The Multi-Sectoral Thirlwall’s Law: evidence from 14 developed European countries using product-level data

Abstract

The paper reports estimates of import and export functions for five technological sectors in 14 developed European countries. These functions have never before been estimated for developed countries adopting a technological classification of sectors. The paper compares estimates of income elasticities found using vector error-correction models employing aggregate deflators, with estimates found using cross-product panels employing product-specific quality-adjusted price indexes recently calculated by Feenstra and Romalis. The results indicate that the income elasticities of imports and exports are higher for medium- and high-tech manufactures, which suggests the importance of moving from the production of simple goods to the production of goods with high technological content. The estimates also suggest that the Multi-Sectoral Thirlwall’s Law holds for the countries analysed, while comparing the estimates revealed that cross-product panels with quality-adjusted prices generate considerably more robust results. The investigation reveals that using a more recent time period generates estimates of income elasticities of demand for primary products and resource-based manufactures that tend to be higher than the estimates found by studies that have used longer time periods, while the opposite holds for low-, medium-, and high-tech manufactures.

Keywords:

• balance-of-payments constraint
• Multi-Sectoral Thirlwall’s Law
• international trade
• sectoral analysis
• demand-led growth
• error correction models
• panel data estimations

JEL Classifications:

• O11
• F43
• E12

Acknowledgements

We would like to thank Professors Ana Maria Hermeto and Gustavo Britto, from Cedeplar-UFMG, and three anonymous referees for their helpful comments. The usual disclaimer applies. This work was supported by the Coordination for the Improvement of Higher Education Personnel (CAPES-Brazil) in a partnership with the Cambridge Trusts, under Grant number 0257-11-7.

Notes

1. Houthakker and Magee’s (1969, 121) seminal work explored differences in income elasticities between US sectors.

2. For simplicity, these equations disregard cross-price elasticities.

3. Lall (2000) classifies SITC (Rev. 2) three-digit product categories into technological sectors. See Lall (2000) for a detailed analysis of the evolution of world trade in each technological sector (across different country groups) between the 1970s and the 2000s. Owing to the poor quality of the data for the sector Other Manufactures and due to the relatively low relevance of this sector (which represents on average around 0.3% of total world exports), data related to this sector were not used in this paper’s tests.

4. Tharnpanich and McCombie (2013) regressed import and export functions by primary and manufacturing products. Nonetheless, the authors do not explore the different levels of technology within manufacturing. In spite of that, they find that manufactured products face higher income elasticities than primary products. Gouvea and Lima (2013), in turn, estimated sectoral elasticities using cross-country panels. However, they adopted the Broad Economic Classification (BEC) instead of Lall’s Technological Classification. Furthermore, they did not estimate the functions for each country separately.

5. See Feenstra and Romalis (2014) for a detailed description of the methodology used to estimate quality-adjusted price indexes.

6. World income is assumed to be exogenous, given that it is unlikely that the exports of one SITC product category from one country to the world generates any significant impact on world income. In addition, local income is also assumed to be exogenous. Although imports are a component of local income, it is unlikely that the imports of one SITC product category exert a significant effect on local income.

7. See Wooldridge (2002) and Baum (2006) for detailed discussions on instrumental variable methods.

8. See Baum, Schaffer, and Stillman (2007) for a detailed discussion of this estimator.

9. See Griffith, Harrison, and van Reenen (2006) and Hausmann, Hwang, and Rodrik (2007) for some examples of works that employ System GMM.

10. Although data are available for more recent years, these data were not used to avoid capturing the short-term effects of the 2007 financial crisis and the subsequent European crisis.

11. Gouvea and Lima (2013, 244) used the average official exchange rate (national currency/US dollar) and the ratio of the implicit US GDP deflator to the countries’ GDP deflator to measure relative prices. This measure is analogous to 1/PPP (from WDI). PPP data, however, are available for a longer period of time. It is also worth noting that similar measures of relative prices were used by Gouvea and Lima (2010) and Tharnpanich and McCombie (2013).

12. Feenstra and Romalis (2014) estimated quality indexes, unit price indexes, and quality-adjusted price indexes for SITC (Rev. 2) four-digit product categories for 185 countries over the period 1984–2011. Focusing on developed countries, the selection of the countries analysed in the current paper was primarily guided by the coverage of the quality-adjusted price indexes calculated by Feenstra and Romalis (2014), since missing quantity data prevent the calculation of prices indexes for all SITC products in all years and countries. The 14 selected countries were the ones for which data with associated price indexes: (i) represents more than 80% of the total value of exports and imports in the whole period, and more than 80% of the total value of exports and imports in each of Lall’s (2000) technological sectors; (ii) presents on average no less than 80 SITC categories within each technological sector (40 for High-Tech Manufacturing); and (iii) presents an average number of SITC categories with no less than 15 years available within each technological sector. The high coverage of the data for these countries in relation to the total data on exports and imports minimises the possibility of sample selection bias. Ireland was excluded from the sample due to the lack of data on GDP before 1995.

13. Most of the empirical literature that employs panel data models uses either five- or ten-year averages. In this paper’s tests, four-year averages were used to maximise the number of time periods in the database (1984–2007).

14. With rare exceptions, all series are I(1) according to either the Phillips-Perron and/or the Augmented Dickey-Fuller tests. Furthermore, in the vast majority of the cases, Johansen's Trace Statistic, the Maximum-Eigenvalue Statistic, and/or HQIC, and SBIC indicate that there is only one cointegrating vector between the series. In the cases where either one of the variables was not I(1) (namely, Finland’s imports, Norway’s low-tech manufactures (LTM) exports, and Portugal’s medium-tech manufactures (MTM) imports), or the number of cointegrating vectors was different from one (Denmark’s high-tech manufactures (HTM) imports, Germany’s HTM imports, Italy’s primary products (PP) exports, and Portugal’s LTM exports and imports), the functions were estimated using OLS in first difference with Newey and West’s (1987) heterogeneity and autocorrelation robust standard errors. See Enders (1995), Johansen (1995) and Becketi (2013) for detailed discussions on time series econometrics.

15. In all regressions, Hausman’s test indicated that the fixed effects estimator was preferable to the random effects estimator.

16. When a model has several endogenous variables, it is not possible to assess how well each endogenous variable is being instrumented. In all the regressions the income elasticity of demand was positive and significant at the 0.1% level. Hansen’s (1982) J Test rejected the null hypothesis of over-identification in only 10 of the 140 regressions, while Kleibergen and Paap’s (2006) LM Tests rejected the null hypothesis of under-identification in all the regressions.

17. In all regressions but one (for PP imports from Switzerland) the income elasticities of demand were positive and significant at the 5% level. Moreover, Hansen’s J Test rejected the null hypothesis of over-identification in only three of the 140 regressions, while Arellano and Bond’s (1991) AR Test rejected the null hypothesis of no autocorrelation in the second lag (the first used as an instrument) in only nine of the regressions at the 5% level.

18. Gouvea and Lima (2010) also found some extremely large income elasticities using VECMs: 10.073 for high-tech exports from Philippines; 8.456 and 12.224 for low- and medium-tech exports from Malaysia, respectively; 5.874 and 6.499 for medium- and high-tech exports from Mexico, respectively; and 8.066 for high-tech imports from Korea.

19. It is important to mention that in the VECMs, although most of the price elasticities of demand for imports were negative, as expected, the opposite was found for exports. This suggests that the aggregate measures of relative prices normally used in the balance-of-payments constrained growth literature are imperfect measures, especially when sectoral export and import functions are estimated. For the cross-product panels, however, both for exports and for imports, the price elasticities were predominantly negative. The IV estimator with Hausman’s instruments is the estimator that generates the highest number of negative price elasticities. Thus, these results indicate the superiority of using quality-adjusted price indexes and Hausman’s instruments.

20. As Thirlwall (2013, 51) argued, ‘there may be at certain times skill bottlenecks, but if the industrial sector of an economy needs more labour, it will find it’. The question, therefore, is the pace of this transfer whenever supply bottlenecks become relevant.

21. The relatively high income elasticities of demand for imports found for Denmark, Germany and Italy (especially in the high-tech sector) seems to be the result of intra-industry trade between highly developed countries (see import shares in Table 3). Again, similar results are observed in Gouvea and Lima’s (2010) work, where Korea has the highest income elasticity of demand for high-tech imports.

22. Taking into account the effect of relative prices and calculating the equilibrium growth rate according to equation (8(8) ) does not change the results. Employing estimates of price elasticities found using IV with Hausman’s instruments (regardless if they are significant or not), the average absolute difference increases from 0.48 to 0.57, while the only change in the predictions is that Germany’s equilibrium growth rate is now below the actual growth rate, which wrongly suggests a current account deficit. Hence, considering price effects worsens the statistical fit of the equilibrium rates.

23. Soukiazis, Cerqueira, and Antunes (2013) have proposed an expanded version of Thirlwall’s Law that distinguishes the import content of aggregate demand (dividing income into consumption, investment and government expenditure), and introduces public deficit and debt measures as determinants of growth. This model seems to be better suited to explain the trade imbalances and the short-term growth rate fluctuations observed in Greece, Portugal and Spain. Focusing on the example of Portugal, the authors show that using the weak version of Thirlwall’s Law (y_BOP = x/π), over the period 1986–2010, the equilibrium rate of Portugal (2.338) is lower than its actual average growth rate (2.728), which correctly predicts current account deficits. Moreover, the equilibrium growth rate predicted by their model (1.995) is lower than the original Thirlwall’s Law, which predicts even higher deficits. However, if the strong version of Thirlwall’s Law (y_BOP = ɛz/π) is employed using the estimates found by the authors, then the equilibrium rate ((2.88*2.5)/2.63=2.738) is actually slightly higher than the actual growth rate of Portugal, which wrongly predicts current account surplus. Moreover, it is important to note that the authors used the rate of growth of the income of OECD countries to measure the growth of foreign demand instead of the growth rate of world income, as carried out in this paper. Hence, this reinforces the argument that a longer timespan is necessary to estimate the long-term growth rate of southern European countries due to the stronger impact that the creation of the Eurozone exerted on these countries.