Hi,
The original post has a lot of interesting problems to deal with.
1. It is seems that there are some unobservable individual effects.
2. Some of the exogenous could be endogenous.
3. Some of the X have small within variation.
I checked very fast the Fama-MacBeth procedure and
I don't think is a feasible solution here. Guido's suggestion about
Hausman-Taylor seems reasonable. The problem is to know
which X are truly exogenous and the correlation between
those guys and the X that are endogenous. You will end up
doing Hausman test to valid the instruments. With more
programming Amemiya-MacCurdy procedure should be
available if you wrote the code for unbalanced panels.
Still, Kit's suggestions about IV procedure can be applied
to HT procedure, manually you could add valid instruments
in the third steps of HT. Requires to code HT, but it is not hard.
Let's call it this estimator HT2.
Taking Andrea's question I think that orthogonal-forward
deviation (OD) as Arellano-Bover (1995) can be considered
to estimate the model by IV with unobservable non-random
component. It would interesting to see the performance of
HT2 versus IV+OD.
Anyway, the third problem does not have a solution here.
I am working in that problem without endogeneity and it
seems that the FE estimator is asymptotically unbiased but
with high MSE in the case that within variation is very-very
small. Could you tell us the ratio within/between?
I did some theoretical analysis (using double asymptotic)
and the worse scenario gets convergence-rate of n instead
of n*T, but for a short panel it doesn't make a huge difference.
I hope to see more discussion here.
Finally, I would like to know why the panel is unbalanced.
Some economic story... to understand the behavior of the
market (sorry to the more statistical-readers).
Rodrigo.
----- Original Message -----
From: "sistoand80" <[email protected]>
To: "statalist" <[email protected]>
Sent: Friday, February 02, 2007 9:13 AM
Subject: Re:st: Re: fixed effects vs random effects
Hi everybody.
Kit Baum suggests an instrumental variable estimator to dela with
endogeneity. But in this case what is the better estimator between IVfixed
effect (xtivreg2, fe) and IV first-difference (xtivreg2, fd with,
optionally, gmm)?
I've the same problem and I'm dealing with a small sample (14 countries, 6
years)
Thank you!
Andrea Sisto,
University of Turin (Italy)
---------- Initial Header -----------
