st: gllamm with numerical problems and the setup with discrete latent vars

[Hey Sky <heyskywalker@yahoo.com>]

hey, all

just read the "ML numerical problems" and thought the problem I met

I am modeling training behavior with discrete latent variable. I assume there is 
only one discrete latent variable, and the code is as following. it works well.  


gen cons=1
eq mu1: cons

gllamm training_decision indep,  i(id) base(5) link(mlogit) family(binom) ip(f) nrf(1) eq(mu1)


the result of upper code for latent variable:
Probabilities and locations of random effects
------------------------------------------------------------------------------
 ***level 2 (id)
 
    loc1: -1.8365, .54899
  var(1): 1.0082459
    prob: 0.2301, 0.7699


my understanding to the loc1 is: the parameters for the discrete latent variable, type of person, 
for all people. that is:

epsilon_i= alfa_0 + alfa_1*mu
mu is type of person which I assume two types here.

but how to model it if  I assume people with different training decision have 
different probility to belong to type 1 or 2 person?


I have tried the following code and hope I can get:
epsilon_i= alfa_edu_0 + alfa_edu_1*mu
epsilon_i= alfa_wrk_t0 + alfa_work_t1*mu
epsilon_i= alfa_wrknt0 + alfa_wrknt1*mu


the code: 
eq mu1: training
eq mu2: work_with_training
eq mu3: work_no_training

gllamm training_decision indep,  i(id) link(mlogit) family(binom) ip(f) nrf(3) eq(mu1 mu2 mu3)

but the computer reports numerical problems. I think some place in the code is wrong,
because of fortran have given result. any suggestions about it? 

I wish I make myself clear and thanks a lot for any reply in advance


Nan
from Montreal




*
*   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/