If you really have dynamic panels with strong time series effect (and
it looks so once you talk about causality -- is it Granger causality
that you want to test?) then both -xtreg, re- and -gllamm- will be of
little help as they assume independence of observations taken at
different time points conditional on the random effect. The focus of
the causality tests is to build a time series model and see the
strength of the relations between two time series.
Yes, -xtreg, [re|mle]- implicitly assumes MAR and provides consistent
and asymptotically efficient estimates (as MLE would, from the general
missing data theory, provided your model is correctly specified). The
reason it is not documented properly is that MAR/MCAR/NMAR concepts
are used in sociology, while panel data models are written mostly for
economists, so there is a lack of communication between the two
industries.
On Tue, 12 Oct 2004 15:57:39 +0200, Michael Ingre
<[email protected]> wrote:
> Hi Binh
>
> > I have read
> > that to avoid the bias and inconsistency caused by
> > heterogeneity I should use a mixed fixed and random
> > effect model.
>
> The term "mixed" refers to the fact that both random and fixed effects
> are estimated and -xtreg- fits mixed effects models (with a random
> intercept) by default or by using the -re- option for the GLS estimator
> and the -mle- option for maximum likelihood estimation.
>
> > my dataset contains many missing
> > observations. My question is should I drop all these
> > observations for the consistent and efficient
> > estimates?
>
> If you drop the observations you not only reduce statistical power you
> also assume that data is missing completely at random (MCAR). This is a
> rather strong assumption and can usually be relaxed by maximum
> likelihood estimation to assume data only being missing at random
> (MAR). MAR does not mean that missing is random (sic!) it may be
> systematic. However, the probability of of a missing value should be
> related to the covariates in the model, not the dependent variable.
> Look at Schafer & Graham (2002) for an introduction.
>
> I think that -xtreg y x , i(id) mle- produce valid estimates under the
> MAR assumption but I have not seen a reference to it in the manual.
> Perhaps some one else on the list can enlighten us.
>
> Another program called -gllamm- is robust when data is MAR but it is
> slow in converging and might not be the optimal solution for you. It
> can however be downloaded directly from within Stata -ssc install
> gllamm-.
>
> Michael
>
> References
>
> Schafer, J. L. and Graham, J. W. Missing data: our view of the state of
> the art. Psychological Methods 2002,7:147-177
--
Stas Kolenikov
http://stas.kolenikov.name
*
* 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/
