The real exchange rate and productivity differentials: a panel cointegration approach

Abstract

Balassa (1964) and Samuelson (1964) have expressed the view that economic development is normally accompanied by a real exchange rate appreciation. The aim of the study is twofold: first, we investigate whether the Balassa–Samuelson (BS) hypothesis holds, and second we appraise the performance of the commonly used productivity differentials proxy. Results from the Prais and Winsten (1954), Im, Pesaran and Shin (2003) and McCoskey and Kao (2001) panel data techniques suggest that it is legitimate to consider the BS hypothesis at different stages of economic development. Additionally, the proxy works quite well.

Notes

¹ The selected countries are Denmark, France, Germany, Italy, Japan, Netherlands, Spain, Sweden and UK. All data are annual observations obtained from the Groningen Growth and Development Centre and International Financial Statistics. Tradable sectors include agriculture, hunting, forestry and fishing; mining and quarrying and manufacturing. Nontradable sectors include public utilities (electricity, gas and water); construction; wholesale, retail trade restaurants and hotels; transport, storage and communication; finance, insurance, real estate and business services; community, social and personal services; and government services. Each sector is assigned a weight corresponding to its contribution to gross domestic product (GDP). Labour productivity is computed as the ratio of gross value added to employment.

² SEs are in parentheses. R ² is the within-R ² for fixed effects (FE) and overall − R ² for random-effects (RE). *, **, *** denotes 1, 5 and 10% significance level, respectively. According to the Hausman specification test, the null hypothesis of no systematic difference in coefficients between the FE and RE models cannot be rejected. Hausman's χ ²(5) = 0.13 [0.9997], χ ²(4) = 0.00 [1.0000] for the SPBM and APBM correspondingly. P-values are in square brackets. The FE model is defined as y _it  = α_i  + γ_t  + βx _it  + ν_i  + ε_it . The constant term α_i varies over individual countries but not with time. α_i can be treated as an additional random error. The RE model can be defined as y _it  = α + βx _it  + ν_i  + ε_it . ν_i is the unit-specific residual. The coefficients are assumed to be constant across individuals and the variance unit-specific error term is zero. The Breusch and Pagan (BP) Lagrangian multiplier test cannot accept the null hypothesis of Var(ν) = 0. The BP χ² (1) = 2191.4 [0.0000] and χ² (1) = 2104.62 [0.0000] for the SPBM and APBM. The Wooldridge's autocorrelation tests report F(1, 8) = 34.066 [0.0004] and F(1, 8) = 8.313 [0.0204] for the SPBM and APBM. The null hypothesis of no autocorrelation cannot be accepted. The Arellano and Bond dynamic panel generalized methods-of-moments (GMM) estimators are computed. The GMM consistency depends on the Sargan test of overidentifying restrictions, which tests the overall validity of the instruments. The absence of second order correlation in the error term is also a pre-requisite. So, H₀ should not be rejected. Their p-values are reported above. The second GMM model satisfies these conditions although the null hypothesis of no first order serial correlation cannot be rejected. Green's groupwise heteroskedasticity tests reveal χ ²(8) = 1302.55 [0.0000] and χ ²(8) = 1623.72 [0.0000] for the SPBM and APBM. The null hypothesis of homoskedasticity is rejected. In case disturbances are not independent and identically distributed PW recommend a panel-corrected SE, which can correct for both correlated and heteroskedastic residuals. We know that an AR(1) process exists from the high significance of the lagged endogenous variable of the GMM models. As such, we can estimate the PW model assuming there is first-order autocorrelation and the coefficient of the AR(1) process is specific to each panel.

³ All unit root tests were carried out using regressions including a constant and trend at level forms. For the first differences, only a constant was included since differencing usually removed the deterministic trends. Detailed results of the IPS tests are available upon the author's request.