RE: st: Comparing the similar model (but one in level and the other after taking the 1st difference)

[Nick Cox <n.j.cox@durham.ac.uk>]

I will pass on your second question. I have never used these commands or even read their documentation recently. 

On your first, I doubt that there is an approved answer waiting to be discovered. You could summarise residuals in some way, e.g. as a root mean square error. Error measured in terms of your response variable is on the same scale as error measured in terms of differences of your response variable. But these measures are not related directly to the criteria by which these models are fitted. Also, a sensible answer to this question will depend on your substantive or scientific aims, which are not explicit here. A better answer is to see what good researchers do in the literature for your field. 

Nick 
n.j.cox@durham.ac.uk 

Almutairi T.

Dear Nick and Maarten

In this case how to compare bewtween the two models if R^2, LL & IC are not appropriate?

Also: Does xtpcse or xtgls commands (with option AR1) are simply taking the first difference of all variables to ensure the treatment 
of autocorrelateion?? 

From: owner-statalist@hsphsun2.harvard.edu [owner-statalist@hsphsun2.harvard.edu] On Behalf Of Nick Cox [n.j.cox@durham.ac.uk]

The first part of the answer is the general reason. The log-likelihood depends on the scale of the variable. Thus comparison of likelihoods for different response variables is not a comparison of comparable cases.

The FAQ is just about a special case which illustrates this principle; no one is saying that it necessarily applies to your problem.

Almutairi T.

I cant see any relation b/w my question and the link (Positive log-likelihood values happen), please explain further.


From: owner-statalist@hsphsun2.harvard.edu [owner-statalist@hsphsun2.harvard.edu] On Behalf Of Maarten Buis [maartenlbuis@gmail.com]

On Tue, Jan 24, 2012 at 1:28 AM, Almutairi T. wrote:
> I estimated the same model twice, but one in level and the other after taking the 1st difference (to get rid of AR1)
>
> I am looking for a goodness of fit measure for pooled time
> series models  other than R^2 (as it is not suitable for model with unsimilar dependent variable)
>
> Does log-likelihood values (and BIC or AIC) suitable to compare b/w these 2 models?

No, that to depends on the scale of the dependent variable, see for
example: <http://blog.stata.com/2011/02/16/positive-log-likelihood-values-happen/>


*
*   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/