International financial integration and real exchange rate long-run dynamics in emerging countries: Some panel evidence

Abstract

The aim of this article is to provide new empirical evidence on the impact of international financial integration on the long-run Real Exchange Rate (RER) in 39 developing countries belonging to three different geographical regions (Latin America, Asia and MENA). It covers the period 1979–2004, and carries out ‘second-generation’ tests for non-stationary panels. Several factors, including international financial integration, are shown to drive the long-run RER in emerging countries. It is found that the new financial environment characterised by international financial integration leads to a depreciation of the RER in the long run. Further, RER misalignments take the form of an under-valuation in most MENA countries and an over-valuation in most Latin American and Asian countries.

Keywords:

• emerging economies
• real exchange rate
• financial integration
• misalignment
• second-generation panel unit-root
• cointegration tests

JEL Classifications:

• E31
• F0
• F31
• C15

Acknowledgements

We are very grateful to two anonymous referees for helpful comments and suggestions on an earlier draft of this article.

Notes

1. Net foreign assets (NFA) represent wealth and affect domestic demand through intertemporal optimisation. Higher demand for goods leads to lower relative prices of non-tradable goods, and an appreciating RER.

2. According to Marquez (1990), elasticities comply with this condition, at least in the long run (see also Morel and Perron 2003).

3. Bénassy-Quéré et al. (2004, p 10).

4. Our approach differs from that of Bénassy-Quéré et al. (2004) for at least three reasons. First, we deal with 39 developing countries belonging to three different geographical regions (Latin America, Asia and MENA), rather than developed ones (G20). Second, our purpose is different since we investigate the long-term behaviour of the RER from the point of view of the new financial architecture, and extend existing studies by considering the possible effect of international financial integration on the RER long-run dynamics. In contrast, Bénassy-Quéré et al. (2004) examine the effect of net foreign asset on the equilibrium RER. Third, Bénassy-Quéré et al. (2004) use first-generation panel unit root and cointegration tests developed on the assumption of cross-sectional independence of panel units, an assumption that is unrealistic in many empirical settings. Our analysis performs instead second-generation tests allowing for various types of dependence across the different units.

5. For further details on these three alternative measures of international financial integration we refer the reader to the paper by Milesi Ferretti et al. (2005, 2006) (from which these three measures were taken), that examines trends in gross and net international investment positions and their components for a large set of advanced and developing economies and that includes in an appendix graphs and statistics about these measures of international financial integration.

6. Note that for each of the different groups of countries (all emerging countries, Latin America, Asia and MENA) we have retained the measure of IFI that was the most statistically significant.

7. It should be noted that, before carrying out the second-generation panel unit-root tests that allow for cross-section dependence, we have implemented the simple test of Pesaran (2004) and have computed the CD statistic to test for the presence of such cross-section dependence in the data. This test is based on the average of pair-wise correlation coefficients of the OLS residuals obtained from standard augmented Dickey-Fuller regressions for each individual unit. Its null hypothesis is cross-sectional independence and it follows asymptotically a two-tailed standard normal distribution. The null hypothesis is always rejected for all series for all emerging countries, as well for each of the three groups of countries, regardless of the number of lags included in the augmented DF auxiliary regression (up to five lags) at the 5% and 10% level of significance. This confirms that the members of our panel are cross-sectionally correlated and that therefore any first generation panel unit root tests (which assume cross-country independence), such as the test by Hadri (2000), would be flawed and cannot be used here.

8. The results are not very sensitive to the size of the bootstrap blocks.

9. The lag order in the individual ADF type regressions is selected for each series using the AIC model selection criterion.

10. Note that to deal with the few borderline cases reported in Table 1 where the joint non-stationary null is (marginally) rejected at the 10% (though not at the 5%) level (for example, MENA countries/public spending), we have investigated the robustness of the results by implementing the recent CIPS panel unit root test developed by Pesaran (2007), who showed that, by augmenting the usual ADF regression with the first difference and the first lag of the cross-sectional mean, one can account for the cross-sectional dependence arising through a single stationary factor. The results of this other second-generation panel unit root tests (available upon request) provide a clear support for the existence of a unit root in all series under consideration at the 5 and 10% level of significance.

11. As pointed out by a referee it is also worthwhile to use some panel cointegration tests with a null of no cointegration (in addition to the test of Westerlund and Edgerton 2007, whose null hypothesis is joint cointegration for all countries), in order to see if these additional tests also reject that null, hence providing evidence that the cointegration results are robust. As mentioned by the same referee, ‘there are at least two reasons for using alternative tests. One is the well-known fact that classical hypothesis testing implies that the null hypothesis is accepted unless there is strong evidence to the contrary. The other reason is that the bootstrap test used here is new and its properties have been investigated only through a limited Monte Carlo exercise with one result that was “somewhat unexpected”. In addition, the Monte Carlo exercise in Westerlund and Edgerton used two different values of T (50, 100) which are bigger than the one used here (T = 26)’.Therefore, in order to take into account possible cross-sectional dependence of panel units we have computed the bootstrap distribution of Pedroni's cointegration test statistic (1999, 2004), thereby generating data-specific critical values. As in Banerjee and Carrion-i-Silvestre (2006), we have of course not used the seven statistics proposed by Pedroni but only the parametric version of the statistics, i.e. the normalised bias and the pseudo t-ratio statistics, and in particular the ADF test statistics. These test statistics (available upon request) always reject the null hypothesis of no cointegration between RER and fundamentals, irrespective of whether the model includes a constant or a linear trend. These results are robust for the four groups of countries considered in our investigation (all emerging countries, Latin America, Asia and MENA) and thus provide additional support for cointegration.

12. The reason why we have chosen to estimate the long-run coefficients by DOLS following Kao and Chiang (2000) and not by FMOLS (fully modified OLS) as proposed by Pedroni (2000) is the following observation of Kao and Chiang (2000, p 216), who notice that: ‘the OLS estimator has a non negligible bias in finite sample, FMOLS estimator does not improve over the OLS estimator. The FMOLS is complicated by dependence of the correction terms upon the preliminary estimator. More seriously, the failure of the non-parametric correction for the FM in panel could be severe. This indicates that DOLS estimator may be more promising than the OLS or FMOLS estimators in estimating cointegrated panel regressions’.

13. We are of course aware of the debate in the literature on the use of filters and that ‘filtering matters’ particularly when using the HP, or the band pass filter. It has been shown (see Guay and St-Amant 2006) that the HP and BK filters do relatively well when applied to series that have a peak in their spectrum at low frequencies, but that they do poorly with series whose spectrum decreases sharply and monotonically at higher frequencies, i.e. series that have the typical spectral shape identified by Granger (1966). Consequently, the following simple strategy should be followed by applied researchers: estimate the spectral (or pseudo-spectral) density of the series of interest, so that the appropriateness of using the HP and BK filters to identify the cyclical component can be evaluated. The use of the HP and BK filters is very problematic when a series has the typical Granger shape.