[Gretl-users] Re: Hausman test and robust estimates
On Fri, 4 Nov 2022, Alison Loddick wrote:
This I'm hoping is an easy question. I'm doing panel
regression
with robust estimates. I have read that you cannot do a Hausman
test on robust estimates. I'm wondering therefore whether Gretl
does the test on the non-robust estimates or whether it does a
robust hausman test?
As explained in the chapter of the Gretl User's Guide titled "Panel
data" there are two ways of carrying out the Hausman test, the
matrix-difference method (which can fail in finite samples) and the
regression method (which will always work).
By default gretl uses the regression method, and handles correctly
the case where a robust covariance estimator was employed for the
random effects model.
Actually the documentation in the User's Guide is out of date on
this point: it implies that gretl always uses sums of squared
residuals in computing the Hausman test via the regression method
(which would not be robust). In fact, in the robust case we compute
a Wald chi-square statistic based on a robust estimator of the
covariance matrix of the augmented regression. So the doc needs
updating; thanks for drawing that to our attention!
Allin Cottrell