Structural changes, cointegration and the empirics of Thirlwall's law

Abstract

Thirlwall's law establishes a relation between the long-run growth rate, the growth of exports and the long-run income elasticity of imports. The estimation of this parameter requires cointegration techniques, which in turn require a large span of data, thus exposing the estimates to risks of structural changes. While this problem has been recognized in the literature, the evidence produced is still partial, being concerned with a very limited number of countries, and in some respects unsatisfactory. In this article we fill this gap by assessing Thirlwall's empirical regularity on a sample of 22 Organization for Economic Cooperation and Development (OECD) countries using econometric techniques that allow for the presence of a shift of unknown date in the long-run parameters. The results are generally supportive of Thirlwall's hypothesis and allow us to reconcile and qualify the evidence provided in the existing literature.

Acknowledgements

Useful comments from Tony Thirlwall and an anonymous referee are gratefully acknowledged. Usual disclaimers apply. This research was supported by an MIUR grant.

Notes

¹ Among the first contributions we recall Thirlwall (1979, 1981, 1986), Thirlwall and Hussain (1982), McCombie (1985, 1989, 1992), McGregor and Swales (1986, 1991), Bairam (1988), Fagerberg (1988), Bairam and Dempster (1991).

² The same result is obtained by assuming that the Marshall–Lerner condition holds as an equality throughout the sample; see Thirlwall (1979).

³ Although under the hypotheses above, specified the two expressions are equivalent at the theoretical level, in empirical studies Equation 4 is preferred, because of its simplicity and also because the growth rate of world demand involved in Equation 3 raises some measurement issues (see Bairam and Ng, 2001), and the study of Bairam (1997) points out that the income elasticity of exports is more likely to undergo structural changes; see Section III.

⁴ Since our analysis rests on Equation 4, this survey will focus on issues concerning the estimates of the long-run import functions.

⁵ The argument of McCombie (1997), according to which, since Thirlwall's law is concerned with long-run growth rates the use of growth rates does not determine a loss of information, clearly misses the point that the law considers a long-run, rather than an impact, elasticity, and in order to get consistent estimates of the long-run parameters we need to utilize the levels of the variables.

⁶ By the way, Equation 4 appears preferable, because it involves only one estimated parameter, instead of the ratio of two estimated parameters, and is therefore less subject to sample variability.

⁷ Previous analyses of the impact of structural breaks on the performance of Thirlwall's law outside the cointegration approach include Atesoglu (1993, 1995).

⁸ The increase in the sample size determined by the adoption of higher frequency data is more apparent than real, since what matters for cointegration analysis is the span, rather than the number, of the observation (Otero and Smith, 2000).

⁹ Due to lack of data, the estimation for Canada, Denmark, France, New Zealand and Switzerland was performed over the reduced sample 1966–2006.

¹⁰Atesoglu (1997) studies the cointegration between two variables (the US GDP and exports). The bias in the Engle and Granger cointegration estimates occurs in the presence of multiple cointegration vectors, which in turn implies that the variable considered are at least three.

¹¹ In the cases of Finland and New Zealand the noncointegration hypothesis is rejected at the 10% level without considering a structural break and at the 5% level once a structural break is taken into account.

¹² An anonymous referee pointed out that these cases could be coped with the test for panel cointegration with multiple structural breaks proposed by Westerlund (2006). We acknowledge gratefully this suggestion. However, Westerlund's (2006) test allows for multiple breaks in the deterministic components of the individual regressions. This corresponds to a panel version of the ‘level shift’ model (Equation 10), and is therefore unsuited to detect the possible presence of multiple ‘regime shifts’ (i.e. breaks in the long-run elasticities).