RE: st: loglikelihood and loglikelihood ratio

[Kit Baum <baum@bc.edu>]

<>
Using a LRT, one can only compare models with the same dependent  
variable (that is, the same observations, not just the same variable  
name). You apparently are comparing the LLF values of one subset with  
that of the whole sample. That makes no sense, as you cannot derive  
the model in the subset as a restricted version of the model for the  
whole sample. LRTs work the same as Wald "subset F" tests of the sort  
performed by -test-. They cannot be used to compare models fit over  
different observations.


Kit Baum   |   Boston College Economics and DIW Berlin   |   http://ideas.repec.org/e/pba1.html
An Introduction to Stata Programming   |   http://www.stata-press.com/books/isp.html
An Introduction to Modern Econometrics Using Stata   |   http://www.stata-press.com/books/imeus.html



On Mar 18, 2009, at 02:33 , Jingjjing wrote:

> My total dataset contains 7 regions. The estimation results for the 7
> regions are fine. All R square are positive, all LR chi() are
> positive, and all degree of freedom are right.
>
>
>
> 2. I choose the last 4 regions of the total 7regions and create a new
> data set(changed the dummy variables).
>
> Here, all the R square are positive, all LR chi() are positive. But
> the degree of freedom are strange.
>
> 4 regions Unrestricted
> -  
> ----------------------------------------------------------------------
> Equation          Obs  Parms        RMSE    "R-sq"       chi2        P
> -  
> ----------------------------------------------------------------------
> lnc                72     40    .0987079    0.9693   7.62e+07   0.0000
> sl                 72      8    .0230819    0.3033     417.70   0.0000
> se                 72      8    .0023162    0.9399    1246.72   0.0000
> sm                 72      8    .0292372    0.5744    1094.54   0.0000
> -  
> ----------------------------------------------------------------------
>  _cons in equation lnc are dropped, no other variable droped
>
>
> 4 regions Restricted
>
> -  
> ----------------------------------------------------------------------
> Equation          Obs  Parms        RMSE    "R-sq"       chi2        P
> -  
> ----------------------------------------------------------------------
> lnc                72     37    .0930231    0.9727   1.47e+07   0.0000
> sl                 72      7    .0195899    0.4982     347.11   0.0000
> se                 72      7    .0022661    0.9425    1275.65   0.0000
> sm                 72      7    .0270912    0.6346    1003.91   0.0000
> -  
> ----------------------------------------------------------------------
> No _cons dropped, no variable dropped
>
> Likelihood-ratio test                                  LR chi2(2)   
> =      6.71
> (Assumption: B nested in A)                            Prob > chi2  
> =    0.0350
>
> I am thinking if the degree of freedom changed from 3 to 2 because of
> _cons in unrestricted model is dropped, but kept in restricted model?
>
>
> 3. I chose the first 3 regions and created them as a new dataset
> (changed the dummy variables). When I estimated equations lnc, sl, sm,
> se, there are two negative R square values. So I changed them to ln,
> sl, sk, se and got one negative R-sq this time. LR chi() here are
> negative.

*
*   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/