I am using xtabond2 to estimate a dynamic panel (N=5,500 and T=12) using
either the GMM in first-differences (GMM-fd) or the system GMM (GMM-s).
With such a short time series and a large cross-sectional dimension, I
would expect the GMM to do well over the downward biased fixed effects
estimator (FE). However, it appears that the autoregressive estimated
coefficient in GMM-fd is much more downward biased than FE. The GMM-s does
better than GMM-fd but still gives estimation values which are not
significantly different from FE.
Notice that to enhance consistency in estimation I make sure I use in each
model the appropriate lagged values for the instruments, as dictated by the
Arellano-Bond test for serial correlation in the disturbances, which is
reported with xtabond2.
Is this a puzzle or am I doing something obviously wrong?
Many thanks in advance for any help I can get.
