Thanks to Rodrigo and Steve for taking the time to comment, I will consider
these in my work.
Regards,
JULIUS
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Rodrigo Alfaro
Sent: 27 November 2005 17:17
To: [email protected]
Subject: st: Re: RE: panel data analysis using xtregar
Julius, first lagged dependent variable (LDV) with xtreg, fe and xtregar is
not the same. xtregar works like Cochrane-Orcutt regression (meaning you
have the first observation... read Autocorrelation section in any textbook
to learn about it)... for that purpose you have to set what kind of rho will
be in your regression (option rhotype)... the default is Durbin-Watson.
Details of the formula appear in the manual reference. If the coefficient of
your LDV is near to 1 maybe you have a Unit Root case (note that this
coefficient is downward-biased if it is positive Nickell 1968). There is a
huge (and maybe more than that) about LDV+FE versus Arellano-Bond (1991, GMM
estimator), Anderson-Hsiao (1981, IV estimator)... it seems to me that the
bias generated by LDV+FE it is not so bad in compare with weak instruments.
Hahn-Kuersteiner (2000) propose an asymptotic correction for that
parameter... all these is available in Arellano's book. Rodrigo.
----- Original Message -----
From: "Steve Stillman" <[email protected]>
To: <[email protected]>
Sent: Saturday, November 26, 2005 8:48 PM
Subject: st: RE: panel data analysis using xtregar
> Julius. It is not exactly clear what your model is here, but, in general,
> panel data models with lagged dependent variables cannot be consistently
> estimated by directly entering the lagged variable as a RHS variable. It
> sounds like you should have a look at xtabond and the user-written
> xtabond2 for your model.
>
> Cheers,
> Steve
>
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]]On Behalf Of Julius
> Fr�d�ric Andr�
> Sent: Sunday, November 27, 2005 7:26 AM
> To: [email protected]
> Subject: st: panel data analysis using xtregar
>
>
> Dear statalist,
>
> Currently researching balanced panel data using insolvency prediction
> estimates for different companies and regressing them (among other
> controlling, time-variant variables) on time dummies to evaluate the
> effects
> of a policy introduction in a certain year. I am using xtreg, fe for
> evaluation, the panel is described by t=12 and i=110. I presume
> autocorrelation of the insolvency predictor (ie. the dependent variable)
> and
> therefore constructed a dataset with a lagged variable for the insolvency
> predictor, hence also had to eliminate the first period of the original
> dataset. It turned out that the estimator of the lagged variable is indeed
> highly significant.
>
> Now I also consider using xtregar. Being new to stata (also to panel data
> and time series analysis), however, I unfortunately could not find the
> underlying formula used by Stata up to now. Would such an autoregressive
> model be similar to the approach I did manually with introducing a lagged
> variable and deleting the values of the first period for each I? The stata
> result being different for the AR(1) model and the fixed-effect model
> including a lagged variable do not corroborate this assumption, so I
> assume
> some other underlying model. I know that an AR appproach incorporates
> past
> values, but does this mean the past error term, the past dependent or the
> past independent variable (or all of them)? It would be of great help if
> someone could provide a formula here or give a brief statement if I should
> resort to xtregar at all if I assume autocorrelation anyways (so that the
> applying AR is simply not necessary anymore).
>
> Additionally, I would like to backup my decision to include a lagged
> variable in the model by testing on autocorrelation, using a Durbin-Watson
> substitute for panel data, such as Bhargava et al. I know that Stata can
> calculate this statistic when using xtregar, however, it seems to me that
> the reported output is the result AFTER applying the AR (1) model,
> indicating only the presence or absence (or degree) of autocorrelation
> left
> after using the AR model? Is this correct, or does the test statistic
> indicate autocorrelation before using the AR model?
>
> Lastly, I am not sure whether I could also use the Baltagi-Wu test score
> alternatively, seeing to it that I have a balanced dataset. Is this test
> statistic only working with unbalanced data?
>
> The three preceding issues certainly are still on a somewhat basic level,
> however, I would definitely appreciate useful comments on any or all of
> these!
>
>
> Thank you very much in advance,
>
> JULIUS ANDRE
>
>
> *
> * 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/
>
>
> *
> * 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/
>
*
* 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/
*
* 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/
