Re: Which matrix is near singular?
Posted: Thu Sep 08, 2016 1:06 pm
by EViews Glenn
It's not the moments that I'm asking about it's the weighting. How do you want the GMM to do the weighting?
Re: Which matrix is near singular?
Posted: Fri Sep 09, 2016 5:48 am
by freedomplaza
Soory didn't notice your response on page 2 of the thread...
Ideally 2-step optimal weighting. For starters, getting results with an identity weighting matrix would be a success.
Re: Which matrix is near singular?
Posted: Fri Sep 09, 2016 9:45 am
by EViews Glenn
Thought about this a bit last night.
The issue with EViews estimation of the TSLS model is that the only algorithm that we currently offer for system estimation is Gauss-Newton. GN doesn't work well in your setting, as the OPG for your model is, by construction, going to be poorly behaved. To see this, note that the projection defined by your single instrument effectively computes the means by equation so there is no variation in the projected gradients within an equation. Put differently, in the IV case with only a ones vector, the OPG for the stacked data configuration is equivalent to one in which there are only two observations.
On a related note, this fact makes it unclear to me as to what optimal two-step weighting may mean in this context.
That said, it is possible to estimate the model with other methods. I've put together a program [attached] which takes your moment specification and estimates it directly using the EViews optimize command to estimate the parameters using BFGS. It hasn't been tested, nor the results verified, but it should get you going.