The last time I wanted to estimate a time-series model in the presence of
autocorrelation and endogenous RHS variables, I used TSP, which had a
nifty AR() command for essentially doing Stata's "prais" or old "corc"
command with instrumental variables. I see that Baum, Schaffer, and
Stillman have expertly provided the Stata community with a new version of
ivreg2, which I think can correct the var-covar matrix for
heteroscedasticity and for autocorrelation.
My concern is that ivreg2 may be getting the standard errors right but not
necessarily the coefficient estimates, at least not in the way traditional
corrections for autocorrelation typically do. I'm concerned because in
the phillips.dta example discussed in the helpfile, the parameter
estimates do not change when the bw() option is used. That is,
. ivreg2 cinf unem
. ivreg2 cinf unem, bw(3)
. ivreg cinf unem
. reg cinf unem
all yield exactly the same coefficient estimates. (The standard errors
are a little different across most but not all of these.) By contrast, if
. prais cinf unem
we obtain different estimates and standard errors. I was under the
impression that this outcome is to be expected after correcting for
autocorrelation. While OLS is technically unbiased, because the direction
of bias is unknown and therefore should be zero on average, in practice
autocorrelation acts as an omitted variable, moving the estimates in one
direction or another, even though it is unknown ex ante. So it would be
odd to see the estimates remain exactly the same after autocorrelation
were corrected for.
Am I wrong to be concerned? Or am I perhaps missing the spirit of the
autocorrelation correction in ivreg2? Is there a different Stata
algorithm that does what I'm looking for?
Ryan D. Edwards, Ph.D.
Postdoctoral Fellow in the Study of Aging, RAND Corporation
and Visiting Scientist, 2004-2005
Department of Population and International Health
Harvard School of Public Health
* For searches and help try: