GMM - Continuously Updating Estimator Puzzle with EVIews 7

[maka]

I noticed that I get very different results if I apply Eviews 7 Generalized Method of Moments (GMM)-Continuously Updating Estimator (CUE) to a given sample, depending on whether I ran the CUE only, whether I ran the CUE after the N-step iterative weight updating, or whether I ran the CUE after the Iterate to convergence weight updating (all with the default settings) . Are there any reasons that can explain this GMM-CUE puzzle with EViews 7?

Thanks,

Alexis

[EViews Glenn]

I've looked at the workfile and equation in question (they were sent via email) and responded in email...

CUE is a highly nonlinear method of estimating a GMM specification and as with all nonlinear estimation methods it is possible that you will get different results for estimation beginning at different starting values. I will also note that your particular specification uses 37 observations to estimate 17 coefficients, with HAC kernel weighting. I suspect that the relative paucity of data for such a complicated specification undoubtedly contributes to the observed instability. As with all nonlinear specifications where there may be multiple local roots, you may use the objective function to discriminate between solutions.