#93 random effects model: F statistic or wald chisq statistic

[KTTK]

I noticed, for the random effects model, no F statistic (test of joint significance of coefficients) is outputted by Gretl. Searching a bit on this topic, it seems like there is no F statistic for random effect models. Stata outputs a chisq test instead and its documentation says: "All that is known about the random-effects estimator is its asymptotic properties, so rather than reporting an F statistic for overall significance, a chisq is reported." (http://www.stata.com/manuals14/xtxtreg.pdf, p. 18). I have not found further references about this topic...maybe you have some at hand?

So, it would be nice if Gretl outputs the Wald chisq by default for RE models. Also, I would be very helpful if there would be a statement of the lacking of the F statistic for joint significance in the manual, ch. 18 (which is really nice!)
Another point in this regard: Are the test for linear restrictions (post estimation Tests -> linear restrictions) valid for RE models? They apply an F test...

[Allin Cottrell]

On Mon, 28 Mar 2016, KTTK wrote:

So, it would be nice if Gretl outputs the Wald chisq by default
for RE models.

Fair enough, but it's pretty easy to do this yourself:

panel y 0 X --random-effects
scalar k = $ncoeff - 1
matrix q = zeros(k, 1)
matrix R = q ~ I(k)
restrict
R = R
q = q
end restrict --silent
X2 = $test * k
pvalue X k X2

[KTTK]

Oh, again thank you for your very quick reaction. This is very handy as I am not experienced with hansl and is surely also welcomed by users who just use the GUI.

As I have still not found another reference (besids the Stata manual), do you have one?

[Sven Schreiber]

Just one last concrete reaction here instead of the mailing list (don't
know if you're subscribed to the list):

I'd say this (the unknown finite-sample behavior) just follows
generically from the fact that the RE estimator is a feasible GLS (FGLS)
estimator, no further or special reference needed.

This doesn't mean that an F-form test is "wrong". Often we hope that the
F-form will approximate the true (but unknown) distribution better in
finite samples. The applied researcher just should know that the F-form
test doesn't have an exact F-distribution in these cases. But the
chi2-form test also doesn't have an exact chi2 distribution, only an
asymptotically justified one.

cheers,
sven

Am 29.03.2016 um 19:10 schrieb KTTK:

Oh, again thank you for your very quick reaction. This is very handy
as I am not experienced with hansl and is surely also welcomed by
users who just use the GUI.

As I have still not found another reference (besids the Stata
manual), do you have one?

[Allin Cottrell]

Please see
http://lists.wfu.edu/pipermail/gretl-users/2016-March/011726.html
for further discussion.

[Sven Schreiber]

Allin just posted on the mailing list that the RE output now includes this overall chi2 test statistic, so I'm closing this ticket. For the other discussion items see the mailing lists.
Thanks,
sven

[KTTK]

By the same argument (unknown properties in finite/small samples): one would also take the normal distribution instead of the t distribution for single coefficent test?

[KTTK]

For reference: the change to z test for single coefficient tests was done simultaneously with the change from F test to Wald chi-square test for joint significance of coefficients. Thank you, Allin!