What does export diversification do for growth? An econometric analysis

Abstract

It is frequently suggested that export diversification contributes to an acceleration of growth in developing countries. Horizontal export diversification into completely new export sectors may generate positive externalities on the rest of the economy as export oriented sectors gain from dynamic learning activities due to contacts with foreign purchasers and exposure to international competition. Vertical diversification out of primary into manufactured exports is also associated with growth since primary export sectors generally do not exhibit strong spillovers. Yet there have been remarkably few empirical investigations into the link between export diversification and growth. This paper attempts to examine the hypothesis that export diversification is linked to economic growth via externalities of learning-by-doing and learning-by-exporting fostered by competition in world markets. The diversification-led growth hypothesis is tested by estimating an augmented Cobb–Douglas production function on the basis of annual time series data from Chile. Based on the theory of cointegration three types of statistical methodologies are used: the Johansen trace test, a multivariate error-correction model and the dynamic OLS procedure. Given structural changes in the Chilean economy, time series techniques considering structural breaks are applied. The estimation results suggest that export diversification plays an important role in economic growth.

Acknowledgements

We thank the Evangelisches Studienwerk e.V. Villigst for financial support.

Notes

¹ The link between export diversification and export earnings instability has been the subject of considerable research in the last two decades. See Stanley and Bunnag (2001) for a review of the theoretical and empirical literature on this topic.

² See Amin Gutiérrez de Piñeres and Ferrantino (2000: Chapter 8) for an endogenous growth model, in which technological or marketing knowledge in one export sector diffuses into other lines of exporting.

³ The authors use several indicators for export diversification, such as, for example, the number of export sectors or the Herfindahl index.

⁴ Balaguer and Cantavella-Jordá (2004) consider the impact of structural transformation from traditional primary exports to nontraditional manufactured exports on Spanish GDP and thus the impact of vertical export diversification on growth.

⁵ Amin Gutiérrez de Piñeres and Ferrantino (2000, Chapter 4) use the Herfindahl index to measure export concentration. The correlation between export concentration and Chilean output turns out to be statistically significant. The coefficient of the Herfindahl index does not have the expected negative sign but is positive, which implies a negative correlation between export diversification and aggregate output.

⁶ The Central Bank of Chile classifies the Chilean manufacturing exports according to the comprehensive definition of manufacturing of the ISIC.

⁷ The declaration refers to three-digit export sectors according to the SITC definition.

⁸ It is empirically not directly observable.

⁹ See, for example, Matsuyama (1992), Chuang (1998).

¹⁰ Augmented Dickey–Fuller and Phillips–Perron tests would indicate that each series is integrated of order 1. (Results are not reported here.) However, the observed unit root behaviour is the result of failure to account for structural changes.

¹¹ The ‘additive outlier model’ implies that the change in the trend function is sudden. The ‘innovational outlier model’ implies that the break in the series does occur gradually.

¹² All our empirical tests have been carried out by EVIEWS 5.0.

¹³ The unit root tests proposed Lumsdaine and Papell (1997) also indicate that real GDP, aggregate capital and employed people are integrated of order 1, whereas the export sector and the industrial share series can be constructed as stationary fluctuations around a breaking trend function. As above, the selected break years in the export sector and the industrial share series are 1971 and 1973. The details of the tests are not reported for brevity, but are available upon request.

¹⁴ Collinearity between lz_t and lix_t was investigated by inspecting the correlation matrix. The correlation coefficient of 0.50 indicates a low degree of collinearity between the detrended series. In contrast, if we compute the correlation matrix of the trended series (LZ_t and LIX_t ) we have a correlation coefficient of 0.96, indicating a very high degree of collinearity.

¹⁵ To assess the structural stability of the trend stationary models, we additionally calculated the recursive residuals. Recursive residual analysis also suggests that there are structural breaks in 1971 and 1973.

¹⁶ Results in the next section further confirm that lz_t and lix_t can be regarded as stationary.

¹⁷ The lag length was determined using the Hannan–Quinn and the Schwarz criterion.

¹⁸ du75 is 1 from 1975 onwards and zero before 1975; d75 is 1 in 1975. The possible reason for du75 and d75 to be important is the deep economic depression in 1975.

¹⁹ t-ratios in parentheses underneath the estimated coefficients. ** and *** denote 5% and 1% level of significance respectively. The number in parenthesis behind the values of the diagnostic tests statistics are the corresponding p-values. JB is the Jarque–Bera test for normality, LM(k), k = 1, 3, are LM tests for autocorrelation based on 1 and 3 lags, respectively and ARCH(k) is an LM test for autoregressive conditional heteroscedasticity of order k = 1, 2, 4. White = White test for heteroscedasticity of the errors.

²⁰ Conventional distributional results are applicable for the t-test statistic since the Bewley-transformed ECM term is stationary (according to the trace test). Additionally, one may argue that the null of no cointegration may be rejected at the 1% significance level, because the t-value of the loading coefficient (−7.27) lies below the critical value for two stochastic regressors (−4.38) according to the test for cointegration suggested inter alia by Ericcson and MacKinnon (2002). However, further stationary variables may influence the distribution of the ECM test statistic under the null of no cointegration.

²¹ Tests for weak exogeneity within the Johansen framework indicate that LK_t is weakly exogenous, while LY_t and LL_t are endogenous. However this test is not invariant to the inclusion of stationary variables, such as Lz_t , Lix_t . Thus, weak exogeneity in the full system (LY_t , LK_t , LL_t , Lz_t , Lix_t ) may differ from weak exogeneity in the subsystem (LY_t , LK_t , LL_t ). Instead of investigating the weak exogeneity status of each of the ‘explanatory’ variables, the DOLS procedure is preferred here.

²² Dummy variables are used to capture the effects of the deep economic crises in 1975; du75 and d75 are defined as in Equation 14.

²³ Following Stock and Watson (1993) the insignificant leads and lags were not dropped. If we follow Hendry's general-to-specific approach the residuals appear not to be as free of autoregressive conditional heteroscedasticity, although the coefficients for the explanatory variables are reasonably similar.

²⁴ However, the results are not directly comparable due to different estimation methods and different economic variables in the estimation equations.