A comparison of FDML and GMM for estimation of dynamic panel models with roots near unity

This thesis compares the performance of the first-differenced maximum likelihood estimator (FDML) and the Blundell-Bond continuously-updating system GMM estimator of the autoregressive parameter in an AR(1) dynamic panel model without exogenous covariates, particularly focusing on the close-to-non-stationary case. This case is far from trivial, as a high degree of persistence is the norm rather than the exception in economic panels. The results of the Monte Carlo simulations show that the absolute mean and median biases of the FDML are higher for low values of N and T in the close-to-non-stationary case. However, the biases become negligible for both estimators as $N$ and $T$ increase. The power of the GMM is generally higher than that of the FDML, while, on the other hand, the GMM suffers from severe size distortions. This problem is magnified both when T increases, as well as when the process approaches non-stationarity. Finally, the GMM estimator is shown to display Cauchy properties when the process is very close to non-stationarity. This produces some peculiar bias results for certain combinations of N and T when using the GMM