Re: st: RE: r-square 4-level-logit-regression xtmelogit

[Martina Brandt <brandt@soziologie.uzh.ch>]

Dear Nick, dear list,

this is what I did - and what I got, using the same samples:

1) gllamm

xi: gllamm y x1...xn , i(l2 l3 l4) link(logit) family (binom) from(a) 
nip(10) adapt robust eform

/*Results (part):
Variances and covariances of random effects
------------------------------------------------------------------------------

***level 2 (l2)

    var(1): 1.5149127 (.55028054)

***level 3 (l3)

    var(1): 2.1240474 (.52440731)

***level 4 (l4)

    var(1): .02896267 (.02613964)
*/
gllapred phat1
sum phat1

/*
    Variable |       Obs        Mean    Std. Dev.       Min        Max
-------------+--------------------------------------------------------
       phat1 |      7680   -2.010795    1.506836  -7.667717   3.125869
*/

2)xtmelogit
xi: xtmelogit y x1...xn || l4: ||  l3: || l2:, variance or

/*
------------------------------------------------------------------------------
  Random-effects Parameters  |   Estimate   Std. Err.     [95% Conf. 
Interval]
-----------------------------+------------------------------------------------
land: Identity               |
                  var(_cons) |    .022862   .0201887      .0040499 .1290577
-----------------------------+------------------------------------------------
hhnr: Identity               |
                  var(_cons) |   2.171575   .4617767       1.43143 
3.294423
-----------------------------+------------------------------------------------
beobnr: Identity             |
                  var(_cons) |   1.434772   .5741192      .6549036 
3.143318
------------------------------------------------------------------------------
*/
predict phat2
sum phat2

/*
    Variable |       Obs        Mean    Std. Dev.       Min        Max
-------------+--------------------------------------------------------
       phat2 |      7680    .2411294    .2457556   .0005897   .9551997
*/

-> the ("explained") variance (SD(phat2)^2) now is much smaller compared 
to the overall variance than in the first example using gllamm. I must 
have overlooked something important, but I have no idea, what that could 
be? is it that predict (option mu) produces somethind different to 
gllapred? would i have to take only the fixed part (predict, xb)?

by the way: the example in the help function also indicates the 
level-1-indicator, but this produces the same results, just takes a lot 
longer. Does anyone know more about that?

Sorry for bothering you that extensively!

martina

_________________________
Martina Brandt
Universit�t Z�rich
Soziologisches Institut
Andreasstr. 15
CH-8050 Z�rich
Tel. +41(0)44 6352347

www.suz.uzh.ch/ages


Nick Cox schrieb:

> If you can demonstrate that results that should be 
> the same, from the same data and the same model, 
> are different from -xtmelogit- and -gllamm-, then 
> there is a problem. 
> I think you need at least to show (examples of) the commands 
> you used and the results you got for experts 
> on these commands to comment. 
> Nick 
> n.j.cox@durham.ac.uk 
> Martina Brandt
>
>
> > Dear Nic, thanks a lot for your answer - i know that 
> > pseudo-r-squares are 
> > quite tricky. But the probem here is, that the same 
> > pseudo-r-square changes 
> > using xtmemixed instead of gllamm because the estimated 
> > variance of phat in 
> > comparison to the level 1 to level 4 variances is much 
> > smaller than it is 
> > using gllamm?!
> >   On Mon, 15 Oct 2007 17:03:56 +0100
> >   "Nick Cox" <n.j.cox@durham.ac.uk> wrote:
> > > There is a entire bestiary of pseudo-R-squares
> > > based on different kinds of analogy to R-square, 
> > > strong, weak and otherwise. There is no
> > > reason in general why they should agree. 
>  
> Martina Brandt
>  
> > > > mc kelvey and zavoina suggest an r2 for multilevel logit 
> > regression, 
> > > > which is the variance of the predicted probabilities 
> > divided by the 
> > > > total variance of the model (=proportion of explained 
> > > > variance). in the 
> > > > four level model this would be 
> > > > (var(phat))/((var(phat)+((pi2)/3))+var(level2)+var(level3)+var
> > > > (level4))
> > > > (see snijders & bosker 1999: 225).
> > > > using gllamm i always had pseudo r2 around 0.20, and now 
> > > > using xtmelogit 
> > > > it is supposed to be only around 0.01. does anyone have an 
> > idea, why 
> > > > this could have happened and how these differences could 
> > be explained?
> *
> *   For searches and help try:
> *   http://www.stata.com/support/faqs/res/findit.html
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/