The Stata 8 (and 7) command xtabond estimates the Arellano and Bond
dynamic panel data model such as
y_it = a * y_{i,t-1} + b * x_it + nu_i + eps_it
In the Stata 8 Cross Sectional Time Series manual page 16, it says that
the nu_i are random effects that are i.i.d. over the panels.
In the Arellano and Bond paper, page 280, they first derive the
estimator for the fixed effects case, where nu_i can be correlated with
x_it. After this, the authors describe how assuming nu_i is a random
effect and not correlated with x_it allows for some additional moment
restrictions with which to increase efficiency, which they then derive.
The rather stylized discussion of the xtabond "Methods and Formulas" on
page 28 and on in the manual does not mention that the estimator
imposes the additional moment restrictions associated with random
effects.
So does anyone know whether xtabond uses the easier fixed effects
assumption, or whether it imposes the additional moment restrictions
derived from the random effects assumption?
