4.2 Dynamic Panel Estimation
Columns (1) and (2) inTable 4 present the one-step and the two-step estimators using the
Arellano and Bond (1991) dynamic panel estimation, respectively. Column (3) presents
the two-step estimator using the Arellano and Bover (1995) dynamic panel estimation.
It is worth nothing that both the Arellano–Bond and Arellano–Bover system GMM
estimations have passed a battery of diagnostic checks. The Sargan test does not reject the
overidentification restrictions, which indicates that the overidentification restrictions are
valid. For the two-step estimator of Arellano–Bond dynamic panel estimation, we failed
to reject the null hypothesis of no first-order serial correlation and second-order serial
correlation. The Sargan statistic is asymptotically distributed as chi-squared with 127
degrees of freedom, i.e. 127 overidentification restrictions for the Arellano–Bond panel
estimation. For the Arellano–Bover system GMM, the fact that the absence of the
first-order serial correlation is rejected while the absence of the second-order serial
correlations is not indicates that there is zero autocorrelation in the first differenced errors.
The real exchange rate has a positive impact in OLS models (Table 3), but the impact is
mixed in dynamic panels (Table 4), varying from positive in one-step estimator to
negative insignificant in the two-step estimator. We then compared the dynamic GMM
panel estimation (columns 1-3) with the OLS fixed effect panel estimation (column 4) to
highlight the information about parameters. Despite the Arellano and Bond (1991)
argument that additional instruments can be obtained in a dynamic panel data model if
one utilises the orthogonality conditions that exist between lagged values of dependent