EViews.com

Hi,

I'm using v7.2.
I've estimated a supply-demand system with 3SLS, FIML and GMM.
But very strangely although the 3SLS and GMM std errors are very small -> p-values <0.01
the FIML p-values are all well above 0.5.
Why is that?

And some questions about the availability of tests:
Why aren't the heteroscedasticity, the RESET and the IV tests available in the system estimation?
Why isn't the autocorrelation test available in LIML and in single-equation GMM?
Why isn't the RESET and heteroscedasicity test available in single-equation GMM?
Why isn't the regressor endogeneity test available in LIML?
Why does the RESET test always crash Eviews after LIML (although Eviews never crashes in any other situation)?

Thanks in advance,

Andrew

I'm not going to comment on the econometric questions.

The reason some tests are available for some estimation types and not for others are either that the test is not econometically valid for that estimator, or that the test takes a different form for that estimator, and has not been added (since it can't use the same code as other estimators), or we are not 100% certain that the test is valid.

The crash is a bug, that was previously unreported, but now that you have reported will be fixed.

OK, Thanks.

The reason I put the first question here was:
Is Eviews correctly calculating FIML standard errors?

Cause I've not seen so striking differences in articles or textbooks
and the 3SLS and FIML should be asymptotically equivalent
(my error distributions are normal, the portmanteau test says no autocorrelation
and there are 50 yearly observations)
and also the parameter estimates are very close for 3SLS, FIML and GMM
and the LIML estimates have much smaller std. errors (also p<0.01 as with 3SLS).

In the manual there is no technical discussion for FIML
(for the other types of estimators there is),
it would be great to have one.

Could someone tell me,
which tests from my questions are not available because they are not valid econometrically and why?

To the best of our knowledge the FIML standard errors are correct (and match the results we use for validation).

Asymptotic equivalence is a messy concept and with only 50 observations it's not clear that I would expect equivalence. Similarly, the IID normality requirement for FIML may be suspect (you say that you have normal errors, but unless this is a simulation, I'm not certain how one would argue that the errors in a demand system are actually Gaussian, not to mention uncorrelated, even with the Portmanteau results). Further, as with any nonlinear specification, FIML admits the possibility of multiple roots.

As to the documentation, we are caught in a bit of a quandry. We obviously can't have full econometric discussions of all of the details of all of the estimators in EViews in our manuals. Moreover, for most users, the full set of details are not important. We try to strike the right balance in our docs and describe enough so that people know what is going on and point people to original sources where we feel appropriate. In some cases we may put the dividing line in the wrong place and we appreciate your suggestion that we add more in this case.

That said, as to FIML we didn't feel that there wasn't much discussion required since given the specification, there weren't really any options to set (this is not the case for the other estimators). That said, given the likelihood and the specification, we use an OPG estimator for the covariance of the coefficients.

I hope this helps.

Thanks.