Re: st: Panel Fixed(random) effects model with autocorrelation and heteroskedasticity

[Joachim Landström <joachim.landstrom@fek.uu.se>]

Lucas,
It sounds as if a PCSE would be in place but such models are  
asymptotic in T and not in N (in other words does such models rest on  
having "small" N/T-ratios)which means that it is out of the question  
for you (since your N/T-ratio is "large"). You need a model that is  
asymptotic in N and which can treat autocorrelation and  
heteroscedasticity.

I suggest that first you do a Prais-Winsten transformation of your  
variables to remove any presence of autocorrelation. Then you might  
add a time dummy to reduce the likelihood of panel-wide  
heteroscedasticity. I think Roodman (2008) discusses the use of time  
dummies to remove panel-wide heteroscedasticity. If you suffer from  
heteroscedasticity on id-level you may have to consider to also  
deflate your variables (i.e., if they still are on levels).

However, you do not specify your model so here I have to make a guess.  
If you use a lagged  dependent variable in your model, you really must  
use a dynamic panel regression model since your FE model most likely  
suffers (badly) of Nickell bias (too small T). That brings you into  
the use of e.g. xtabond or similar models.

Reference
Roodman, D. 2008. How to do Xtabond2: An Introduction to Difference  
and System GMM in Stata. In Center for Global Development Working  
Paper No. 103: Center for Global Development.

Regards,
Joachim


Quoting Lucas Bremer <bremer@bwl.uni-kiel.de>:

> Dear all,
>
> I read a lot in the stata archive, but I didn't find the right answer for my
> problem.
>
> In fact, I have two panel datasets (each 700 obs, 6 time periods). So first
> of all I tested random vs. fix effects with the hausman test. For one Panel
> I should use random effects, for the other one fixed effects. Then I tested
> for serial correlation and heteroskedasticity with positive results
> (http://www.stata.com/support/faqs/stat/panel.html).
>
> Now I search for the right estimation procedure to handle serial correlation
> & heteroskedasticity for random effects and for fixed effects.
>
> Is there a procedure that can do the corrections for fe & re? I think it is
> better to use the same estimation procedure in both datasets and do not
> switch between them to be able to compare the results.
>
> I tried to use xtgls ... , panels(hetero) corr(ar1) for the random effects
> model. For the fixed effects model I added for each ID a personal dummy, but
> that didn't worked.
>
> Thank you in advance,
> Lucas
>
>
>
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/



--
Joachim Landström


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/