Dear Jon,
(1) If your variable has some variation over time, then the within variation
is small. As long as the within variation converges to a positive number,
you can obtain a fixed-effects estimator for that variable. This estimator
is consistent, but the standard error is huge. You can quickly see the
sample within-variation using -xtsum-. I haven't seem the definition of
highly stable that you provide, could you give me that reference?
I am not sure what you suggest with the autocorrelation value of that
variable. If you have an autocorrelated variable [pretend AR1: x(i,t) =
c(i)+r*x(i,t-1)+e(i,t)] the variance of x increase with r. So far, I haven't
found a solution for a small time-variation variable, but it seems that FE
are still good estimators.
(2) I reply your question with other: why you want to use instruments? Do
you have another variable that is correlated with the error term? with FE
you already drop the cross-sectional unobservable.
Rodrigo.
----- Original Message -----
From: "Jon O'Brien" <jobrien2@nd.edu>
To: <statalist@hsphsun2.harvard.edu>
Sent: Sunday, June 04, 2006 11:51 PM
Subject: st: Fixed effects and stable variables: xtivreg or ivreg2?
Hello Statalist members,
I tried this message 2x before and it didn't go through - hopefully 3rd
times a charm! I have 2 related questions. First question:
When using fixed effects, variables that are invariant over time will get
washed out by the firm/subject level fixed effect. I've read that variables
that are highly stable (i.e., show very little variation over time) may also
largely get washed out by the fixed effect. However, I cannot find any
references that provide guidelines for determining when a [key theoretical]
variable is "too stable" to be accurately estimated via fixed effects. For
example, if I have 10 years of annual data and a key variable is correlated
0.91 with it's lag, is that "too stable"?
Second, if a Hausman test suggests that random effects are inappropriate,
what might be a reasonable alternative to xtivreg with the fe option (note:
the stable variable is not the endogenous one)? Would ivreg2 with gmm and
cluster options be reasonable in this case ?
Thanks for any suggestions you can offer,
Jon O'Brien
Assistant Professor
University of Notre Dame
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/