st: Xtoverid and negative Hausman Test

[May Ster <mayfrank@gmail.com>]

Dear all,


I’m currently working on Panel Data . In my data, panel var = 90 units
, time var = 38 years.
For your informaiton I use ;
Xtreg lnY lnA1 lnA2 lnB1 lnB2 lnC lnD  Dummy1 Dummyf2 Dummy3 Dummy4
Dummy5 Dummy6
LnY = Dependent Var.
LnA1, LnA2, LnB1, LnB2, LnC, Dummy1,2  = time-variant variables
LnD, Dummy3, 4, 5,6 = time-invariant variables.
It is assume to be a linear regression containing individual effects ;
I then have checked for the appropriateness of the individual effects
model to be use. [The Hausman test proves fixed effects model is
preferred over the random effects model]
Then, I took the benefit of the Hausman-Taylor estimation as according
to the literatures there should be correlations between some of
explanatory variables and individual effects. As well, if using fixed
effects model, time-invariant variables could not be observed. Thus,
this will not bring any benefits to the interpretation of the effects
since coefficient on, for example  from above lnDij will not be read.
My question is, is it necessary to perform another Hausman Test to
state again that the Hausman-Taylor is a consistent estimator? In
certain literatures, the researcher’s choice choosing  the
Hausman-Taylor is simply to acquire coefficient on time-invariant
variables in which the fixed effects model couldn’t perform.
However, I actually did perform the Hausman Test and found the test
showing negative value. I then consulted Statalist so far on this
issue. A lot of them say this case is common and I should do as STATA
suggested; using ‘Suest’. But, ‘Suest’ can’t be applied with ‘Xtreg’,
and then I became very doubtful.
Another way is to use the command ‘XTOVERID’ after ‘Xthtaylor’ can be
the way to obtain positive value of Sargan-Hansen Statistic. But if I
understand it correctly, ‘XTOVERID’ is an over-identifying test to
sort of signal that instruments I have chosen are appropriate. I still
do not see how ‘XTOVERID’ can be used to compare whether the
Hausman-Taylor is more consistent than the fixed-effects in this case.
I am certain that some use ‘XTOVERID’ to decide the fixed effect
versus the random effects models. But, how is it done about comparing
the Hausman-Taylor verus the fixed effects model?
Some says, I do not have to compare these two models; as if given the
fact that the fixed-effects estimator was consistent, why I would have
to use the Hausman-Taylor. I am not so sure about this. I wonder
whether there is such a test to conclude that the Hausman Taylor is
more consistent than the fixed effects estimator (which lacks lots of
efficiency in this case).
Best Regards,

Orachat

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/