aggregate IIT I now turn to the description of changes in IIT over time, based on the "wide coverage" sample, which offers comparable data over the full sample period. Figure 4 provides the main picture. It shows how, irrespective of the level of categorical aggregation, global IIT has exhibited a secular upward trend that has leveled out in the mid-1990s. 19 In this narrower country sample, more than a third of global trade is now IIT if measured at the 5digit level, and more than half if measured at the 3-digit level. The upward trend in IIT 18 The classification at the 5-digit level is taken from the United Nations' Broad Economic Categories, concorded to the SITC, Rev. 1. Table 2 shows this grouping at the 3-digit level. The full (5-digit) classification can be provided on request. 19 Measured IIT in 2004 and 2005 is somewhat biased downward due to the fact that in those years COMTRADE data attribute a significant share of EU imports to the EU as a whole rather than to the individual destination countries. This reduces observed import volumes of EU member states in those two years.
suggests a process of world-wide structural convergence: economies are becoming more similar over time in terms of their sectoral compositions.
As a complement to the time series of Figure 4, which is based on data for the 46 predominantly higher-income countries for which consistent import data are available, I show aggregate IIT levels for the "wide coverage" data set in Figure 5. It is unsurprising that IIT shares are lower in Figure 5 than in Figure 4, as the latter omits most low-income countries. Nonetheless, the broadly increasing share of IIT in world trade is as evident in  Again, it becomes apparent that the rise in IIT has been a very general phenomenon, as it is observed for all three product groups. Primary products, not surprisingly, have consistently exhibited the lowest IIT shares and also recorded the slowest increase. Average IIT levels   Figure 15 illustrates the evolution of IIT within and between country income groups.
Because the poorest countries are underrepresented in the "long coverage" data set (see Appendix Table 1), I combine the World Bank's "low income" and "lower middle income" categories into a single "low" group. Again a positive correlation between income levels and IIT is clearly apparent, with IIT among high-income countries far outstripping IIT among all other country groups. There has, however, been some marked convergence in global IIT patterns, with IIT shares among all country groups trending upwards since around 1980, and IIT shares involving middle-income and low-income countries growing more rapidly than IIT among high-income countries.
One conspicuous pattern in Figure 15 is a leveling-off in all IIT series, coinciding roughly with the turn of the millennium. A similar, though less pronounced, trend break is also visible in the aggregate IIT time paths shown in Figure 3. Figure 15 shows that the recent stagnation in aggregate IIT growth is not due to the increased integration into world trade of emerging economies and an associated inter-industry "re-specialization", because all country groups exhibit slowdowns. 20 One possibility is that IIT has leveled of because of the recent increase in the share of primary goods in the value of world trade. Only some 6 percent of global trade in primary goods were intra-industry in 2006 (see Figure 2).
Being based on the "long coverage" sample, Figure 15 offers a continuous time series, but it does not take account of most of the world's poorest countries. Figures 16 and 17, being based on the "wide coverage" data set, address this issue. The exclusion from global IIT by the poorest countries emerges starkly from Figure 16. Among countries categorized as "low income" by the World Bank, the intra-group IIT share has remained stuck below a derisory 0.5 percent since 1962. The convergence in global IIT levels is clearly a middle-income country phenomenon. The surge in IIT among the lower middle income countries from 2.2 percent in 1990 to 13.9 percent in 2006 is particularly striking.
The polarized global geography of IIT is also apparent in Figure 17, where I report the evolution of IIT between income groups: everybody's average IIT is highest with the highincome countries and lowest with the low-income countries.

IIT and Regional Integration
In light of the ongoing proliferation of regional integration agreements (RIAs), I report some relevant evidence for the EU and for four Sub-Saharan African RIAs. of African RIAs having stimulated neither substantial regional trade nor structural convergence.

Some Simple Regressions: IIT, Income and Distance over Four Decades
As a complement to the descriptive statistics that represent the main contribution of this paper, I report some simple regression results to quantify the sensitivity of IIT to bilateral distance as well as its relation to per-capita income levels.
where GL cd is the aggregate bilateral GL index between countries c and d as defined in (2), pcGDP is per-capita GDP, dist is the geodesic distance between the two countries' main cities and contig is a dummy variable set to one for countries that share a common land border. The dependent variable is a log transformation of the GL index, which centers it symmetrically around zero and makes it unbounded. 21 Specification (7) contains the main variables featuring in most cross-country IIT regressions: the joint income level of the country pair, which is commonly associated with high IIT; the difference in income levels, which is associated with low IIT; and distance measures, which are also associated with low IIT. 22 while there are instances of statistically significant positive as well as negative coefficients, the large majority of estimates are not statistically significantly different from zero.
The main output from this exercise is Figure 24, which traces the annual estimated coefficients on distance and on average GDP per capita over the sample period. Two 21 In order not to lose bilateral observations with no IIT, I have set GL cd = 0.0001 for all country pairs with zero recorded IIT, this number being slightly lower than the smallest observed non-zero bilateral GL index. The qualitative results are fairly robust to the particular choice of this number. 22 See, e.g., Hummels and Levinsohn (1995), and Bergstrand and Egger (2006

Marginal IIT
Figures 25 to 29 illustrate the broad patterns of global MIIT, computed using definitions (5) and (6), and Table 4     Upper Middle Inc.
Lower Middle Inc.

Low Income
High Income GL, 5-digit, total trade GL, 5-digit, intra-region trade Notes: Country grouping according to World Bank categorization (see Table 1); "wide coverage" data set

Hi-Hi Low-Hi Low-Low Low-Med Med-Hi Med-Med
Notes: Country grouping according to World Bank categorization (see Table 1, "Low" category is combination of LIC and LMC); "long coverage" data set; data converted into constant prices using US GDP deflator; base and end periods are averages of three adjacent years .06 .08 .1 Hi-Hi Low-Hi Low-Low Low-Med Med-Hi Med-Med Notes: Country grouping according to World Bank categorization (see Table 1, "Low" category = LIC + LMC); product grouping according to United Nations "Broad Economic Categories"; "long coverage" data set; data converted into constant prices using US GDP deflator; base and end periods are averages of three adjacent years Hi-Hi Low-Hi Low-Low Low-Med Med-Hi Med-Med Notes: Country grouping according to World Bank categorization (see Table 1, "Low" category = LIC + LMC); product grouping according to United Nations "Broad Economic Categories"; "long coverage" data set; data converted into constant prices using US GDP deflator; base and end periods are averages of three adjacent years Hi-Hi Low-Hi Low-Low Low-Med Med-Hi Med-Med Notes: Country grouping according to World Bank categorization (see Table 1, "Low" category = LIC + LMC); product grouping according to United Nations "Broad Economic Categories"; "long coverage" data set; data converted into constant prices using US GDP deflator; base and end periods are averages of three adjacent years 35