I am not sure what the -r2_mz- command does, but in general there are a number of theoretical problems with calculating
a single R2 for logistic random effects models. Certainly I would not expect any single metric of "variance explained"
to be comparable between the -logit- and -xtlogit- models; for the -xtlogit-, there is variance explained at 2 different
levels, and surprising things happen when you add explanatory variables. See for example [Snijders T & Boskers R,
Multilevel Analysis: An introduction to basic and advanced multilevel modelling. Sage. 2012.], Chapter 17.
If your goal is to compare the performance of the two models, I suggest using the C-statistic, which can be compared
directly.
Jeph
On 4/25/2013 8:25 AM, Andreas Schiffelholz wrote:
For both calculations I used Dirk's -r2_mz- comand, calculating the McKelvey & Zavoina's R². So if it is working with
both -logit- and -xtlogit-, I think it is calculationg the same. I just wanted to make sure the comand is working
correctly in both cases, that's why I asked.
Andreas
Am 25.04.2013 14:08, schrieb Nick Cox:
Among many other issues, you need to be clear that you are calculating
exactly the same thing. There are numerous different definitions of
Pseudo R-square. Only the same formula applied in different situations
is comparable.
Nick
njcoxstata@gmail.com
On 25 April 2013 12:42, Andreas Schiffelholz
<Andreas.Schiffelholz@gmx.de> wrote:
Dirk,
thank you very much, that helps a lot!
When I use this comand to calculate the Pseudo R2 for the -logit- and the
-xtlogit, re-, they differ a lot (e.g. 0.09 vs. 0.19). Is it normal that the
r2 differs so much between two models using the same independent variables
or can your command only be used for -xtlogit- and the -logit- is not
correct?
Cheers,
Andreas
* You can calculate McKelvey & Zavoina's R² using -r2_mz- from SSC, see
-findit r2_mz-.
*
* Dirk
*
* Wed, 24 Apr 2013 09:49:31 +0200, Andreas
Schiffelholz<Andreas.Schiffelholz@gmx.de> wrote:
Hello everybody,
I have a very basic question. I'm currently working on a logistic panel
regression with random effects -xtlogit, re-. While I'm able to
calculate pseudo R² for my pooled models with cluster-robust standard
errors -logit, cluster (...)-, I'm not able to calculate these for the
- -xtlogit, re- model.
Is there a way to calculate these pseudo R²? If not, what other
comparable criteria would you use instead?
Thanks a lot in advance!
Andreas Schiffelholz
