Re: st: RE: RE: Xtgee - noconstant option

[Nick Cox <njcoxstata@gmail.com>]

How much difference you will see in practice clearly depends on how
close your means are to zero. I can't see that you should find this
impressive or that you have a discovery here. If you change the
criterion of evaluation, the evaluation changes.

Nick

On Thu, Mar 22, 2012 at 6:03 PM, Solon Moreira <sm.ino@cbs.dk> wrote:
> Thanks a lot Nick and Maarten,
>
> I also tested other models and apparently the effects of increasing the significance also increases when I use noconstant. I am thinking that this effect may also being caused by the fact that I have a high number of observations clustered around the origin, so maybe that is why it is fitting better.
>
> Best,
> Solon
>
> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Nick Cox
> Sent: 22. marts 2012 15:19
> To: 'statalist@hsphsun2.harvard.edu'
> Subject: st: RE: Xtgee - noconstant option
>
> The complication here of using -xtgee- is secondary to the main issue. Consider a very simple regression y = a + bx. The key question is whether the regression using x as predictor improves on a null model in which the mean of y is used to predict y. Change the regression to y = bx and the key question is whether the regression using x as predictor improves on a null model in which zero is used to predict y. On the second criterion almost any model does better. The exceptions are when the mean is exactly zero and the models end up the same. But the comparison is spurious. Any model does better compared with a lower benchmark than compared with a higher benchmark.
>
> This is not to say that regression through the origin makes no sense. It sometimes does. Your case does not sound like one of them.
>
> A Google reveals many longer explanations. It is nice that what comes first on Googling "r square no intercept" is from our Stata-loving friends at UCLA and is at http://www.ats.ucla.edu/stat/mult_pkg/faq/general/noconstant.htm
>
> Nick
> n.j.cox@durham.ac.uk
>
>
> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Solon Moreira
> Sent: 22 March 2012 13:44
> To: statalist@hsphsun2.harvard.edu
> Subject: st: Xtgee - noconstant option
>
> Dear Statalist members,
>
> I'm trying to run an xtgee model predicting firm's market-share on t+1 using a group of explanatory variables regarding firm characteristics (continuous) and contractual aspects of technology transfer contracts (dummy). I am using the following specifications: link (identity) family (Gaussian) corr (exchangeable). My issue here is related with the fact that when I set the option noconstant (go through the origin) all the coefficients become much more significant. Although I would expect that for some variables it makes sense, I do not fully understand this effect as well if it is right to use in that option for an xtgee model. Would someone have a suggestion if the use of noconstant is not indicated in this case?
>
> *
> *   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/
>
> *
> *   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/
>
>
>
> *
> *   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/

*
*   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/