Working Paper Series: Special Edition of 2016 to 2018 Interns

[( _ , − − , − −1 ) ∗ ( + )] = 0 (11) [( , − − , − −1 ) ∗ ( + )] = 0 (12) [ , − − , − −1 ) ∗ ( + )] = 0 (13)

[1,4] [1,4] [1,4]

=1, t

=1, t

=1, t

According to Bonnefond (2014), the consistency of the system GMM estimator relies on two hypotheses. First, the set of instrumental variables must not correlate with the error terms. This hypothesis is tested using Sargan/Hansen 39 test of over-identifying restrictions. Second, the absence of second-order autocorrelation (AR2) in residuals must be verified, given that a negative first-order autocorrelation (AR1) may is spotted. This second hypothesis is tested using Arellano- Bond tests for AR(1) and AR(2).

5. Empirical Results and Analysis 5.1 Preliminary Tests

Two preliminary estimations were conducted without considering the exchange rate regimes, using a pooled OLS and a panel regression with fixed effects. Results can be seen in table 5 . The OLS estimator cannot account for potential correlation between the countries’ errors over time and does not address the problem of unobserved heterogeneity. The FE panel regression estimator fixes this issue by taking autogenous transformation of the data, however, it suffers from Nickell bias 40 due to data’s time series being too short. Cognizant of the fact that these estimates will be inconsistent, they were done to compare coefficients with that of the subsequent system GMM estimator.

39 The Hansen test was used in this paper over the Sargan test because the Sargan test works best in the face of homoscedasticity while the Hansen’s test is preferred in the face of cross-country heteroskedasticity, which this paper’s data set consist of. 40 Baun C.F (2013), “Dynamic Panel Data Estimators”, Boston College

133