Dear Mark
Thank you very much for your message. The problem is that (with xtoverid) I
do not know any way to ascertain whether the possibly endogenous variable is
exogenous or whether I have a weak instruments problem (or whether the
random effects estimates are preferred over the fixed). With xtoverid, is
there any way to know the estimates I have to rely on?. In fact, using the
ivreg2 command with the endog( ) option shows that the variable is not
endogenous, but this is not a panel data estimation and I do not know
whether, from the ivreg2 estimation, I can simply conclude that there is not
an endogeneity problem. In any case, I still would have to solve the problem
of getting the coefficients of the time-invariant dummies if the Hausman
test indicates that the fixed effects is the preferred estimation (could
xthaylor provide a consistent solution?).
On the other hand, I have been suggested to estimate GMM System through
xtabond2, but reading David Roodman's paper, it seems to me that the context
in which this is applied is different (1. I have dummy variables that could
bias the results; 2. I have 59 firms followed an average of 25 quarterly
periods; 3. I have a good external instrument; 4. I do not have lags of
dependent variables as regressors). Please, any advise on this?
Thanks !
Jaime.
-----Mensaje original-----
De: [email protected]
[mailto:[email protected]] En nombre de Schaffer, Mark E
Enviado el: jueves, 08 de octubre de 2009 16:22
Para: [email protected]
Asunto: st: RE: Instrumental variables and panel data
Jaime,
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]] On Behalf Of
> Jaime Gómez
> Sent: 06 October 2009 23:13
> To: [email protected]
> Subject: st: Instrumental variables and panel data
>
> Dear Statalisters
>
> We have a model in which firm performance depends on (1) the
> order of entry and (2) a possibly endogenous variable and (3)
> other variables, including time dummies. First, we were
> suggested to use instrumental variable techniques and to
> provide HAC standard errors, something we have already done
> with the ivreg2 command in Stata and using an external
> instrument. We tested for the exogeneity of the possibly
> endogenous variable through the endog( ) option and the test
> shows that the variable could be considered exogenous.
>
> In a second step, we have been suggested to use the panel
> structure of our data and, simultaneously, to consider the
> endogeneity problem. Ideally, we would like (1) to estimate a
> panel data model with instrumental variables and HAC errors,
> (2) to test for the exogeneity of our possible endogenous
> variable and (3) to check whether the fixed or random effects
> model is appropriate. So, it seems that the xtivreg or
> xtivreg2 commands could be the solution. Nevertheless, we
> have several problems:
>
> 1) the order of entry is represented through time invariant
> dummies (pioneer, second mover, third mover, ...) that drop
> when we estimate a fixed effects model, but we are (very)
> interested in the values of the coefficients. So it seems
> that the only way of getting these coefficients is to
> estimate a random effects model and check whether this is
> appropriate with a Hausman test (If I reject the random
> effects model, ¿could I get the order of entry coefficients
> through another panel data technique?)
>
> 2) Before doing so we have to find the way of getting HAC
> standard errors. I think I would know how to do this with
> xtivreg2 (I am assuming that the options are similar to the
> ones in ivreg2), nevertheless it seems that there is no way
> of estimating a random effects model with xtivreg2. The
> problem with using xtivreg seems that the estimation and
> postestimation options are much more restricted than with
> xtivreg2 (for example, how do I get HAC errors? How do I test
> for the endogeneity of the regressor? Should I use xtoverid
> for testing for the appropriateness of the random effects model?).
>
> In summary, is there any way for treating all these issues
> (possibly omitted variables that advise the use of panel data
> techniques, time invariant variables of interest, HAC
> standard errors and instrumental variables) at the same time?
> Alternatively, could you suggest another strategy to tackle
> all the problems with Stata (perhaps sequentially?)?
A couple of thoughts...
1. You can use -xtoverid- with the undocumented -noisily- option to
estimate a random effects model with various types of robust SEs. There
have been several threads on Statalist about it, so it should be pretty easy
to find. (I really have to get around to making -xtivreg2- do random
effects....)
2. Cluster-robust SEs are robust to arbitrary within-cluster correlation as
well as heteroskedasticity, and you can think of them as a variety of HAC
SEs. The main difference between them and the usual kernel-based HAC SEs
(as supported by -xtivreg2- et al.) is that the asymptotics for
cluster-robust SEs have the number of clusters going off to infinity; the
asymptotics for the usual kernel HAC SEs (Bartlett kernel aka Newey-West and
all those guys) is that they require time to go off to infinity. Most
panels these days are small-T-large-N, so chances are you would be better
off with cluster-robust. Of course, it's up to you.
Cheers,
Mark
> Thanks a lot
> Sincerely
> Jaime Gómez
> Universidad de Zaragoza
>
>
> *
> * 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/
>
--
Heriot-Watt University is a Scottish charity
registered under charity number SC000278.
*
* 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/
*
* 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/
