Stephen Soldz,
You appear to have a single observation for each client
(175 observations and 175 clients) and you are specifying
a logistic regression model with a random intercept for client.
Unfortunately, the model is not identified (or at most very weakly).
If you had several observations per client, the within client
correlations would provide information on the random intercept
variance, but with a single obervation per client, the variance
cannot really be estimated: vastly different values will give nearly
the same log-likelihood. That is the reason you get a 'flat region'
error message in gllamm, as well as the message 'condition number
could not be computed'. You can estimate models without
random effects in gllamm using the init option.
I would be happy to help with your problem (please email
me privately). I also saw your message to the multilevel list
- it looks like gllamm could be useful for that problem as well.
Best regards,
Sophia
----- Original Message -----
From: "Stephen Soldz" <[email protected]>
To: <[email protected]>
Sent: Tuesday, September 24, 2002 2:36 PM
Subject: st: GLLAMM
> I'm trying to use GLLAMM for structural equation modeling with a binary
> outcome (school attendance 12 months after entering substance abuse
> treatment). I'm primarily interested in a path model, with manifest
> variables, though there may be one latent variable. What I really need to
> be able to do is translate a path diagram into gllamm code. I've been
going
> over the manual, but can't see how to do it. I don't really have a
> multilevel problem, so the i() variable is just client id.
>
> Even for an extremely simple model, predicting school attendance at 12
> months from attendance at admission, my model doesn't converge, whereas a
> simple logistic model does.
>
> ***********************
>
> --------------------------------------------------------------------------
--
> ---
> log: C:\Data\MIST\MISTStata\glamm_test.log
> log type: text
> opened on: 24 Sep 2002, 17:08:28
>
> .
> . logit inschool12 inschool,
>
> Iteration 0: log likelihood = -111.1575
> Iteration 1: log likelihood = -104.90761
> Iteration 2: log likelihood = -104.86177
> Iteration 3: log likelihood = -104.86177
>
> Logit estimates Number of obs =
> 175
> LR chi2(1) =
> 12.59
> Prob > chi2 =
> 0.0004
> Log likelihood = -104.86177 Pseudo R2 =
> 0.0566
>
> --------------------------------------------------------------------------
--
> --
> inschool12 | Coef. Std. Err. z P>|z| [95% Conf.
> Interval]
> -------------+------------------------------------------------------------
--
> --
> inschool | 1.203973 .3422187 3.52 0.000 .5332364
> 1.874709
> _cons | -.071459 .2674319 -0.27 0.789 -.5956158
> .4526978
> --------------------------------------------------------------------------
--
> --
>
> .
> . gllamm inschool12 inschool, i(twistid) family(binom) link(logit) dots
> nip(20)
>
> ................Iteration 0: log likelihood = -104.90004
> .....................................Iteration 1: log likelihood
> = -104.86177
> > (not concave)
>
............................................................................
> ...
> > ..............................numerical derivatives are approximate
> flat or discontinuous region encountered
>
............................................................................
> ...
> > .........................numerical derivatives are approximate
> flat or discontinuous region encountered
>
............................................................................
> ...
> > ...........................numerical derivatives are approximate
> flat or discontinuous region encountered
> ....Iteration 2: log likelihood = 77082.381 (not concave)
>
............................................................................
> ...
> > ..............................numerical derivatives are approximate
> flat or discontinuous region encountered
>
............................................................................
> ...
> > .........................numerical derivatives are approximate
> flat or discontinuous region encountered
>
............................................................................
> ...
> > ...........................numerical derivatives are approximate
> flat or discontinuous region encountered
> ....Iteration 3: log likelihood = 77161.125 (not concave)
>
............................................................................
> ...
> > ........................................................numerical
> derivatives
> > are approximate
> flat or discontinuous region encountered
>
............................................................................
> ...
> > ...........................numerical derivatives are approximate
> flat or discontinuous region encountered
> ....Iteration 4: log likelihood = 77163.479 (not concave)
>
............................................................................
> ...
> > ...............................numerical derivatives are approximate
> flat or discontinuous region encountered
>
............................................................................
> ...
> > .........................numerical derivatives are approximate
> flat or discontinuous region encountered
> ....................................................Iteration 5: log
> likeliho
> > od = 77163.479 (not concave)
>
> number of level 1 units = 175
> number of level 2 units = 175
>
> Condition Number could not be computed
>
> gllamm model
>
> log likelihood = 77163.479
>
> --------------------------------------------------------------------------
--
> --
> inschool12 | Coef. Std. Err. z P>|z| [95% Conf.
> Interval]
> -------------+------------------------------------------------------------
--
> --
> inschool | 4.53e+07 811588.2 55.86 0.000 4.37e+07
> 4.69e+07
> _cons | -1.39e+08 7423013 -18.72
> 0.000 -1.54e+08 -1.24e+08
> --------------------------------------------------------------------------
--
> --
>
>
> Variances and covariances of random effects
> --------------------------------------------------------------------------
--
> --
>
>
> ***level 2 (twistid)
>
> var(1): 4.143e+13 (0)
> --------------------------------------------------------------------------
--
> --
>
>
>
> .
> . log close
> log: C:\Data\MIST\MISTStata\glamm_test.log
> log type: text
> closed on: 24 Sep 2002, 17:12:50
> --------------------------------------------------------------------------
--
> ---
>
> Using adaptive quadrature seems to improve things;
>
>
> . gllamm inschool12 inschool, i(twistid) family(binom) link(logit) dots
> nip(20)
> > adapt
> ....
> ........................................Iteration 0: log likelihood
> = -104.86
> > 177
>
..................................................................Iteration
> 1:
> > log likelihood = -104.86177 (not concave)
>
> number of level 1 units = 175
> number of level 2 units = 175
>
> Condition Number could not be computed
>
> gllamm model
>
> log likelihood = -104.86177
>
> -------------------------------------------------------------------------
--
> ---
> inschool12 | Coef. Std. Err. z P>|z| [95% Conf.
> Interval]
> -------------+-----------------------------------------------------------
--
> ---
> inschool | 1.249865 .3549912 3.52 0.000 .5540955
> 1.945636
> _cons | -.0745689 .2784373 -0.27 0.789 -.6202959
> .4711581
> -------------------------------------------------------------------------
--
> ---
>
>
> Variances and covariances of random effects
> -------------------------------------------------------------------------
--
> ---
>
>
> ***level 2 (twistid)
>
> var(1): .17140917 (0)
> -------------------------------------------------------------------------
--
> ---
>
>
> ***********
>
> Among other questions, what is the meaning of the level 2 variance, and
why
> are the coefficients different than the simple logit model.
>
> Any guidance apreciated.
>
> Stephen Soldz
> Boston Graduate School of Psychoanalysis
> 1581 Beacon St.
> Brookline, MA 02146
> (617) 277-3915
>
> *
> * 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/
