Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
J. Int. Trade & Economic Development 13:4 397–418 Does convergence cause trade, or does trade cause convergence? Teresa Cyrus Department of Economics, Dalhousie University, Halifax, Nova Scotia, Canada Abstract This paper examines the direction of causality between international trade and cross-country income differences in several ways. First, instruments for income are used in pooled gravity regressions to determine the effect of income differences on bilateral trade, and instruments for trade are used in regressions to determine the causes of income dispersion. Results of these cross-country estimations show that more similar countries trade more, while trade appears to increase dispersion. However, fixed-effects regression, random-effects regression, and Granger causality tests show that trade reduces income differences over time. Thus, while the postwar era has seen increasing trade and conditional convergence, the causality is bi-directional: convergence causes trade, and trade causes convergence. Keywords Trade, convergence, income, gravity, causality, instruments. 1. INTRODUCTION Over the postwar period, the world has become increasingly integrated through international trade. At the same time, the countries of the world have experienced convergence in income per capita, holding capital accumulation and population growth constant. Has convergence caused trade to increase? Has trade caused convergence? Is some other factor causing both? This paper explores the direction of causality between trade and income differences. From 1965 to 2000, the average share of trade (exports plus imports) in income has risen fairly consistently for the 56 countries considered in this paper, from 43.94 per cent in 1965 to 81.18 per cent in 2000.1 For the same group of countries, convergence conditional on country-specific attributes (saving rates and population growth rates) has occurred.2 Address for Correspondence Teresa Cyrus, Department of Economics, Dalhousie University, Halifax, Nova Scotia, Canada, B3H 3J5. E-mail: [email protected] The Journal of International Trade & Economic Development ISSN 0963-8199 print/ISSN 1469-9559 online ª 2004 Taylor & Francis Ltd http://www.tandf.co.uk/journals DOI: 10.1080/0963819042000300573 398 The Journal of International Trade & Economic Development How might convergence affect trade? Both old- and new-style trade theories give reasons for links from income to trade. Heckscher – Ohlin – Samuelson trade theory states that trade is based on differences in factor proportions; differences in capital – labour ratios will increase trade. By contrast, the new trade theory, due predominantly to Krugman (1979) and Helpman (1981), based on monopolistic competition and economies of scale, implies that similarities in income between countries will increase trade. How might trade affect convergence? The literature on knowledge dissemination and technology transfer shows how trade may lessen income differences between countries.3 One possibility, illustrated in Grossman and Helpman (1991), is that goods may have ideas or knowledge embedded in them. Trade allows these ideas to flow across borders, thereby allowing follower countries to catch up to those more advanced. Another possibility is that exposure to international trade may increase competitive pressures, which force domestic firms to increase their productivity, resulting in higher income. In contrast, other models show that trade may in fact increase income differences between countries. For example, as in Young (1991), poor countries may become trapped in producing goods that offer little scope for technology growth or development. Thus, the trade literature provides reasons why income differences will affect trade, but the sign of the effect is unclear. Similarly, the convergence literature provides reasons why trade will affect income differences, but it is not clear whether trade causes income differences to rise or to fall. Obviously, there is a question of causality here. If we believe that income dispersion affects trade, and that trade affects dispersion, then it may be difficult to know the true sign and magnitude of these effects. A few others have tried empirically to link trade and income convergence, but have not established causality. Ben-David (1993) finds a relationship between the timing of trade liberalization and income convergence in the European Community. Helliwell and Chung (1990) examine 19 industrial countries, and find evidence that countries that have increased their openness to international trade experience faster convergence. Helpman (1987) shows that, for a sample of 14 OECD countries over the postwar period, dispersion has fallen as the trade – income ratio within the group has risen. In contrast, Slaughter (2001) uses a differences-in-differences approach to test whether trade-liberalizing countries experience a greater change in convergence than countries that have not liberalized; he finds no strong evidence that this is the case. None of these papers addresses causality, however. This paper uses data on bilateral trade, while most authors who examine trade use the ratio of exports or total trade to GDP. Using Does convergence cause trade or does trade cause convergence? 399 bilateral trade provides for a richer dataset; it allows for many more observations to be used in estimation and, more importantly, if we believe that convergence happens because countries learn from their trading partners, then including information on those trading partners is essential. It should be noted that the word ‘convergence’ here refers to a fall in cross-country income differences, not beta or sigma convergence as described by, for example, Barro and Sala-i-Martin (1991). I address the dilemma of causality in three ways. First, I estimate two equations, one for trade and one for income differences, using instruments for these two variables. In the bilateral trade equation, the instruments I use for per-capita income differences include differences in initial income per capita and in factor accumulation rates (physical and human capital accumulation rates and the population growth rate). In the income-differences equation, the instruments I use for bilateral trade are distance and adjacency, which are standard gravity variables;4 dummy variables representing whether the two countries speak the same language or belong to the same trading bloc; and the sum of the two countries’ populations. Second, similar regressions will be run using the fixed-effects methodology. This has the advantage of taking into account country-specific fixed effects. However, since many of the variables do not vary over time, they cannot be included in these regressions, so the equations differ somewhat from the OLS and instrumental variables equations. While the randomeffects estimator is most likely inconsistent due to correlation between the explanatory variables and the unobserved effects, the results of randomeffects regressions are reported for the sake of comparison. Finally, Granger causality tests will be performed to determine whether past values of income differences are useful in predicting current trade flows, and whether past values of bilateral trade are useful in predicting current levels of income differences. This paper contributes to this literature by seeking to determine the direction of causality between income convergence and trade. The results robustly show that income differences lower trade. However, the results are mixed as to whether trade raises or lowers income differences; crosssectional instrumental variables regressions show that trade raises differences, but the time-series results show the opposite. At any given point in time then, increased trade is associated with higher income differences, but over time, trade causes income differences to contract. In Section 2, the theoretical links between trade and income differences are presented. Section 3 presents the equations to be estimated and describes the data. The empirical results are presented in Section 4, and Section 5 considers robustness and specification tests. Section 6 concludes. 400 The Journal of International Trade & Economic Development 2. THEORETICAL CONSIDERATIONS 2.1 The impact of income differences on trade How do income differences affect trade? The factor-proportions theory originated by Heckscher (1919) and Ohlin (1924), and refined by Samuelson (1948, 1949) and Jones (1956), among others, makes predictions different from the new trade theory of Krugman (1979) and Helpman (1981). The seminal work of Heckscher and Ohlin, like the earlier theory of Ricardo, states that differences between countries drive trade. The factorproportions theory links a country’s relative factor endowments and the factor content of its trade flows, stating that a country will export the commodities that abundantly use its intensive factors; countries will engage in trade only if their capital – labour ratios differ. Since there is a high correlation between capital – labor ratios and output per capita,5 the factorproportions theory then implies that countries will engage in trade only if their levels of output per person differ. The factor-proportions theory does not explain trade between similar countries, and yet, in the postwar era, a large proportion of trade has been between industrial countries. In addition, much trade has been in differentiated products. Krugman (1979, 1980, 1981) and Helpman (1981) explain these empirical regularities with models of monopolistic competition, economies of scale, and intra-industry trade, showing that even countries with identical factor endowments will engage in international trade. In the factor-proportions model, of course, identical factor endowments in two countries imply no bilateral trade whatsoever. Both Krugman and Helpman mention the similarity of their conclusions to those of Linder (1961), who posits that countries that have more similar demand structures will trade more. Linder points out that the level of average income is the most important factor affecting the demand structure, stating that ‘per-capita income differences are a potential obstacle to trade’ (p. 98); thus, countries with more similar levels of income per capita should trade more.6 Including per-capita differences in income as an explanatory variable in a gravity regression will allow us to see the impact of dispersion on trade. If differences increase trade, then the factor-proportions theory is sufficient to explain trade. If, on the other hand, differences reduce trade, then the new trade theory is supported. 2.2 The impact of trade on income differences Trade may affect the differences in per-capita income between countries, but it is not clear whether trade causes dispersion to fall over time, or to rise. While many recent papers have explored the theoretical relationship Does convergence cause trade or does trade cause convergence? 401 between trade and convergence, links from trade to both convergence and divergence have been found. Those who claim that trade lessens income differences often assert a link through technology. This is due, for example, to international transfers in knowledge, where knowledge is embedded in goods. Countries at a lower level of development may be able to increase their productivity by importing goods from richer countries and making use of the technology and ideas contained therein. Poor countries can thus close the gap between their low levels of technology and the higher levels in other countries. Grossman and Helpman (1991), in an analysis similar to that of RiveraBatiz and Romer (1991), assert that trade allows countries to learn about innovation in products and techniques. They produce a theory in which knowledge diffusion, once channels of communication have been opened between countries, increases the rates of innovation and growth in each country. When they allow for the integration of product markets, they find that trade has both level and growth effects through ‘the elimination of duplicative research’ (p. 245). Ben-David and Loewy (1996) construct a model in which technology grows at a fixed rate in a closed economy but, in an open economy, grows instead at an endogenous rate that depends on trade. Their analysis is focused on the trade liberalization process, but Ben-David (1996b) amends the Ben-David – Loewy model to focus on bilateral trade volumes. BenDavid (1996a and 1996b) find convergence in TFP and output per worker when countries are grouped together based on trade flows; random groupings show no convergence. Another group of papers finds that trade may cause income differences to rise, not fall, over time. While Grossman and Helpman (1991) find that, in general, growth rates will be faster as a result of trade, there are three cases in which trade may increase differences between countries by dampening incentives for research. First, a small country facing increased competition from abroad may find that the profitability of investments in knowledge has fallen. Second, a country that begins with a disadvantage in research productivity and then enters into international competition with a more advanced country will find a slowing of innovation and growth. Third, a country highly endowed with unskilled labour may specialize in traditional manufacturing and may experience a fall in its overall growth rate of manufactured output. Grossman and Helpman (1995), Helpman (1984), and Ethier (1982) show that, if there are increasing returns to scale on the national level, a country that specializes in the constant-returns-to-scale sector may lose from trade. Feenstra (1990) presents a model with two countries of different sizes and no knowledge spillovers, in which trade causes a slowing in the rates of product creation and of growth in the smaller country. In Young’s (1991) model of learning by doing, poorer countries are forced by competition with richer 402 The Journal of International Trade & Economic Development countries into producing goods in which learning by doing has been exhausted; thus, poorer countries will experience slower rates of technical progress and GDP under free trade than under autarchy. Including trade as an explanatory variable for income differences will allow the impact of trade on income differences to be seen. If the knowledge diffusion that accompanies trade in goods pulls poorer countries toward richer countries, then we will see that trade lowers dispersion. On the other hand, if poorer countries find themselves worse off as a result of trade, then we will find that trade increases dispersion. 3. SPECIFICATIONS AND DATA 3.1 Trade and income differences equations To determine the impact of income differences on trade, and the impact of trade on income differences, two equations will be estimated: a gravity regression that allows differences in income per capita to affect bilateral trade, and a dispersion regression that allows trade to affect differences in income per capita. The bilateral trade equation is as follows: ln Tradeij ¼ b0 þ b1 ðln GDPi þ ln GDPj Þ þ b2 jðln GDPi =Popi -ln GDPj =Popj Þj þ b3 ðln Distanceij Þ þ b7 ðln Remotenessi þ ln Remotenessj Þ þ b4 ðAdjacencyij Þ þ b5 ðLanguageij Þ þ b6 ðTrading Blocij Þ þ eij ð1Þ The dependent variable is the log of bilateral trade (exports plus imports). The independent variables are the sum of the log of country i’s total GDP and the log of country j’s total GDP; the absolute value of the difference between the log of country i’s income per capita and the log of country j’s income per capita; the log of the great-circle distance between the two countries; the sum of the log of the two countries’ remoteness terms, which, for each country, is a GDP-weighted average of its distance from all trading partners; a dummy variable for whether the two countries share a common border; a dummy variable representing whether the same language is spoken in the two countries;7 and a dummy variable for whether the two countries are members of the same trading bloc.8 Note that if b2 is negative, then the new trade theory is supported, since countries with more different incomes trade less than countries with more similar incomes. The factor-proportions model would predict the opposite, that trade is based on differences, not similarities; in this case, b2 would be positive. The presence of ln GDPi + ln GDPj in the regression is intended to normalize for size. The concern here is not with the effects of being a large or a small country per se, but instead to isolate the effects of differences or Does convergence cause trade or does trade cause convergence? 403 similarities in per-capita GDP. Obviously, holding distance constant, a hypothetical country’s trade with a smaller country like Belgium will be less than its trade with a larger country like Germany, simply because Belgium has a lower GDP and thus has less to trade. The income differences equation is as follows: jðln GDPi =Popi ln GDPj =Popj Þj ¼ b0 þ b1 ðln Tradeij Þ þ b2 jðln GDP60;i =Pop60;i -ln GDP60;j =Pop60;j Þj þ b3 jðln ski -ln skj Þj þ b4 jðln shi -ln shj Þj ð2Þ þ b5 j½ln ðni þ g þ dÞ-ln ðnj þ g þ dÞj þ mij Here, the dependent variable is the absolute value of the difference between the log of country i’s income per capita and the log of country j’s income per capita. The independent variables are the log of bilateral trade; the absolute value of the difference in the two countries’ log GDPs per capita in 1960; the absolute value of the difference in their log investment shares (sk); the absolute value of the difference in their log secondary-school enrolment ratios (sh); and the absolute value of the difference in the log growth in their populations (n, plus 0.05, to account for technological progress and depreciation). Note that if b1 is negative, then trade causes convergence (by lowering income differences); if b1 is positive, then trade causes divergence. Data on bilateral trade, income, and the other variables are available for 1965, 1970, 1975, 1980, 1985, 1990, 1995 and 2000, for the 56 countries listed in Appendix B. Data sources are found in Appendix A. 3.2 Ordinary least squares and instrumental variables The first test of the trade – income differences relationship involves a careful estimation of equations (1) and (2). Uncertainty as to the direction of causality between trade and income differences tells us that ordinary least squares is not the correct estimator. For example, if higher trade causes smaller income differences, then income differences may be significant in a trade regression, even if this is not the correct direction of causality. More generally, it may be that a third factor – similar free-market government policies in two countries, for example – causes both high bilateral trade and low income differences. To solve this problem, instrumental variables will be used for both income differences and for trade. These two equations represent a system, so that the variables in the trade regression other than income may be used as instruments for trade in the income – differences equation, and the variables in the income – differences regression other than trade may be used as instruments for income in the trade equation. 404 The Journal of International Trade & Economic Development Good instruments for income differences in a trade regression must be correlated with dispersion and uncorrelated with the error term. The instruments used here are the factor accumulation variables from Mankiw et al. (1992). Since these variables are instrumenting for income differences, the form of the instruments used here will be the absolute value of the difference in each variable between country i and country j. In their augmented Solow model, Mankiw et al. show that factor accumulation rates over the period 1960 to 1985 do a good job of explaining GDP per worker in 1985. The equation estimated is: ln yi ¼ b0 þ b1 ðln ski Þ þ b2 ðln shi Þ þ b3 ½ln ðn1 þ g þ dÞ þ eij ð3Þ where sk is the physical capital accumulation rate (the average share of investment in GDP), sh is the human capital accumulation rate (the average share of the working-age population in secondary school), and n is the population growth rate (the growth rate of the working-age population); and g + d accounts for (worldwide) technological progress and depreciation, set equal to 0.05. Mankiw et al. show that factor accumulation rates are highly correlated with income, but are they uncorrelated with the error term in a trade equation? To be correlated with the error term, these variables must affect trade in some way other than through income (or through the other righthand side variables). It is difficult to imagine how differences in the share of investment in GDP, or in the secondary-school enrolment ratio, or in the growth of the working-age population, could affect trade other than by affecting income differences. Good instruments for trade in an income-differences regression must be correlated with trade and uncorrelated with the error term. The instruments used here include geographic, cultural, and size variables. The geographic variables, the log of distance (measured as the great-circle distance between main cities) and a dummy for adjacency, have been shown in standard gravity regressions to be highly correlated with trade.9 The cultural variables include dummy variables representing whether the two countries speak the same language, and whether the two countries belong to the same trading bloc. All of these variables are highly correlated with trade; however, it is possible to argue that they are not uncorrelated with the error term in an income-differences regression. For example, while adjacent countries undoubtedly trade more, they also probably have similar incomes. Similarly, countries that speak the same language may share a common colonial history, and may thus have similar incomes. The correlation of these variables with the error term in an income-differences regression is not likely to be large, however. In fact, regressing the residual of the income- Does convergence cause trade or does trade cause convergence? 405 dispersion equation on these variables shows that they are not significant in explaining the residual (this is an over-identifying restrictions test, which will be described in more detail later). Finally, the size variable is the sum of the two countries’ log populations. This is intended as an instrument for the sum of the countries’ total GDPs, since the sum of total GDPs is plagued by the same endogeneity problems as the difference in per-capita GDPs. It is likely that population will affect income differences only by affecting trade. 3.3 Fixed and random effects Since the available data constitute a panel across countries and over time, panel data techniques can also be used to examine the relationship between bilateral trade and income differences. However, since the fixed-effects technique does not allow for any time-invariant variables to be included, many variables must be left out of such a regression: distance, remoteness, adjacency, common language, and common trading bloc in the trade equation; and initial per-capita income differences in the income equation. The fixed effects regressions thus do not exactly test the joint trade – income differences relationship as presented here. However, this technique does allow for country-pair heterogeneity and gives each country-pair its own intercept; the removal of these country-specific fixed effects from consideration allows for a close look at the relationship between trade and convergence. It is possible to use instrumental variables in a fixed-effects regression, but again, the problem here is that many of the planned instruments do not vary over time and so cannot be used. A possible solution is to use a lagged value of the endogenous variable as its instrument. Thus, in the fixed-effects trade regression, lagged income differences are used as an instrument for current income differences; in the fixed-effects income differences regression, lagged trade is used as an instrument for current trade. A Hausman test for equality of the fixed effects and random effects coefficients shows that a random-effects technique produces inconsistent estimates; still, the random-effects coefficients are reported for the sake of comparison. The advantage to using the random-effects technique is that both time-series and cross-sectional variation are used, so time-invariant variables can be included. 3.4 Causality tests Another way to examine this relationship is to use Granger causality tests. One variable can be said to Granger-cause another if the first variable is useful in predicting future values of the second variable. Like the fixed- and random-effects regressions, this kind of test makes good use of the time- 406 The Journal of International Trade & Economic Development series properties of this dataset. For the trade equation, a regression can be run of current trade on past trade and past income differences. If the lagged income differences are significant in explaining current trade, then income differences Granger-cause trade. Similarly, if lagged trade is significant in explaining current income differences, then trade Granger-causes income differences. 4. EMPIRICAL RESULTS 4.1 The impact of income differences on trade Table 1 presents the effect of income differences on bilateral trade. Data from 1965 to 2000, at five-year intervals, are used.10 Dummy variables representing each year were included, but are not reported in order to save space. Columns 1 and 2 present the results for pooled ordinary least squares (OLS) and instrumental variables (IV) regressions; the fixed-effects (FE) and FE-IV results are in columns 3 and 4; and the random-effects (RE) results are in column 5. The OLS and IV results show that total GDP, differences in GDP per capita, distance, remoteness, a common border, a common language, and membership in the same trading bloc are highly significant in explaining bilateral trade. The instrumental variables equation shows that a 1 per cent increase in the difference of two countries’ GDP per capita lowers trade by 0.15 per cent. Of course, a change in a country’s GDP affects both the sum of total GDPs as well as the difference in per-capita GDPs. The total effect on trade depends on whether the richer or poorer country (as measured by per-capita GDP) experiences a rise in income. If the richer country’s GDP increases by 1 per cent, then trade rises by just 0.43 per cent; the increase in GDP raises trade by 0.58 per cent, reflecting the fact that the country is now producing more goods to trade, but lowers trade by 0.15 per cent, since the country’s income per capita becomes more different from that of its trading partners. However, if it is the poorer country that receives a 1 per cent boost to GDP, trade rises by 0.73 per cent (0.58 per cent + 0.15 per cent); the country has more goods to trade, and its GDP per capita is now more similar to that of its trading partner. All of the other explanatory variables are highly significant and have the expected signs. Distance and remoteness enter negatively; adjacency, sharing a common language, and membership in a common trading bloc have a positive effect on bilateral trade. For the last three variables, which are dummy variables, the coefficients should be exponentiated; for example, belonging to the same trading bloc more than doubles bilateral trade, since e0.79 = 2.20. The fixed effects regression, in column 3, includes only the sum of the two countries’ total incomes and the difference in their per-capita incomes as explanatory variables (as well as year dummy variables, which are not 0.78** (0.0077) 7 0.22** (0.019) 7 0.61** (0.027) 7 0.78** (0.061) 0.061 (0.084) 0.60** (0.039) 0.73** (0.059) 7 6.44** (1.12) 0.68 9642 0.58** (0.010) 7 0.15** (0.023) 7 0.49** (0.028) 7 1.20** (0.065) 0.33** (0.084) 0.69** (0.041) 0.79** (0.062) 7.25** (1.23) 0.66 9642 IV 7 32.70** (0.93) 0.78 9642 0.99** (0.025) 7 0.48** (0.029) FE 7 32.66** (1.12) 0.75 7499 0.98** (0.030) 7 0.48** (0.041) FE-IV 0.87** (0.015) 7 0.40** (0.025) 7 0.67** (0.060) 7 0.86** (0.13) 7 0.096 (0.20) 0.69** (0.086) 0.63** (0.14) 7 7.78** (2.40) 0.68 9642 RE Notes: The dependent variable is the log of bilateral trade in 1965, 1970, 1975, 1980, 1985, 1990, 1995, and 2000. OLS refers to ordinary least squares, IV to instrumental variables, FE to fixed effects, and RE to random effects. Time dummy variables are included but not reported here. Standard errors are in parentheses (robust standard errors for the OLS and IV equations). **, * and # denote significance at the 99%, 95%, and 90% levels, respectively. The R2 reported is the adjusted R2 for the OLS and IV equations, the ‘within’ R2 for the FE regression, and the ‘overall’ R2 for the RE regression. The instruments are described in the text. R2 N Constant Trading Blocij Languageij ln Remotenessi + ln Remotenessj Adjacencyi ln Distanceij jln GDPi/Popi - ln GDPj/Popjj ln GDPi + ln GDPj OLS Table 1 The impact of income differences on trade Does convergence cause trade or does trade cause convergence? 407 408 The Journal of International Trade & Economic Development reported), since the other gravity variables do not change over time. Here, the coefficients become larger in size but remain strongly significant. The random-effects coefficients, in column 4, are also similar, although adjacency is no longer significant. It should be noted that, in all cases, the coefficient on per-capita income differences is negative and significantly different from zero. Similarities, not differences, increase trade; convergence, then, does cause trade. Results specific to OECD countries are found in Table 2. For the OLS and IV equations, I include a dummy variable indicating that both of the trading countries are members of the OECD. To allow the coefficients to vary for OECD countries, I also include interaction terms of all of the variables with the OECD dummy. The first two columns show the OLS results; the second two show the IV results. For each specification, the first column shows the coefficients and standard errors on the variable other than the OECD dummy and its interactions; these are shown in the second column. The effects of each variable on trade between OECD countries can thus be seen as the sum of the usual coefficient and the OECD interaction coefficient. For countries in the OECD, total GDP has a smaller effect on trade than for non-OECD countries, but it is still significant. Interestingly, differences in GDP per capita now enter positively for non-OECD countries, but the coefficient is insignificant; however, income differences for OECD countries become even more important. In fact, the magnitude of the coefficient on the difference in per-capita income exceeds that of the coefficient on total income. Thus, a 1 per cent increase in the richer country’s GDP actually causes trade to fall, not rise, by 0.24 per cent (0.30 per cent - 0.54 per cent); an increase in the poorer country’s GDP causes trade to rise by 0.84 per cent. The effects of distance, adjacency, and speaking the same language are stronger in the OECD, as evidenced by coefficients on the interaction terms with the same positive sign as the overall coefficients. However, remoteness now raises trade for OECD countries, and the effect of belonging to the same trading bloc is now very small. This may be due to the fact that many of the 22 OECD countries in this sample belong to such a bloc, and thus there is little variation in the sample. The final column shows the fixed-effects results. Using an OECD dummy, which remains constant over time, would not be possible here, so instead I run a fixed-effects regression on only the OECD countries. The results are similar to the overall FE regression in Table 1. 4.2 The impact of trade on income differences The results in Table 3 show that differences in initial income per capita and in factor accumulation rates are significant in explaining differences in current income per capita. In the OLS specification, bilateral trade is 0.044** (0.013) 7 0.79** (0.071) 7 0.050** (0.048) 7 0.13** (0.043) 0.21 (0.15) 0.37** (0.078) 0.20** (0.061) 0.15 (0.095) 0.75** (0.0091) 7 0.090** (0.023) 7 0.68** (0.030) 7 0.53** (0.067) 7 0.040 (0.12) 0.54** (0.043) 0.40** (0.084) 7 9.38** (1.19) 0.69 9642 0.67** (0.026) 0.032 (0.027) 7 0.66** (0.034) 7 0.73** (0.070) 0.29* (0.12) 0.62** (0.044) 7 0.074 (0.14) 7 3.20* (1.46) 0.65 9642 Variable IV 1.25** (0.068) 7 0.30** (0.094) 7 0.37** (0.10) 7 0.57** (0.14) 7 0.055 (0.076) 0.86** (0.25) 0.27 (0.19) 0.48** (0.10) 0.11 (0.095) 0.26# (0.13) 7 42.06** (2.53) 0.92 1797 FE OECD Notes: The dependent variable is the log of bilateral trade in 1965, 1970, 1975, 1980, 1985, 1990, 1995 and 2000. OLS refers to ordinary least squares, IV to instrumental variables, and FE to fixed effects. Time dummy variables are included but not reported here. Standard errors are in parentheses (robust standard errors for the OLS and IV equations). **, * and # denote significance at the 99%, 95% and 90% levels, respectively. The R2 reported is the adjusted R2 for the OLS and IV equations, and the ‘within’ R2 for the FE regression. The instruments are described in the text The total effect of each variable for OECD countries is the sum of the ‘variable’ and ‘OECD’ coefficients. R2 N constant Trading Blocij Languageij ln Remotenessi + ln Remotenessj Adjacencyij ln Distanceij jln GDPi/Popi - ln GDPj/Popjj ln GDPi + ln GDPj OECD Variable OLS Table 2 The impact of income differences on trade: OECD Does convergence cause trade or does trade cause convergence? 409 7 0.019** (0.0021) 0.72** (0.0072) 0.51** (0.013) 0.20** (0.011) 0.094** (0.0052) 7 0.13** (0.013) 0.70 9642 0.028** (0.0035) 0.72** (0.0073) 0.56** (0.013) 0.22** (0.012) 0.082** (0.0054) 7 0.30** (0.017) 0.69 9642 IV 0.12** (0.010) 7 0.19** (0.011) 0.017** (0.0050) 1.28** (0.016) 0.11 9642 7 0.069** (0.0036) FE 0.061** (0.013) 7 0.15** (0.014) 0.011# (0.0058) 1.69** (0.035) 0.01 7499 7 0.19** (0.011) FE-IV 7 0.048** (0.0028) 0.81** (0.013) 0.21** (0.010) 7 0.12** (0.011) 0.038** (0.0049) 0.27** (0.020) 0.65 9642 RE Notes: The dependent variable is the absolute value of the difference between country i and country j’s log GDP per capita in 1965, 1970, 1975, 1980, 1985, 1990, 1995 and 2000. OLS refers to ordinary least squares, IV to instrumental variables, FE to fixed effects, and RE to random effects. Time dummy variables are included but not reported here. Standard errors are in parentheses (robust standard errors for the OLS and IV equations). **, * and # denote significance at the 99%, 95% and 90% levels, respectively. The R2 reported is the adjusted R2 for the OLS and IV equations, the ‘within’ R2 for the FE regression, and the ‘overall’ R2 for the RE regression. The instruments are described in the text. R2 N Constant jln (ni + g + d) - ln (nj + g + d)j jln shi - ln shjj jln GDP60,i /Pop60,i – ln GDP60,j/Pop60,jj jln ski - ln skjj ln(Tradeij) OLS Table 3 The impact of trade on income differences 410 The Journal of International Trade & Economic Development Does convergence cause trade or does trade cause convergence? 411 significant in reducing income differences; but when instruments are provided for trade, trade increases income differences. How can this be reconciled with the common result of cross-country growth regressions (such as Frankel and Romer, 1999) that a higher trade share increases income? The answer could be that trade does increase income, but it increases rich countries’ income more than poor countries’. Interestingly, the fixed-effects and random-effects regressions provide the same answer as OLS: trade reduces income differences. Initial income differences and all of the factor accumulation variables are significant in all of the specifications. Table 4 presents results for the OECD. In both the OLS and the IV equations, differences in initial income, physical capital accumulation rates, and population growth rates are less important for OECD than non-OECD countries, while differences in human capital accumulation rates are more important for OECD members. The OECD dummy is negative in the IV specification, indicating that OECD countries have smaller income differences, all else held constant, than non-OECD countries. As in the standard regressions, trade lowers income differences in the OLS specification but raises them once instruments are used for trade. Still, the negative coefficient on the interaction term with trade shows that the divergent aspects of trade are smaller for OECD countries than others. Again, the fixed-effects regression, which is run only for OECD countries, shows that trade reduces income differences. 4.3 Causality tests To determine whether income differences Granger-cause trade, it is necessary to regress trade on lagged trade and on lagged income differences; if the coefficient on lagged income differences is significant, then income differences Granger-cause trade. Similarly, if lagged trade is significant in an income-differences regression, then trade Granger-causes income differences. An important part of the Granger test is determining how many lags to include. Usually, with data that are yearly or less frequent, only one or two lags are appropriate. The Akaike Information Criterion (AIC) confirmed that the correct lag length is one. Before testing for Granger causality, it is important to determine whether the variables in question have unit roots, as this would affect the validity of the results. Therefore, the Im – Pesaran – Shin (2003) test for unit roots in the context of panel data was undertaken for both trade and income differences. The finding was that trade is stationary, but income differences are integrated of order 1. The Granger causality tests must then be amended in the manner of Toda and Yamamoto (1995), with one additional lag added due to the fact that one of the variables was found to be I(1). A standard Wald test is applied, but only to the first lag, not the additional lag; the Wald statistic then has an asymptotic w2 distribution. 7 0.015** (0.0032) 7 0.14** (0.019) 7 0.18** (0.036) 0.094* (0.026) 7 0.030* (0.0091) 0.015 (0.019) 7 0.0055* (0.0027) 0.70** (0.0078) 0.48** (0.014) 0.17** (0.012) 0.098** (0.0061) 7 0.083** (0.014) 0.71 9642 0.062** (0.0051) 0.68** (0.0084) 0.53** (0.015) 0.19** (0.013) 0.079** (0.0065) 7 0.25** (0.019) 0.68 9642 Variable IV 7 0.013* (0.0053) 7 0.15** (0.022) 7 0.22** (0.040) 0.11** (0.029) 7 0.16 (0.010) 7 0.14** (0.026) OECD 0.0027 (0.018) 7 0.037* (0.016) 0.0056 (0.0045) 0.55** (0.030) 0.20 1797 7 0.022** (0.0063) FE Notes: The dependent variable is the absolute value of the difference between country i and country j’s log GDP per capita in 1965, 1970, 1975, 1980, 1985, 1990, 1995 and 2000. OLS refers to ordinary least squares, IV to instrumental variables, and FE to fixed effects. Time dummy variables are included but not reported here. Standard errors are in parentheses (robust standard errors for the OLS and IV regressions). **, * and # denote significance at the 99%, 95% and 90% levels, respectively. The R2 reported is the adjusted R2 for the OLS and IV equations, and the ‘within’ R2 for the FE regression. The instruments are described in the text. The total effect of each variable for OECD countries is the sum of the ‘variable’ and ‘OECD’ coefficients. R2 N Constant jln (ni + g + d) - ln (nj + g + d)j jln shi - ln shjj jln GDP60,i /Pop60,i – ln GDP60,j/Pop60,jj jln ski - ln skjj ln(Tradeij) OECD Variable OLS Table 4 The impact of trade on income differences: OECD 412 The Journal of International Trade & Economic Development Does convergence cause trade or does trade cause convergence? 413 Results from the Granger causality tests are presented in Table 5. Part (a) shows the results from regressing bilateral trade on lagged values of trade (five and ten years in the past) and lagged income differences. It can be seen that past income differences are significant in explaining current bilateral trade flows. The w2 statistic shown is for the significance of the first lagged income differences variable, and we reject the null hypothesis that income differences cannot be used to predict current trade flows. Income differences thus Granger-cause trade. Part (b) of the table shows the results from regressing bilateral income differences on lagged values of income differences and lagged values of trade. Again, we reject the null hypothesis of Granger non-causality, and we see that trade Granger-causes income differences with a negative sign, showing that trade leads to lower income differences. 5. ROBUSTNESS AND SPECIFICATION TESTS 5.1 Robustness tests The OLS and IV equations test for contemporaneous correlation. Another way to perform a similar test is to use a three-stage least squares system of Table 5 Granger causality (a) Do income differences granger-cause bilateral trade? ln(Tradeij) jln GDPi /Popi – ln GDPj/Popjj Adj. R2 w2-statistic Prob 4 w2 (b) Does bilateral trade granger-cause income differences? jln GDPi /Popi – ln GDPj/Popjj ln(Tradeij) Adj. R2 w2-statistic Prob 4 w2 Lag 1 Lag 2 0.70** (0.013) 7 0.23** (0.063) 0.87 13.96 0.0002 0.22** (0.012) 0.17** (0.063) 1.30** (0.012) 7 0.011** (0.0024) 0.96 19.01 0.0000 7 0.31** (0.012) 0.0090** (0.0024) Notes: In (a), the dependent variable is the log of bilateral trade. In (b), the dependent variable is the absolute value of the difference between country i and country j’s log GDP per capita. **, * and # denote significance at the 99%, 95% and 90% levels, respectively. 414 The Journal of International Trade & Economic Development equations. When this kind of test was performed, results were very similar to the IV regressions: income differences were shown to lower trade, and trade was found to raise income differences. Tests were performed to check whether a single instrument was driving the results. For each instrumental variables regression, additional regressions were performed which excluded each instrument. The results were robust to these tests. The coefficients on income differences and on trade (as well as the other variables) varied only slightly in each regression, and the significance levels were unchanged. 5.2 Specification tests Hausman (1978) tests were performed for all regressions. The Hausman test is a test of equality between the OLS estimates and the IV estimates. The purpose of the test is to determine whether OLS provides consistent and efficient estimates, which will be the case if there is no correlation between the regressors and the error term. If this is true, then the IV estimates should simply equal the OLS estimates plus noise. The null hypothesis of the Hausman test, then, is that there is no endogeneity and so OLS is the correct specification. For both the trade equation and the income-differences equation, the null hypothesis that the OLS and IV coefficients are the same is soundly rejected at the 99 per cent level. The Hausman tests indicate a need for instruments, but can the instruments that have been chosen be considered valid? To determine this, tests of overidentifying restrictions were performed to discover whether the chosen instruments are indeed uncorrelated with the error term. In an overidentifying restrictions test, the residual of the OLS equation is regressed on the proposed instruments. The R2 of this regression is multiplied by the number of observations to obtain a w2 statistic with degrees of freedom equal to the number of endogenous variables less the number of instruments. In all cases, the w2 statistic is significant at a 90 per cent level or higher, showing that there is at least some correlation of the instruments with the error term, and thus that they are not perfect instruments. However, it should be noted that even almost perfect instruments may have difficulty in passing this test, since the high number of observations in these regressions pushes up the w2 statistic. Any correlation of the instruments with the error, no matter how small, will cause the instruments to fail the test if the sample size is sufficiently large. It may instead be useful to examine the R2 of the equations that regress the OLS residual on the instruments. If the R2 is low, then it may be inferred that, even if particular instruments are significant, together they have little overall power to explain the error term. The R2 in these regressions is 0.015 for the trade instruments and 0.036 for the income differences instruments, showing that, in general, the instruments explain little of the error term. Does convergence cause trade or does trade cause convergence? 415 6. CONCLUSION This paper provides several ways of estimating the relationship between trade and convergence. First, instruments for income differences and for trade are provided in order to determine their true contemporaneous relationship. Using initial income and factor accumulation variables as instruments for income, it is shown that countries with more similar incomes per capita trade more, thus supporting the new trade theory over theories of trade based solely on factor proportions. Trade is instrumented by gravity and cultural variables and seems to increase income differences. However, in tests utilizing the time-series nature of the dataset, trade is found to close the gap between rich and poor countries. Fixed- and random-effects regressions show that income differences reduce trade and that trade reduces income differences. Granger causality tests reinforce this result and show that the causality is bi-directional. At a particular point in time, trade is associated with greater differences in per-capita income, but over time, trade causes those differences to diminish. OECD countries are examined in some detail; since this group has seen the greatest amount of both trade and convergence, perhaps the process works differently for them. The results for OECD countries differ in some small ways, but the fundamental linkages are basically the same as those of non-OECD countries. There does not appear to be a ‘convergence club’ at work here. NOTES 1 2 3 4 5 6 Data are from the Penn World Table. The 101 countries in the Penn World Table for which data are available have experienced a similar rise in the average trade share. See Mankiw et al. (1992) for a discussion of conditional convergence. See Grossman and Helpman (1995) for a survey. The gravity equation of bilateral trade, analogous to Newton’s law regarding the gravitational force between two objects, states that two countries’ trade is directly related to the size, or GDP, of the two countries, and inversely related to the distance between them. Data from an earlier version of the Penn World Table show that, for the 56 developing and industrial countries in this sample in 1990, the correlation between capital per worker and output per capita is 86 per cent. The predictions of Helpman and Krugman and Linder are similar, but are not quite the same. To be specific, Helpman and Krugman predict that the product of two countries’ GDPs will have a positive effect on their trade, whereas Linder predicts that the difference of two countries’ per-capita GDPs will have a negative effect. To be able to choose between these predictions, a gravity regression was run which included income and income per capita for both trading countries. It was found that both predictions are upheld. The sum of the logs of the two countries’ GDPs raises their trade, and the difference between their per-capita GDPs lowers trade. 416 7 8 9 10 The Journal of International Trade & Economic Development The languages included are Arabic, Chinese, Dutch, English, French, German, Japanese, Portuguese and Spanish. The trading blocs included are EU/EFTA, NAFTA, ASEAN, Andean Pact, Mercosur, US-Israel FTA, and Australia-New Zealand FTA. Following Frankel (1997), the trading bloc dummy extends back to the beginning of the sample, even for countries that formed trading blocs later. The rationale is that the ties that cause the formation of a trading bloc tend to exist long before the bloc formally comes into being. See, for example, Frankel et al. (1995). Although 12,320 observations are possible for the 1540 country pairs (56 6 55 / 2) over 8 time periods, only 9642 observations are used. This is because the remaining observations are instances of zero bilateral trade, which drop out when logs are taken. Common ways to include the zero-trade observations, including the use of tobit and of re-coding zero trade to a small number, resulted in very similar results. REFERENCES Barro, R. and Sala-i-Martin, X. (1991) ‘Convergence across states and regions’, Brookings Papers on Economic Activity 1, 107 – 58. Ben-David, D. (1993) ‘Equalizing exchange: trade liberalization and income convergence’, Quarterly Journal of Economics 108, 653 – 79. Ben-David, D. (1996a) ‘Trade and convergence among countries’, Journal of International Economics 40, 279 – 98. Ben-David, D. (1996b) ‘Technological convergence and international trade’, CEPR Discussion Paper No. 1359. Ben-David, D. and Loewy, M. (1996) ‘Knowledge dissemination, capital accumulation, trade and endogenous growth’, CEPR Discussion Paper No. 1335. Ethier, W. (1982) ‘Decreasing costs in international trade and Frank Graham’s argument for protection’, Econometrica 50, 1243 – 68. Feenstra, R. (1990) ‘Trade and uneven growth’, NBER Working Paper No. 3276. Frankel, J. (1997) Regional Trading Blocs in the World Economic System, Washington, DC: Institute for International Economics. Frankel, J. and Romer, D. (1999) ‘Does trade cause growth?’ American Economic Review 89, 379 – 99. Frankel, J., Stein, E. and Wei, S. (1995) ‘Trading blocs and the Americas: the natural, the unnatural, and the supernatural’, Journal of Development Economics 47, 61 – 95. Grossman, G. and Helpman, E. (1991) Innovation and Growth in the Global Economy, Cambridge, MA: MIT Press. Grossman, G. and Helpman, E. (1995) ‘Technology and trade’, In Grossman, G. and Rogoff, K. (eds) Handbook of International Economics, Vol. III. Amsterdam/New York: North-Holland, 1279 – 337. Hausman, J. (1978) ‘Specification tests in econometrics’, Econometrica 46, 1251 – 71. Heckscher, E. (1919) ‘The effect of foreign trade on the distribution of income’, Ekonomisk Tidskrift 21, 497 – 512. Reprinted (1991) in Flam, H. and Flanders, M. (eds) Heckscher-Ohlin Trade Theory, Cambridge, MA: MIT Press, 43 – 69. Helliwell, J. and Chung, A. (1990) ‘Macroeconomic convergence: international transmission of growth and technical progress’, NBER Working Paper No. 3264. Does convergence cause trade or does trade cause convergence? 417 Helpman, E. (1981) ‘International trade in the presence of product differentiation, economies of scale and monopolistic competition: a Chamberlin – Heckscher – Ohlin approach’, Journal of International Economics 11, 305 – 40. Helpman, E. (1984) ‘Increasing returns, imperfect markets, and trade theory’, In Jones, R. and Kenen, P. (eds) Handbook of International Economics Vol. I. Amsterdam/New York: North-Holland, 325 – 65. Helpman, E. (1987) ‘Imperfect competition and international trade: evidence from fourteen industrial countries’, Journal of the Japanese and International Economies 1, 62 – 81. Im, K.S., Pesaran, M.H. and Shin, Y. (2003) ‘Testing for unit roots in heterogeneous panels’, Journal of Econometrics 115, 53 – 74. Jones, R. (1956) ‘Factor proportions and the Heckscher – Ohlin theorem’, Review of Economic Studies 24, 1 – 10. Krugman, P. (1979) ‘Increasing returns, monopolistic competition, and international trade’, Journal of International Economics 9, 469 – 79. Krugman, P. (1980) ‘Scale economies, product differentiation, and the pattern of trade’, American Economic Review 70, 950 – 59. Krugman, P. (1981) ‘Intraindustry specialization and the gains from trade’, Journal of Political Economy 89, 959 – 73. Linder, S. (1961) An Essay on Trade and Transformation, New York: Wiley. Mankiw, N., Romer, D., and Weil, D. (1992) ‘A contribution to the empirics of economic growth’, Quarterly Journal of Economics 107, 407 – 37. Ohlin, B. (1924) ‘The theory of trade’, Reprinted (1991) in Flam, H. and Flanders, M. (eds) Heckscher – Ohlin Trade Theory. Cambridge, MA: MIT Press, 75 – 214. Rivera-Batiz, L. and Romer, P. (1991) ‘Economic integration and endogenous growth’, Quarterly Journal of Economics 106, 531 – 55. Samuelson, P. (1948) ‘International trade and the equalisation of factor prices’, Economic Journal 58, 163 – 84. Samuelson, P. (1949) ‘International factor-price equalisation once again’, Economic Journal 59, 181 – 97. Slaughter, M. (2001) ‘Trade liberalization and per capita income convergence: a difference-in-differences analysis’, Journal of International Economics 55, 203 – 28. Toda, H.Y. and Yamamoto, T. (1995) ‘Statistical inference in vector autoregressions with possibly integrated processes’, Journal of Econometrics 66, 225 – 50. Young, A. (1991) ‘Learning by doing and the dynamic effects of international trade’, Quarterly Journal of Economics 106, 369 – 405. APPENDIX A. DATA SOURCES The trade data are from the IMF Direction of Trade Statistics. The adjacency and common language data were constructed by Shang-Jin Wei (www.nber.org/*wei) and were provided by Jeffrey Frankel and Shang-Jin Wei. The distance and remoteness variables were calculated using the greatcircle method. The GDP, GDP per capita, physical capital accumulation, and population growth data were obtained from the Penn World Table 6.1 (pwt.econ.upenn.edu). The human capital accumulation data were obtained from the UNESCO Statistical Yearbook and the World Bank’s World Development Indicators. The common trading-bloc dummy variable was constructed by the author. 418 The Journal of International Trade & Economic Development APPENDIX B. LIST OF COUNTRIES USED IN REGRESSIONS OECD: Asia: Australia Austria Belgium Canada Denmark Finland France Germany Greece Italy Iceland Ireland Japan Netherlands New Zealand Norway Portugal Spain Sweden Switzerland Turkey United Kingdom United States Africa: Algeria Egypt Ethiopia Ghana Kenya Morocco Nigeria South Africa Tunisia China Hong Kong India Indonesia Iran Israel South Korea Malaysia Pakistan Philippines Singapore Taiwan Thailand Latin America: Argentina Bolivia Brazil Chile Colombia Ecuador Mexico Paraguay Peru Uruguay Venezuela