1. For a thorough review of the theory and evidence on the relationship between exports and growth see Edwards (1993 Edwards, S. 1993. Openness, trade liberalization, and growth in developing countries. Journal of Economic Literature, 31: 1358–1393. [Web of Science ®] , [Google Scholar]) and Greenaway and Sapsford (1997).

2. Rodrik (1999 Rodrik, D. 1999. The New Global Economy and Developing Countries: Making Openness Work, Washington, DC: Overseas Development Council.  [Google Scholar]) argues that paying for imports is the only role for exports. Esfahani (1991 Esfahani, H. S. 1991. Exports, imports and economic growth in semi-industrialized countries. Journal of Development Economics, 35: 93–116. [Crossref], [Web of Science ®] , [Google Scholar]) included intermediate imports in his study of 31 semi-industrialised countries and found that the coefficient on exports falls and for some periods becomes insignificant. The coefficient on imports was always found to be positive and significant however.

3. Other studies examine differences in the relationship between exports and growth based on world market conditions, such as oil crises (examples include Balassa (1985 Balassa, B. 1985. Exports, policy choices and economic growth in developing countries after the 1973 oil shock. Journal of Development Economics, 18: 23–35. [Crossref], [Web of Science ®] , [Google Scholar]) and Ram (1985 Ram, R. 1985. Exports and economic growth: some additional evidence. Economic Development and Cultural Change, 33: 415–425. [Crossref], [Web of Science ®] , [Google Scholar])).

4. Whilst the threshold technique employed here improves on those employed in many previous studies, it still shares the limitation that it imposes a discrete break in the data rather than allowing for the possibility of continuous coefficient change. One method that may allow for such a possibility would be to employ a smooth transitions analysis (see Granger and Terasvirta, 1993 Granger, C. W. J. and Terasvirta, T. 1993. Modelling Nonlinear Economic Relationships, Oxford: University Press, Oxford.  [Google Scholar]) that allows for a smooth transition between regression regimes over time rather than a single structural break. Gonzalez, Terasvirta and van Dijk (2005 Gonzalez, A., Terasvirta, T. and van Dijk, D. 2005. “Panel smooth transition regression models, SSE/EFI Working Paper Series in Economics and Finance no. 604”.  [Google Scholar]) have recently begun developing this technique for panel regressions which offers a suitable alternative to the techniques of Hansen and in our setting would allow the parameter associated with exports to change smoothly as a function of a third variable, such as initial income or export growth.

5. The 43 countries are Algeria, Angola, Benin, Botswana, Burkina Faso, Burundi, Cameroon, Cape Verde, Central African Republic, Chad, Congo, Democratic Republic of the Congo, Cote d'Ivoire, Egypt, Ethiopia, Gabon, Gambia, Ghana, Guinea, Guinea-Bissau, Kenya, Lesotho, Madagascar, Malawi, Mali, Mauritania, Mauritius, Morocco, Mozambique, Namibia, Niger, Nigeria, Rwanda, Senegal, Seychelles, Sierra Leone, South Africa, Tanzania, Togo, Tunisia, Uganda, Zambia and Zimbabwe.

6. The coefficient on population growth is expected to be negative and significant. In the existing (mainly cross-section) literature however both positive and negative coefficients have been obtained, with Levine and Renelt (1992 Levine, R. and Renelt, D. 1992. A sensitivity analysis of cross country growth regressions. American Economic Review, 82: 946–963. [Web of Science ®] , [Google Scholar]) finding the coefficient on this variable to be ‘fragile’. We find that excluding the country fixed effects and estimating a pooled or a random effects model give us the more usual negative coefficient on population growth, though the coefficient tends to be insignificant. One possibility therefore is that population growth is capturing some form of market size effect.

7. The differing results for thresholds based on the level of exports and the growth of exports do not appear to be driven by the notion that countries with currently low levels of exports have higher growth rates of exports. The correlation between those countries in the low regime according to the level of exports and those according to the growth of exports is only 0.05.

8. In the two threshold model the estimate of λ ₁ is no longer asymptotically efficient since it was estimated from a sum of squared errors function that was contaminated by the presence of a neglected regime. Bai (1997 Bai, J. 1997. Estimating multiple breaks one at a time. Econometric Theory, 13: 315–352. [Crossref], [Web of Science ®] , [Google Scholar]) suggests a refinement estimator for λ ₁, which involves fixing the second threshold at the estimate for λ ₂ and searching for the first threshold again, now including the second threshold. In Table 4 we report the refined estimator for λ ₁.

9. We have data on manufactured exports in total merchandise exports for less than two thirds of our observations. The average value of this variable for the countries classified in our threshold model as high income was significantly greater than that for those countries classed as low income in our analysis (20.1 versus 6.9 per cent).

10. Some support for this explanation is found by considering a threshold based on the share of manufacturing exports in total merchandise exports for the reduced sample of 185 observations. Here we find a significant threshold at a value of 23.6 per cent. For observations in the low manufacturing share regime we find a positive coefficient on exports, while for observations in the high manufacturing share regime we find a negative coefficient. The coefficients tend to be insignificant in both regimes however. These results are available on request.