EViews.com

Hello,

I have two questions regarding GMM and would be very grateful for any comments and help:

When I estimate several GMM equations, I may get "near singular matrix" result even when I am re-estimating equation which I have previsouly estimated without encountering this note... I thought this might be related to c and resid vectors - when I reset the c vector to 0 vector (default when new workfile is opened), the estimation works just fine...

Thus, I would like to aks whether I am correct in linking the problem to the c vector and if so, why?

Also, when the GMM estimation is run and then re-run (by pressing estimate again), the coefficients are altered slightly (after several successive re-estimations the coefficients seem to converge) - why is this so and should we use the "converged" version (usually the difference is of a small magnitude and would not matter due to rounding).

Thank you very much in advance for your help.

"Bad" coefficient estimates can cause a singular matrix error in a nonlinear model. Think what happens with y = x/(1-c(1)) if c(1) happens to be near 1.0. So your surmise about the C vector is quite right.

When you redo GMM it has to do at least one new estimate to find out that it has converged. That will move the coefficients, hopefully by a very small amount. That's because "converged" means that step sizes have gotten very small, but not to zero.

As an aside, GMM should be less sensitive to starting values than most estimators, since EViews actually uses TSLS to get starting values for GMM. The C vector is only used for starting values for TSLS (if the TSLS problem is itself non-linear).

Thank you for both of the replies :D

So would it be all right to simply change the c vector back to 0 vector whenever the non-singularity problem arises? - of course, if this does not help there is a problem with the coefficients as you have illustrated, I suppose

Thank you once more for the helpful comments - I really appreciate your help