AR-terms/Cochrane-Orcutt in models with panel data

[karsten_kohler]

Dear all,

I've got a question about the inclusion of ar-terms in a panel context.
What I would like to do is a Cochrane-Orcutt estimation (or some similar method) with panel data. As far as I understood, Cochrane-Orcutt estimation is what EViews does when one includes ar-terms in a regression equation with time series data. More precisely, EViews uses some iterative maximum likelihood estimation method ('ARMA Maximum Likelihood (BFGS)') to estimate the parameters of the model. However, when I use ar-terms in a regression equation with panel data, EViews uses the default estimation method 'Panel Least Squares'. Nevertheless, it does seem to apply some iterative procedure as it reports the number of iterations.

My question is: What is the difference in the estimation methods between time series and panel data when one uses ar-terms? What kind of estimation method does EViews use with panel data and does it amount to something equivalent or similar to Cochrane-Orcutt?

Your help is really appreciated! Thank you!


NB: I've conducted Cochrane-Orcutt manually (by quasi-differencing the data using the rho from a regression of the residuals on its first lag) on time series and panel data, respectively. I've compared the results to the output I get when I use ar-terms and they are very close (although not identical, but I also didn't iterate manually). That suggests that EViews does something similar to Cochrane-Orcutt also in a panel context, but I'm not sure.

[EViews Glenn]

Basically equivalent to iterated C-O, but with the differencing built into the model (and consequently the gradients) and the iteration done using nonlinear least squares. Note that if you have fixed effects in the specification, this doesn't remove the asymptotic bias, which is why the GMM estimators are supported.

[karsten_kohler]

Thank you!